NEUCOM - 2512

• (2025). GDViT: Group-level decorrelation-based vision transformer for domain generalization. NEUCOM, 657, 131624. (www)

  @article{ , abstract = {Causality-inspired domain generalization aims to improve model generalization by removing correlations between relevant and irrelevant features. However, a key challenge lies in effectively distinguishing the two. Existing methods, lacking explicit feature grouping, often eliminate all feature correlations indiscriminately, which disrupts the internal structure of relevant features and degrades generalization performance. In this work, we propose a group-level decorrelation-based vision Transformer that explicitly separates features (tokens) into relevant and irrelevant groups. This design preserves the internal correlations within relevant features while removing the correlations between the two groups. To achieve this, we introduce a feature grouping module that guides the separation process, followed by a grouping Transformer encoder that performs inter-group decorrelation, enabling the model to focus more on task-relevant information. Additionally, a supervised contrastive loss is employed to further enhance generalization. Extensive experiments demonstrate that our method significantly improves out-of-distribution performance. Visual analysis further shows that our model suppresses attention to irrelevant features, mitigating spurious correlations and resulting in more stable predictions. Our approach achieves strong performance in both multi-source and single-source domain generalization settings.}, archive = {J_NEUCOM}, author = {Wenqiang Tang and Zhouwang Yang and Yanzhi Song}, doi = {10.1016/j.neucom.2025.131624}, journal = {Neurocomputing}, month = {12}, pages = {131624}, shortjournal = {Neurocomputing}, title = {GDViT: Group-level decorrelation-based vision transformer for domain generalization}, volume = {657}, year = {2025}, }
• (2025). Distributed GNE seeking for aggregative games under event-triggered communication: Predefined-time convergent algorithm design. NEUCOM, 657, 131623. (www)

  @article{ , abstract = {Within the context of distributed generalized Nash equilibrium (GNE) seeking in aggregative games, it is challenging yet interesting to design fast GNE seeking algorithms using energy-efficient communication strategy. However, most existing distributed GNE seeking algorithms can only achieve asymptotic convergence under continuous-time communication setting, resulting in a slower convergence rate and greater consumption of communication resources. In this paper, by exploiting two time-varying gain feedback functions, we present a new kind of distributed GNE seeking algorithm by integrating predefined-time control law with event-triggered communication strategy. It is theoretically shown that the proposed algorithm can solve the predefined-time GNE seeking problem for aggregative games with Zeno behavior being avoided during the seeking process. Compared with the existing algorithms, the present one exhibits several salient features: 1) the convergence time can be preset according to task requirements; 2) the communication resources can be significantly saved by the event-triggered mechanism; and 3) the proposed algorithms exhibit simplicity in their structures and possess the advantage of easy implementability.}, archive = {J_NEUCOM}, author = {Zhijun Guo and Lingwei Zeng and Jinlei Cheng and Pengwen Xiong and Qian Li}, doi = {10.1016/j.neucom.2025.131623}, journal = {Neurocomputing}, month = {12}, pages = {131623}, shortjournal = {Neurocomputing}, title = {Distributed GNE seeking for aggregative games under event-triggered communication: Predefined-time convergent algorithm design}, volume = {657}, year = {2025}, }
• (2025). Graph-based multi-modal MRI analysis with probabilistic attention for stroke lesion detection. NEUCOM, 657, 131620. (www)

  @article{ , abstract = {Stroke lesion detection in brain MRI remains challenging as existing deep learning methods process single modalities and ignore anatomical boundaries, limiting clinical adoption. We develop a graph-based framework that integrates neuroanatomical priors and multi-modal imaging for automated stroke lesion detection. Our approach uses anatomically-constrained supervoxel generation and graph attention networks with probabilistic attention attribution for interpretable lesion detection. Evaluated on the SOOP dataset (1715 subjects including 1461 stroke patients), our method achieves a Dice coefficient, sensitivity, and ROC-AUC of 0.85 ± 0.03, 0.88, and 0.94, respectively, outperforming CNN baselines by 15 %. The framework provides clinically meaningful attention maps and accurate automated stroke analysis.}, archive = {J_NEUCOM}, author = {Luis R. Mercado-Diaz and Derek Aguiar and Hugo F. Posada-Quintero}, doi = {10.1016/j.neucom.2025.131620}, journal = {Neurocomputing}, month = {12}, pages = {131620}, shortjournal = {Neurocomputing}, title = {Graph-based multi-modal MRI analysis with probabilistic attention for stroke lesion detection}, volume = {657}, year = {2025}, }
• (2025). LEAD: LLM-enhanced deep reinforcement learning for stable decision-making in critical autonomous driving scenarios. NEUCOM, 657, 131619. (www)

  @article{ , abstract = {Large Language Models (LLMs) have demonstrated significant potential in addressing decision-making problems in the field of autonomous driving due to their strong reasoning capabilities. However, deploying LLMs in real-world driving scenarios often encounters challenges such as high computational requirements, elevated costs, and increased latency. On the other hand, Deep Reinforcement Learning (DRL) exhibits strong adaptability to decision-making tasks in autonomous driving with a relatively smaller parameter scale. Nevertheless, DRL agents often suffer from low exploration efficiency and high sensitivity to parameter variations. To address the above issues, we propose an LLM-Enhanced Autonomous Driving (LEAD) training framework, which integrates a high-level agent based on LLMs into the training process of DRL models, effectively improving the policy learning efficiency and generalization capability of DRL models. During the early stage of training, a dynamic intervention mechanism is introduced to identify key decision points within the DRL model, and a predefined expert guidance algorithm is utilized to integrate high-level decision strategies from LLMs into these critical nodes. During the later stage of training, the DRL model transitions into an autonomous optimization phase, where the agent, enhanced with LLM priors, continuously interacts with the environment to further refine the policy network, ultimately surpassing the performance of the LLM-based agent. Experimental results demonstrate that the LEAD-PPO model, built upon the proposed framework, reduces collision rates by 49.49 % and 59.4 % in low-density and high-density scenarios, respectively, during training compared to the baseline model. In the testing phase, the DRL model optimized through LEAD achieves task completion rates that are 9.60 %, 35.94 %, and 65.63 % higher than those of the baseline model in simple, moderate, and difficult scenarios, respectively. Overall, the proposed LEAD framework significantly improves the robustness, sample efficiency, and generalization ability of DRL models.}, archive = {J_NEUCOM}, author = {Dongwei Xu and Enwen Qiao and Tongcheng Gu and Hongda Fu and Chengju Sun and Haifeng Guo and Yuqing Liu}, doi = {10.1016/j.neucom.2025.131619}, journal = {Neurocomputing}, month = {12}, pages = {131619}, shortjournal = {Neurocomputing}, title = {LEAD: LLM-enhanced deep reinforcement learning for stable decision-making in critical autonomous driving scenarios}, volume = {657}, year = {2025}, }
• (2025). Fixed/Preassigned-time synchronization of delayed fuzzy memristive neural networks with reaction–diffusion terms. NEUCOM, 657, 131618. (www)

  @article{ , abstract = {How to effectively handle the synchronization of reaction–diffusion fuzzy memristive neural networks (MNNs) and shorten their synchronization time has become a worthwhile and meaningful issue to study. This paper mainly studies fixed-time synchronization (FXTS) and preassigned-time synchronization (PATS) problems for delayed fuzzy memristive neural networks (DFMNNs) with reaction–diffusion terms. First, a DFMNNs model with reaction–diffusion terms is introduced, which can effectively describe the spatial distribution characteristics of the network. Second, through the Lyapunov stability theory, the FXTS criterion and the upper limit of the settling-time (ST) are obtained. Subsequently, a state feedback controller is proposed to ensure that the system achieves synchronization within a specified time, and the synchronization time is independent of initial conditions and control parameters, which gives the designed controller a wider range of applications. Finally, two examples are presented to illustrate the effectiveness of the results.}, archive = {J_NEUCOM}, author = {Hanrui Chen and Dongbing Tong and Qiaoyu Chen}, doi = {10.1016/j.neucom.2025.131618}, journal = {Neurocomputing}, month = {12}, pages = {131618}, shortjournal = {Neurocomputing}, title = {Fixed/Preassigned-time synchronization of delayed fuzzy memristive neural networks with reaction–diffusion terms}, volume = {657}, year = {2025}, }
• (2025). Beyond sparsity: An empirical study of structured collaboration in modular AI. NEUCOM, 657, 131616. (www)

  @article{ , abstract = {In Mixture-of-Experts (MoE) architectures, the prevailing paradigm emphasizes sparse expert activation for computational efficiency. This paper explores an alternative architectural approach centered on structured collaboration, hypothesizing that the quality and nature of inter-module interactions are as significant as computational cost. We present a series of targeted proof-of-concept experiments to validate three distinct principles of structured interaction, inspired by cognitive science. First, we demonstrate that a hierarchical fusion mechanism, modeled on the brain's segregated visual pathways, enhances compositional reasoning on the VQA v2.0 benchmark. Second, by employing a redesigned reinforcement learning task in MiniGrid, we demonstrate that a system-wide differentiated credit assignment (SDCA) mechanism, with conflict detection learned end-to-end, facilitates more robust policy learning. Third, we ascertain that integrating reasoning "tools" as co-adaptive modules offers superior out-of-distribution robustness on the DROP dataset compared to a more powerful baseline agent utilizing external LLM-based tools. Our work provides concrete validation for these principles, highlighting a series of trade-offs between performance, robustness, and efficiency, and suggesting that prioritizing cognitive synergy over simple sparsity offers a promising direction for future research in modular AI.}, archive = {J_NEUCOM}, author = {Xiaofei Zhou and Soohong Kim and Yiru Wang and Kailin Zhang}, doi = {10.1016/j.neucom.2025.131616}, journal = {Neurocomputing}, month = {12}, pages = {131616}, shortjournal = {Neurocomputing}, title = {Beyond sparsity: An empirical study of structured collaboration in modular AI}, volume = {657}, year = {2025}, }
• (2025). A graph transformer-driven reinforcement learning based on popularity for mining complex network key nodes. NEUCOM, 657, 131614. (www)

  @article{ , abstract = {Key nodes in complex networks play a crucial role in maintaining the stability, functionality, and robustness of networked systems. Accordingly, the accurate identification of such nodes is of fundamental importance. Their significance spans multiple domains, including communication systems, transportation infrastructures, life sciences, and social networks. Existing algorithms for key node identification typically rely on heuristic measures or standard deep reinforcement learning frameworks. However, these approaches often suffer from limited feature extraction capabilities, high computational complexity, and insufficient generalizability, and a lack of dynamic adaptability. To overcome these limitations, this study proposes a novel architecture, GTRP (Graph Transformer-Driven Reinforcement Learning Based on Popularity). GTRP extends Epidemic-aware Heterogeneous Graph Transformer (GT) by introducing distinct attention mechanisms for both nodes and edges, enabling the integration of local structural features and global propagation properties. In addition, GTRP incorporates Dual-dynamics Reward Optimization (DR) to identify key nodes based on a network disintegration strategy. The model is trained on randomly generated Barabási–Albert (BA) networks and evaluated on synthetic networks of varying scales as well as multiple real-world network scenarios. Comparative experiments with six representative algorithms demonstrate that GTRP achieves substantial performance improvements—outperforming existing methods by 6.30 % in unweighted networks and 15.90 % in weighted networks. These results underscore the potential of GTRP to advance key node detection in complex network analysis.}, archive = {J_NEUCOM}, author = {Kaili Wang and Muqing Wu and Min Zhao}, doi = {10.1016/j.neucom.2025.131614}, journal = {Neurocomputing}, month = {12}, pages = {131614}, shortjournal = {Neurocomputing}, title = {A graph transformer-driven reinforcement learning based on popularity for mining complex network key nodes}, volume = {657}, year = {2025}, }
• (2025). An efficient classification and error correction knowledge distillation framework for remaining useful life prediction of bearings in air turbine starter. NEUCOM, 657, 131611. (www)

  @article{ , abstract = {Prognostics and Health Management (PHM) is critical for industrial equipment maintenance, whose core task is to predict the Remaining Useful Life (RUL) of a system or component accurately. However, traditional deep learning-based approaches often demand significant computational and memory resources, limiting their feasibility for edge deployment. As an effective model compression technique, Knowledge Distillation (KD) has emerged as a core strategy for enabling edge intelligence by transferring knowledge from a teacher model to a lightweight student model. However, traditional KD methods exhibit a high dependency on the teacher model's output. This dependency limits the student model's capacity for autonomous error correction, impairing its distillation performance. To solve these problems, this paper proposes a novel Classification and Error Correction Knowledge Distillation (CEKD) framework. The framework employs Gaussian kernel-based feature entropy to dynamically evaluate teacher models' predictive capabilities, facilitating comprehensive assessment and sample differentiation. Furthermore, the knowledge self-reflection learning strategy extends error correction to continuous dynamic adjustment, enabling deep optimization of complex data. Experimental results on the air turbine starter bearing datasets show that CEKD surpasses KD methods by improving MAE and RMSE by 79.8 % and 78.6 % on average, while reducing memory consumption and inference time by nearly 10 × and 8 × , respectively, enabling deployment on resource-constrained devices.}, archive = {J_NEUCOM}, author = {Runxia Guo and Jingxu Yi and Xianfeng Luo}, doi = {10.1016/j.neucom.2025.131611}, journal = {Neurocomputing}, month = {12}, pages = {131611}, shortjournal = {Neurocomputing}, title = {An efficient classification and error correction knowledge distillation framework for remaining useful life prediction of bearings in air turbine starter}, volume = {657}, year = {2025}, }
• (2025). Finite-time lag consensus and finite-time h∞ lag consensus for nonlinear multi-agent systems under communication delay. NEUCOM, 657, 131609. (www)

  @article{ , abstract = {In this research, the finite-time lag consensus (FTLC) and finite-time H ∞ lag consensus (FTHLC) problems for first-order multi-agent systems (MASs) are studied. On the one hand, a new state feedback controller considering the communication delay between agents is proposed. Besides, based on several inequality scaling techniques and finite-time stability theory, a sufficient criterion is derived to guarantee the FTLC of MASs. On the other hand, an adaptive state feedback controller and the corresponding adaptive law are put forward, which can also help MASs realize lag consensus in finite time without any additional conditions. Moreover, to address inevitable external disturbances in practice, the proposed control strategies are further enhanced to achieve FTHLC, which further expands the application range of the research results. Finally, the effectiveness of these provided FTLC and FTHLC control schemes in different scenarios is manifested through some numerical simulation examples.}, archive = {J_NEUCOM}, author = {Song Gao and Jin-Liang Wang and Kun Ling and Shun-Yan Ren and Ming-Zhu Wei and Bei Peng}, doi = {10.1016/j.neucom.2025.131609}, journal = {Neurocomputing}, month = {12}, pages = {131609}, shortjournal = {Neurocomputing}, title = {Finite-time lag consensus and finite-time h∞ lag consensus for nonlinear multi-agent systems under communication delay}, volume = {657}, year = {2025}, }
• (2025). Distributed generalized nash equilibrium computing based on dynamic tracking mechanism for nonsmooth aggregative games over time-varying networks. NEUCOM, 657, 131608. (www)

  @article{ , abstract = {This paper investigates the computation of generalized Nash equilibrium in aggregative games with coupling constraints over time-varying networks. The player’s cost objective comprises a differentiable function dependent on the aggregate of all players’ decisions and a possibly non-smooth term with a linear mapping. In this context, designing solution methods for such game formulation is relatively scarce. We thus develop a fully distributed equilibrium-seeking algorithm that accommodates time-varying communication networks while circumventing the need for global decision information. The proposed algorithm synergistically embeds dynamic tracking of aggregate decisions through a consensus-based mechanism with projected pseudo-gradient updates, augmented by a proximal splitting scheme to handle non-smooth components. Theoretically, we establish convergence guarantees to the variational equilibrium through a new operator splitting framework. Finally, numerical experiments are conducted to substantiate the validity of the proposed algorithm.}, archive = {J_NEUCOM}, author = {Liang Ran and Huaqing Li and Zheng Wang and Lifeng Zheng and Jun Li and Zhe Li}, doi = {10.1016/j.neucom.2025.131608}, journal = {Neurocomputing}, month = {12}, pages = {131608}, shortjournal = {Neurocomputing}, title = {Distributed generalized nash equilibrium computing based on dynamic tracking mechanism for nonsmooth aggregative games over time-varying networks}, volume = {657}, year = {2025}, }
• (2025). Neural network-based distributed adaptive synchronization control for multi-weighted complex networks with aperiodic intermittent coupling. NEUCOM, 657, 131604. (www)

  @article{ , abstract = {This paper studies the synchronization control for multi-weighted complex networks (MWCNs) with unknown disturbances and aperiodic intermittent coupling. Firstly, an adaptive neural network strategy is used to approximate the unknown components derived from nonlinear function, while a novel continuous function is proposed by utilizing the idea of time-varying boundary layer technique to deal with the influence of approximation errors. Secondly, different from the continuous coupling and periodic intermittent coupling mechanisms in existing works, an aperiodic intermittent coupling mechanism is taken into consideration in MWCNs. Thirdly, to synchronize MWCNs under aperiodic intermittent coupling, adaptive strategies are developed to adjust coupling strengths and coupling gains based on all edges and partial edges, respectively. Note that these two strategies are fully distributed, i.e., they do not require any global information. Finally, some numerical simulations are provided to verify the effectiveness of theoretical results.}, archive = {J_NEUCOM}, author = {Bin Zhang and Dan Liu and Binrui Wang and Kaibo Shi}, doi = {10.1016/j.neucom.2025.131604}, journal = {Neurocomputing}, month = {12}, pages = {131604}, shortjournal = {Neurocomputing}, title = {Neural network-based distributed adaptive synchronization control for multi-weighted complex networks with aperiodic intermittent coupling}, volume = {657}, year = {2025}, }
• (2025). An effective framework with hybrid augmentation for visual reinforcement learning generalization. NEUCOM, 657, 131602. (www)

  @article{ , abstract = {Current studies in Visual Reinforcement Learning focus on developing policies to acclimate to unknown environments through data augmentation. This paper aims to develop a new methodology to improve upon existing results. To this end, we first categorize existing methods into three groups based on the focus of augmentation: Task-Aware Augmentation, Image-Processing Augmentation, and Image-Scene Augmentation. Subsequently, we establish a unified framework that integrates these three augmentation categories. The core of our framework is hybrid data augmentation, which enhances data diversity. In this framework, we employ hyperspherical space and regularization techniques to address the side effects of such augmentation, specifically the discrepancy between augmented and original data, as well as the instability associated with hybrid augmentation. Finally, we evaluate the proposed framework across three benchmarks, demonstrating its significant advantages over current state-of-the-art methods. Notably, our framework outperforms existing approaches by an average of 4.59 % across 10 tasks in DMC-GB, 28.81 % across 6 tasks in Robosuite, and 20.50 % across 4 tasks in Adroit. The code for our framework will be released at https://github.com/csufangyu/MuHA .}, archive = {J_NEUCOM}, author = {Yu Fang and Xuehe Zhang and Haoshu Cheng and Xizhe Zang and Changle Li and Jie Zhao}, doi = {10.1016/j.neucom.2025.131602}, journal = {Neurocomputing}, month = {12}, pages = {131602}, shortjournal = {Neurocomputing}, title = {An effective framework with hybrid augmentation for visual reinforcement learning generalization}, volume = {657}, year = {2025}, }
• (2025). Enhancing robust generalization through appropriate adversarial example attack intensity. NEUCOM, 657, 131599. (www)

  @article{ , abstract = {Deep Neural Networks (DNNs) are notoriously susceptible to adversarial examples. To mitigate the impact of well-designed adversarial attacks on network models, researchers have developed various defense mechanisms, among which adversarial training has emerged as one of the most effective strategies to date. Adversarial training aims to augment training data with adversarial examples, thus giving DNNs a certain degree of robustness to defend against adversarial attacks. However, while obtaining adversarial robustness, this method comes at the cost of reducing the generalization performance, manifested in the reduced classification effect of clean test datasets. Researchers have been actively seeking to counter the balance between adversarial robustness and model generalization. We believe that the key to balancing these two aspects lies in identifying appropriate adversarial examples. Overly potent examples can lead to a decline in clean accuracy, whereas weaker examples may offer limited robustness. Based on our analysis, a new adversarial example generation algorithm called Denoising Projection Gradient Descent (DPGD) was proposed. DPGD adds a purification module and a constraint in generating adversarial examples, the former is used to limit the influence of too strong adversarial examples on model training and the latter is used to ensure the necessary attack intensity. Combining DPGD with the framework of traditional adversarial training, we obtain the Diffusion Adversarial Training (DifAT) approach. To verify the effectiveness of our proposed method, we conducted extensive experiments on benchmark datasets, including CIFAR-10, CIFAR-100, and Tiny-Imagenet. Our results demonstrate the effectiveness of DifAT in improving the robustness of DNNs while maintaining or even improving their generalization performance.}, archive = {J_NEUCOM}, author = {Xiaoguo Ding and Liangjian Zhang and Qiqi Bao and Yaguan Qian and Bin Wang and Zhaoquan Gu and Yanchun Zhang}, doi = {10.1016/j.neucom.2025.131599}, journal = {Neurocomputing}, month = {12}, pages = {131599}, shortjournal = {Neurocomputing}, title = {Enhancing robust generalization through appropriate adversarial example attack intensity}, volume = {657}, year = {2025}, }
• (2025). User identification based on the topology consistency of cross-layer common neighbors in social network. NEUCOM, 657, 131591. (www)

  @article{ , abstract = {Nowadays, it has become a common practice to create multiple accounts on various social networks for online recreation. When accounts in different networks share similar structural features, i . e . , they have topology consistency, they may belong to the same individual. However, having multiple accounts for the same person across different networks can be inconvenient and uncertain, leading to difficulties in accurate recommendations. Therefore, some researchers have focused on identifying users within single network layers, but without involving the information from various network platforms, resulting in identification confusion and reduced algorithm accuracy. This paper proposes a novel topology consistency-based link prediction algorithm (Topology Consistency: TC) for user identification, combining separate layers of a multilayer network into a single-layer network to include more layer information. TC applies topology information from the cross-layer common neighbors produced in layer combination to distinguish target node pairs and utilizes matrix computation to reduce complexity. Furthermore, to address controversial identification situations appearing after layer combination, the edges between the cross-layer common neighbors are innovatively considered. Finally, experimental results in real-world and artificial networks show that TC has superior performance to state-of-the-art algorithms and has applicability and practicality.}, archive = {J_NEUCOM}, author = {Yujie Yang and Shuai Cao and Long Wang and Dong Liu and Marcus Kaiser}, doi = {10.1016/j.neucom.2025.131591}, journal = {Neurocomputing}, month = {12}, pages = {131591}, shortjournal = {Neurocomputing}, title = {User identification based on the topology consistency of cross-layer common neighbors in social network}, volume = {657}, year = {2025}, }
• (2025). Implementation of linear optics network for pattern recognition task via the generation of continuous variable entangled states. NEUCOM, 657, 131588. (www)

  @article{ , abstract = {Linear optics neural networks or optical neural networks offer potential advantages over traditional electronic neural networks in terms of speed, energy efficiency, scalability, and improved parallelism, particularly for high-bandwidth applications. The use of photonics allows for more compact and integrated neural network designs, potentially enabling the development of larger and more complex networks. A linear optics network is developed to implement a quantum classifier. Indeed, the designed network is a quantum circuit consisting of some Gaussian gates such as displacement, noiseless linear amplification (NLA), squeezer and Green machine. At first, the classical inputs are encoded with the help of position-displacement operator to prepare single-mode coherent states. Then, the amplitudes of the coherent states are amplified by passing through NLA elements followed by squeezer gates that may transform classical coherent states into nonclassical ones. Finally, the transformed coherent states are fed into the Green machine which provides entangled states as the outcome of the network. As a primary goal of this work, the network generates a multi-mode entangled state by applying the displacement operator on the vacuum state encoded classical data. Besides, it is shown that the output state of the circuit may possess squeezing characteristics as another nonclassical feature. In the continuation, as a practical application, the network is implemented to perform some pattern recognition tasks. At first, the Bayes theorem is employed to define discriminant functions to perform a general classification task, then the outcome distribution of the network is utilized to classify some corrupted LEDs that display English letters. Finally, we show that the outcome of the circuit may be manipulated to embed classical neural networks into a continuous-variable variational quantum circuit (VQC). The network is trained via the logistic regression algorithm with the MNIST database. The results show that the digits can be recognized with relatively high accuracy. Ongoing research is focused on developing new linear optical techniques for machine learning, improving the efficiency and scalability of optical networks, and exploring new applications of linear optics in machine learning and quantum computing. Hence, such a quantum circuit may also be used to design novel high-accuracy pattern recognition devices.}, archive = {J_NEUCOM}, author = {Ebrahim Ghasemian and Mohammad Kazem Tavassoly and Habib Rostami}, doi = {10.1016/j.neucom.2025.131588}, journal = {Neurocomputing}, month = {12}, pages = {131588}, shortjournal = {Neurocomputing}, title = {Implementation of linear optics network for pattern recognition task via the generation of continuous variable entangled states}, volume = {657}, year = {2025}, }
• (2025). Learning generic and specific prompts with contrastive constraints for multi-task visual scene understanding. NEUCOM, 657, 131586. (www)

  @article{ , abstract = {Multi-task learning has emerged as a crucial research direction in the field of computer vision, offering improved performance and efficiency across multiple tasks. Recent studies have incorporated prompt learning into multi-task learning to enhance the interaction between prompt vectors and image representations. However, these studies fail to consider the inter-task and intra-task relations of prompt vectors under multi-task scenarios. To address this issue, we propose learning Generic and Specific Prompts (GSPrompt) with contrastive constraints for multi-task visual scene understanding. Our approach assumes that each task possesses both commonality and individuality, leading us to design two distinct types of prompt vectors: task-generic prompts and task-specific prompts. By constraining the prompt vectors through pulling task-generic prompts and pushing task-specific prompts, we enable multi-task models to learn prompts capable of adapting to multiple tasks simultaneously. Extensive experiments on NYUD-v2, PASCAL-Context, and Cityscapes show that GSPrompt learns effective prompts and achieves state-of-the-art performance. The code is publicly available at https://github.com/teeyohan/GSPrompt-main .}, archive = {J_NEUCOM}, author = {Tianyu Han and Zhimin Xu and Wanying Li and Haohao Hu and Xinxin He and Song He and Peng Zan and Xiaochen Bo}, doi = {10.1016/j.neucom.2025.131586}, journal = {Neurocomputing}, month = {12}, pages = {131586}, shortjournal = {Neurocomputing}, title = {Learning generic and specific prompts with contrastive constraints for multi-task visual scene understanding}, volume = {657}, year = {2025}, }
• (2025). Vision-based passenger head detection for carriage crowd density estimation. NEUCOM, 657, 131584. (www)

  @article{ , abstract = {Carriage crowd density monitoring is a key component in developing intelligent transportation systems, such as maglev transportation system. Surveillance images captured by sensors, such as carriage monitoring cameras, serve as a new solution for estimating crowd density inside the carriage due to their wide coverage and real-time updates. In this study, a passenger head detection dataset (PHD) is developed using 3717 images acquired from carriage surveillance. Based on these images, over 67,215 head instances are precisely annotated manually. To address the issue of insufficient feature fusion in existing detection algorithms, an efficient cross-scale feature enhancement (CFE) module is proposed and introduced into the advanced YoloX model. The PHD dataset is, to the best of our knowledge, the first public dataset of surveillance images for carriage crowd density estimation. To prove the usability of the PHD dataset and the validity of the proposed method, 12 different versions of detectors are applied and compared. The results demonstrate the performance of these algorithms in the detection of passenger heads. Our research offers a new approach for carriage crowd density estimation. The dataset is publicly available at: https://github.com/Xujiajing111/PHD .}, archive = {J_NEUCOM}, author = {Jiajing Xu and Mingda Zhai and Yuan Tian and Jun Wu}, doi = {10.1016/j.neucom.2025.131584}, journal = {Neurocomputing}, month = {12}, pages = {131584}, shortjournal = {Neurocomputing}, title = {Vision-based passenger head detection for carriage crowd density estimation}, volume = {657}, year = {2025}, }
• (2025). Single-client GAN-based backdoor attacks for asynchronous federated learning. NEUCOM, 657, 131580. (www)

  @article{ , abstract = {Federated Learning (FL) enables distributed collaborative training while preserving data privacy; however, it demonstrates significant vulnerability to backdoor attacks. Existing attack methodologies predominantly require control of numerous malicious clients to achieve efficacy and largely neglect asynchronous FL scenarios. In response to these limitations, we propose a novel GAN-based backdoor attack framework capable of injecting effective and covert backdoors with minimal malicious client participation, functioning efficiently across both synchronous and asynchronous environments. Our framework operates effectively with a single malicious client, eliminating the need for coordination among multiple adversarial participants or prior knowledge of benign client data distributions. This reduction in resource requirements enhances the framework's practicality in real-world FL implementations. The malicious client employs a Generative Adversarial Network to synthesize adversarial samples containing predefined triggers, which are subsequently incorporated into local training datasets. The concurrent training on legitimate and triggered data enhances attack effectiveness, while gradient injection—manipulating differences between local and global gradients to introduce strategic noise—facilitates backdoor embedding with improved stealth characteristics. Empirical evaluations demonstrate that in a configuration of 200 clients with a single attacker, our framework achieves attack success rates of 98.66 % on MNIST and 86.29 % on CIFAR-10 datasets. Comprehensive experimentation across both datasets substantiates the framework's effectiveness, imperceptibility, and resilience in synchronous and asynchronous FL environments. This research contributes significant insights into backdoor attack strategies in FL, particularly within asynchronous contexts, and underscores the imperative for developing robust defensive countermeasures.}, archive = {J_NEUCOM}, author = {Siyu Guan and Chunguang Huang and Hai Cheng}, doi = {10.1016/j.neucom.2025.131580}, journal = {Neurocomputing}, month = {12}, pages = {131580}, shortjournal = {Neurocomputing}, title = {Single-client GAN-based backdoor attacks for asynchronous federated learning}, volume = {657}, year = {2025}, }
• (2025). UA-PDFL: A personalized approach for decentralized federated learning. NEUCOM, 657, 131579. (www)

  @article{ , abstract = {Federated learning (FL) is a privacy preserving machine learning paradigm designed to collaboratively learn a global model without data leakage. Specifically, in a typical FL system, the central server solely functions as a coordinator to iteratively aggregate the collected local models trained by each client, potentially introducing single-point transmission bottleneck and security threats. To mitigate this issue, decentralized federated learning (DFL) has been proposed, where all participating clients engage in peer-to-peer communication without a central server. Nonetheless, DFL still suffers from training degradation as FL does due to the non-independent and identically distributed (non-IID) nature of client data. And incorporating personalization layers into DFL may be one of the most effective solutions to alleviate the side effects caused by non-IID data. Therefore, in this paper, we propose a novel unit representation aided personalized decentralized federated learning framework, named UA-PDFL, to deal with the non-IID challenge in DFL. By adaptively adjusting the level of personalization layers through the guidance of the unit representation, UA-PDFL is able to address the varying degrees of data skew. Based on this scheme, client-wise dropout and layer-wise personalization are proposed to further enhance the learning performance of DFL. Extensive experiments empirically prove the effectiveness of our proposed method.}, archive = {J_NEUCOM}, author = {Hangyu Zhu and Yuxiang Fan and Zhenping Xie}, doi = {10.1016/j.neucom.2025.131579}, journal = {Neurocomputing}, month = {12}, pages = {131579}, shortjournal = {Neurocomputing}, title = {UA-PDFL: A personalized approach for decentralized federated learning}, volume = {657}, year = {2025}, }
• (2025). A novel two-stage hybrid feature selection: Exploiting ubiquitous intrinsic feature groups. NEUCOM, 657, 131574. (www)

  @article{ , abstract = {Feature selection (FS) has emerged as an important data pre-processing technology to solve the challenging task of data mining. However, most existing FS methods primarily focus on exploiting the independent contributions provided by individual features, while neglecting the critical contributions inherent in ubiquitous intrinsic feature groups. As a result, they may fail to fully capture the potentially valuable information embedded in data. This issue becomes more pronounced in modern large-scale data environments, such as JointCloud, where cross-organizational collaborative data analysis over non-shared datasets is often required. To address this issue, this paper proposes a N ovel T wo- S tage H ybrid FS approach (NTSHFS) that jointly considers the informative contributions of both individual features and collaborative feature groups, enabling a comprehensive evaluation of feature relevance, redundancy and discriminative capability. In the first stage, the correlation coefficient-based co-association matrix (CC-CAM) is developed to ensemble the results obtained by different univariate and structured regularization techniques. Then, a CC-CAM-based embedded FS is proposed to select and rank representative features, achieving strong relevance prioritization and redundancy elimination. In the second stage, a quasi fuzzy-rough set (QFRS) model is designed by integrating the similarity relations at both individual-feature level and multi-feature level through the intersection operation. Based on this model, a QFRS-based filter FS is presented to determine the final feature subset with stronger discriminative capability using internal rankings of feature groups. Experimental results on 24 datasets demonstrate that the proposed approach typically outperforms the compared methods (i.e., achieving an average classification accuracy improvement ranging from 1.77 % to 7.69 %), highlighting its effectiveness, robustness and generalization in data mining.}, archive = {J_NEUCOM}, author = {Lin Qiu and Xingwei Wang and Bo Yi and Yanpeng Qu and Min Huang and Kaimin Zhang}, doi = {10.1016/j.neucom.2025.131574}, journal = {Neurocomputing}, month = {12}, pages = {131574}, shortjournal = {Neurocomputing}, title = {A novel two-stage hybrid feature selection: Exploiting ubiquitous intrinsic feature groups}, volume = {657}, year = {2025}, }
• (2025). Multi-view semi-supervised feature selection with multi-order similarity and tensor learning. NEUCOM, 657, 131573. (www)

  @article{ , abstract = {Multi-view data has attracted extensive attention because it can better characterize samples, and multi-view semi-supervised feature selection can not only effectively improve multi-view performance, but also maintain the original real structure of the data. To this end, many scholars have proposed various models to achieve this goal. However, most of the existing methods rely on the graph structure constructed from the original data and use the constructed graph as a guide for feature selection. This not only ignores multi-order domain knowledge, but also ignores the high-order relations between views. Therefore, this study effectively integrates multi-order domain information with graph learning, and performs tensor low-rank learning on the graph structure between multiple views. A multi-view semi-supervised feature selection method based on multi-order similarity and tensor learning is proposed, which not only integrates multi-order domain information, but also takes into account the relationship between views. Based on this, we propose an iterative method to solve the objective function and prove the superiority of our method on multiple basic datasets.}, archive = {J_NEUCOM}, author = {Hangyu Chen and Xijiong Xie and Yujie Xiong}, doi = {10.1016/j.neucom.2025.131573}, journal = {Neurocomputing}, month = {12}, pages = {131573}, shortjournal = {Neurocomputing}, title = {Multi-view semi-supervised feature selection with multi-order similarity and tensor learning}, volume = {657}, year = {2025}, }
• (2025). Feature selection via risk-bound utility maximization. NEUCOM, 657, 131572. (www)

  @article{ , abstract = {The ultimate goal of supervised feature selection is to identify a feature subset that minimizes classification risk. Contemporary methods, however, often rely on heuristic or model-dependent proxy criteria that lack a direct theoretical connection to this fundamental objective. To bridge this gap, we introduce a new feature selection framework that directly optimizes a model-agnostic utility function grounded in statistical learning theory. Our approach defines the utility of a feature subset based on the 1-Wasserstein distance between class-conditional distributions. This metric is theoretically powerful as it can be used to construct an upper bound on the Bayes classification error, allowing us to construct a utility function that is a direct surrogate for this risk bound. We instantiate this framework with a subset search strategy that effectively captures feature interactions by maximizing this risk-bound utility. Extensive experiments on real-world datasets demonstrate that our method not only achieves state-of-the-art classification performance but also demonstrates superior robustness and interpretability, providing a principled and powerful alternative to traditional feature selection methods, confirming our framework’s theoretical soundness.}, archive = {J_NEUCOM}, author = {Chunxu Cao and Qiang Zhang}, doi = {10.1016/j.neucom.2025.131572}, journal = {Neurocomputing}, month = {12}, pages = {131572}, shortjournal = {Neurocomputing}, title = {Feature selection via risk-bound utility maximization}, volume = {657}, year = {2025}, }
• (2025). Projected variable three-term conjugate gradient algorithm for enhancing generalization performance in deep neural network training. NEUCOM, 657, 131568. (www)

  @article{ , abstract = {Deep learning optimization faces a fundamental trade-off between convergence efficiency and generalization. First-order methods such as stochastic gradient descent (SGD) and adaptive moment estimation (Adam) tend to find flatter minima but converge slowly, while higher-order methods converge rapidly but are often drawn to sharp minima that generalize poorly. To address this, we introduce the projected variable three-term conjugate gradient (PVTTCG) algorithm. Motivated by the geometric instabilities in modern networks that use techniques such as batch normalization (BN), PVTTCG integrates an orthogonal projection into the higher-order optimization framework. This mechanism eliminates radial components from the search direction, inherently guiding the optimization toward flatter regions without requiring additional regularization terms or hyperparameters. The effectiveness of PVTTCG is validated across diverse tasks, including language modeling, large-scale image classification, and a real-world engineering application. In complex scenarios, PVTTCG consistently improves upon its higher-order baseline, achieving up to a 3.92 percentage point gain on CIFAR-100 while remaining competitive with leading first-order methods. A systematic analysis reveals that PVTTCG demonstrates superior robustness to batch size variations, particularly excelling at larger batch sizes. This robustness enables the algorithm to process batch sizes up to 2,048 in engineering applications, achieving a 35.9% test loss reduction compared to Adam. These findings establish PVTTCG as an effective solution for bridging the convergence-generalization trade-off.}, archive = {J_NEUCOM}, author = {Sanghyuk Kim and Hansu Kim and Namwoo Kang and Tae Hee Lee}, doi = {10.1016/j.neucom.2025.131568}, journal = {Neurocomputing}, month = {12}, pages = {131568}, shortjournal = {Neurocomputing}, title = {Projected variable three-term conjugate gradient algorithm for enhancing generalization performance in deep neural network training}, volume = {657}, year = {2025}, }
• (2025). DVC2: Deep video cascade clustering from video structures. NEUCOM, 657, 131565. (www)

  @article{ , abstract = {Video clustering is a critical unsupervised learning task, where category labels are unavailable, unlike in supervised video classification. The primary challenge is learning meaningful video representations without annotations to effectively group similar videos. Most existing methods extract frame-level features and apply standard clustering algorithms such as K-means, but they often fail to capture temporal relationships inherent in video data. In this paper, we introduce Deep Video Cascade Clustering ( DVC 2 ), a novel unsupervised video learning paradigm. Unlike image-based clustering methods, DVC 2 first learns an initial video representation through frame clustering, which serves as guidance, and then aligns video clustering results with both long-term and short-term structures as well as nearest neighbors. We evaluate DVC 2 on benchmark datasets, including UCF101 and Kinetics-400, achieving state-of-the-art results. Notably, even in annotation-free scenarios where self-supervised learning with K-means already yields reasonable clustering, DVC 2 demonstrates significantly superior performance.}, archive = {J_NEUCOM}, author = {Zihua Wang and Siya Mi and Yu Zhang}, doi = {10.1016/j.neucom.2025.131565}, journal = {Neurocomputing}, month = {12}, pages = {131565}, shortjournal = {Neurocomputing}, title = {DVC2: Deep video cascade clustering from video structures}, volume = {657}, year = {2025}, }
• (2025). SiFH: Siamese frequency harmonization self-supervised learning for motion forecasting. NEUCOM, 657, 131563. (www)

  @article{ , abstract = {Motion forecasting is a fundamental component of the autonomous driving system and plays an important role in ensuring safety. While supervised learning methods have achieved promising performance in many domains, their model capacity is typically limited by the availability of annotated data. Some previous works have tried to introduce this paradigm into motion forecasting. However, these early studies neglect the task-specific characteristics. To address this, we incorporate two key insights into motion forecasting tasks within the self-supervised paradigm and propose a novel self-supervised motion forecasting framework. First, we design a Frequency Information Harmonization pretext task that explicitly encourages the model to integrate frequency domain features with their time domain counterparts, making them work harmoniously. Second, we introduce an Implicit Scene Alignment task, which enables the model to learn scene-level semantics by aligning masked and unmasked views through shared prototypes. By jointly optimizing these objectives, the model is encouraged to leverage abundant unlabeled data and capture rich spatio-temporal representations. Extensive experiments conducted on the challenging Argoverse 2 and Argoverse 1 benchmarks demonstrate that our proposed model outperforms previous state-of-the-art baselines and can produce more accurate and reliable predictions.}, archive = {J_NEUCOM}, author = {Chunyu Liu and Zeyu Liu and Tiechui Yao and Shijie Li}, doi = {10.1016/j.neucom.2025.131563}, journal = {Neurocomputing}, month = {12}, pages = {131563}, shortjournal = {Neurocomputing}, title = {SiFH: Siamese frequency harmonization self-supervised learning for motion forecasting}, volume = {657}, year = {2025}, }
• (2025). RosMS-ECDBTM: Efficient channel attention enabled distributed deep learning model for mental state detection using EEG signal. NEUCOM, 657, 131559. (www)

  @article{ , abstract = {Currently, mental stress is emerging as a common social problem triggering different health disorders, including nervousness, heart attacks, strokes, and depression. Specifically, the Electroencephalography (EEG) signal, capable of reflecting the variations in brain activity, is highly used for mental state detection. Despite their promising performance, the existing EEG-based detection methods fail to capture the inherent characteristics of highly intricate and nonstationary EEG signals. In order to address the drawbacks of existing methods, this research proposes the Deep Learning model namely, Rosmarus Migrative Search Optimized Efficient channel attention enabled Distributed Bi-directional Long Short-Term Memory (RosMS-ECDBTM) model for precise mental state detection. More specifically, the efficient channel attention facilitates the proposed model to dynamically highlight the important parts of the signal characteristics while suppressing the irrelevant regions. Besides, the distributed DL architecture improves the learning capability and scalability of the proposed model to process the large datasets, through the parallel processing of sequential data. Further, the proposed approach exploits the Rosmarus Migrative Search Optimization (RosMS) algorithm for optimizing the RosMS-ECDBTM architecture, resulting in improving the training process. Ultimately, the proposed model, combining efficient channel attention and distributed learning mechanism, captures the intricate patterns and reduces the computational complexity of mental state detection. In addition, the Hybrid Discrete Wavelet transform (DWT) approach decomposes the EEG signal into several frequency components for capturing the hidden patterns and anomalous states in the EEG signal, providing key insights for improving the mental state detection. Extensive experiments show that the RosMS-ECDBTM method provides superior performance, achieving a high accuracy of 96.78 %, precision of 98.99 %, recall of 95.91 %, and F1-score of 97.42 % for the Mental Stress Detection dataset compared to other state-of-the-art methods. Ultimately, these findings reveal the high learning efficiency of the proposed deep learning approach in enhancing the mental state detection accuracy, significantly contributing to advancing the field of mental health monitoring.}, archive = {J_NEUCOM}, author = {Mandar Nitin Kakade}, doi = {10.1016/j.neucom.2025.131559}, journal = {Neurocomputing}, month = {12}, pages = {131559}, shortjournal = {Neurocomputing}, title = {RosMS-ECDBTM: Efficient channel attention enabled distributed deep learning model for mental state detection using EEG signal}, volume = {657}, year = {2025}, }
• (2025). Efficient semantic segmentation of remote sensing images through dynamic feature enhancement and multimodal alignment fusion. NEUCOM, 657, 131555. (www)

  @article{ , abstract = {Multimodal fusion has shown promising applications in integrating information from different modalities. However, existing multimodal fusion approaches in remote sensing face two main challenges: First, multimodal fusion models relying on Convolutional Neural Networks (CNNs) or Visual Transformers (ViTs) have limitations in terms of remote modeling capabilities and computational complexity, while state-space model (SSM)-based fusion models are prone to feature redundancy due to the use of multiple scanning paths, and similarly suffer from high computational complexity. Second, existing methods do not fully address inter-modal heterogeneity, leading to poor multimodal data fusion. To address these issues, we propose an efficient multimodal fusion network, AFMamba, based on the state-space model (SSM) for semantic segmentation of remote sensing images. Specifically, we design the Efficient Dynamic Visual State Space (EDVSS) module, which enhances the efficiency of the standard Mamba model by dynamically improving local features and reducing channel redundancy. Furthermore, we introduce the Cross Attention Alignment Fusion (CAAFM) module, which combines cross-image attention fusion and channel interaction alignment to effectively improve the accuracy and efficiency of cross-modal feature fusion and mitigate feature inconsistency. Experimental results demonstrate that in multimodal hyperspectral image semantic segmentation, the proposed model reduces computational complexity, measured in GFLOPs, by at least 61 % while maintaining a low parameter count, achieving optimal overall accuracy (OA) of around 92 %, and effectively balancing performance and computational efficiency.}, archive = {J_NEUCOM}, author = {Wenqian Chen and Wendie Yue and Kai Chang and Hongzhi Wang and Kaijun Tan and Xinyu Liu and Xiaoyi Cao}, doi = {10.1016/j.neucom.2025.131555}, journal = {Neurocomputing}, month = {12}, pages = {131555}, shortjournal = {Neurocomputing}, title = {Efficient semantic segmentation of remote sensing images through dynamic feature enhancement and multimodal alignment fusion}, volume = {657}, year = {2025}, }
• (2025). A multi-step minimax Q-learning algorithm for two-player zero-sum markov games. NEUCOM, 657, 131552. (www)

  @article{ , abstract = {An interesting iterative procedure is proposed to solve two-player zero-sum Markov games. Under suitable assumptions, the boundedness of the proposed iterates is obtained theoretically. Using results from stochastic approximation, the almost sure convergence of the proposed multi-step minimax Q-learning is obtained theoretically. More specifically, the proposed algorithm converges to the game theoretic optimal value with probability one, when the model information is not known. Numerical simulations authenticate that the proposed algorithm is effective and easy to implement.}, archive = {J_NEUCOM}, author = {Shreyas S.R. and Antony Vijesh}, doi = {10.1016/j.neucom.2025.131552}, journal = {Neurocomputing}, month = {12}, pages = {131552}, shortjournal = {Neurocomputing}, title = {A multi-step minimax Q-learning algorithm for two-player zero-sum markov games}, volume = {657}, year = {2025}, }
• (2025). Constrained multi-agent evasion using deep reinforcement learning. NEUCOM, 657, 131550. (www)

  @article{ , abstract = {Designing effective evasion strategies in pursuit–evasion scenarios is challenging, particularly when the pursuer’s model is unknown and inaccessible. This limitation hinders the application of conventional evasion policy design methods. To overcome this challenge, especially when evaders have constrained maneuverability against unrestricted pursuers, we propose a novel multi-agent evasion algorithm based on deep reinforcement learning. Our approach employs a staged learning framework, progressively guiding evaders from simpler to more complex tasks to refine their evasion strategies. Crucially, our algorithm enables evaders to infer pursuers’ intentions even without prior knowledge of pursuers’ objectives, allowing for optimal decision-making despite mobility constraints. Simulation results demonstrate that our method significantly enhances evasion success, validating the effectiveness of learning-based strategies. Additionally, the algorithm exhibits strong adaptability to environmental changes, ensuring reliable performance across diverse pursuit–evasion scenarios.}, archive = {J_NEUCOM}, author = {Bowei Yan and Runle Du and Xiaojun Ban and Di Zhou}, doi = {10.1016/j.neucom.2025.131550}, journal = {Neurocomputing}, month = {12}, pages = {131550}, shortjournal = {Neurocomputing}, title = {Constrained multi-agent evasion using deep reinforcement learning}, volume = {657}, year = {2025}, }
• (2025). STD-explain: Generalizing explanations for spatio-temporal graph convolutional networks based on spatio-temporal decoupled perturbation. NEUCOM, 657, 131539. (www)

  @article{ , abstract = {Spatio-temporal graph convolutional networks utilize an alternating combination of one-dimensional ordinary convolution and graph convolution to extract spatio-temporal features. This alternation intertwines temporal and spatial features closely, leading to a tight coupling between them. The presence of spatio-temporal coupling complicates the analysis of spatio-temporal data, posing challenges for existing explainability algorithms to effectively separate and interpret these intertwined features. Therefore, we propose STD-Explain, an explainable algorithm based on spatio-temporal decoupled perturbation, which employs a two-stage perturbation approach considering subgraph and node-level explanations. Firstly, targeting the spatio-temporal coupling issue in spatio-temporal graph convolutional networks, the algorithm proposes a temporal perturbation algorithm based on Slice Graph and a spatial perturbation algorithm aimed at important subgraph node features. Secondly, to avoid introducing additional semantic information when extracting temporal subgraphs, we propose a method for generating temporal subgraphs in spatio-temporal decoupling, slicing human skeleton sequences with discrete masks to ensure each subsequence maintains spatial structure integrity without introducing additional edges. Furthermore, to ensure the maximum correlation between the interpreted subgraphs and model predictions, we propose a temporal important subgraph discrimination strategy to select the most relevant subgraphs to model predictions. Experimental results demonstrate that STD-Explain performs well in qualitative and quantitative analysis.}, archive = {J_NEUCOM}, author = {Yanshan Li and Ting Shi and Suixuan He and Zhiyuan Chen and Li Zhang and Rui Yu and Weixin Xie}, doi = {10.1016/j.neucom.2025.131539}, journal = {Neurocomputing}, month = {12}, pages = {131539}, shortjournal = {Neurocomputing}, title = {STD-explain: Generalizing explanations for spatio-temporal graph convolutional networks based on spatio-temporal decoupled perturbation}, volume = {657}, year = {2025}, }
• (2025). MGSA: Enhancing transferability of multimodal adversarial attacks via multi-granularity semantic alignment. NEUCOM, 657, 131534. (www)

  @article{ , abstract = {Vision–language pretraining (VLP) models have demonstrated exceptional performance across a wide range of image–text multimodal tasks. Despite their prominence, research confirms that these systems retain significant susceptibility to adversarial manipulation. Existing multimodal adversarial attack methods often fail to fully exploit sample-specific semantic structures, resulting in suboptimal cross-modal alignment and limited transferability of adversarial examples. To overcome this limitation, we propose MGSA—a Multi-Granularity Semantic Alignment Attack framework that enhances adversarial perturbation transferability by jointly disrupting cross-modal semantics at both global and fine-grained levels. MGSA captures coarse-grained alignment using overall representations and fine-grained correspondence by selectively aggregating key image regions and words based on importance. This dual-level joint optimization effectively perturbs both holistic consistency and detailed correspondences, thereby significantly enhancing attack effectiveness in white-box scenarios and transferability to black-box models. Extensive experiments conducted across diverse model architectures and multimodal tasks demonstrate that our method achieves strong performance in white-box settings while significantly improving black-box attack success rates. The results highlight the vulnerability of current VLP models and the effectiveness of our approach in generating transferable and semantically grounded adversarial examples.}, archive = {J_NEUCOM}, author = {Xinyu Liu and Haohua Zhou and Zhidong Shen and Hui Sun}, doi = {10.1016/j.neucom.2025.131534}, journal = {Neurocomputing}, month = {12}, pages = {131534}, shortjournal = {Neurocomputing}, title = {MGSA: Enhancing transferability of multimodal adversarial attacks via multi-granularity semantic alignment}, volume = {657}, year = {2025}, }
• (2025). Urban multi-domain mixing (UMDMix) based unsupervised domain adaptation for LiDAR semantic segmentation. NEUCOM, 657, 131526. (www)

  @article{ , abstract = {3D semantic maps generated from Light Detection and Ranging (LiDAR) point clouds enable scene understanding in diverse applications such as autonomous driving and urban planning. However, existing deep learning models struggle when tested on different domains, worsened by limited labeled data. Unsupervised Domain Adaptation (UDA) can bridge this gap, but existing UDA methods often face adaptation challenges due to domain shifts arising from variations in the physical environment, data sparsity, and sensor differences. To address these limitations, we propose UMDMix , a novel UDA architecture that operates on the mixing of multiple labeled source domains with unlabeled target domains to make the predictive model robust to cross-domain variations. UMDMix integrates a teacher–student learning scheme to produce a robust teacher model and an adaptable student model. The performance of the teacher model in the source domain is further strengthened by a position-aware loss that assigns greater significance to semantically rich neighborhoods. A combination of entropy regularization and KL-divergence loss in the target domain updates the knowledge of the teacher model to the student model during adaptation. Our extensive experiments across diverse environments show that UMDMix achieves an average improvement of 13 % on minor classes such as bicycle, traffic sign, and person in target domain datasets, outperforming previous State-Of-The-Art (SOTA) UDA methods.}, archive = {J_NEUCOM}, author = {Anurag Nihal and Pyare Lal and Vaibhav Kumar}, doi = {10.1016/j.neucom.2025.131526}, journal = {Neurocomputing}, month = {12}, pages = {131526}, shortjournal = {Neurocomputing}, title = {Urban multi-domain mixing (UMDMix) based unsupervised domain adaptation for LiDAR semantic segmentation}, volume = {657}, year = {2025}, }
• (2025). Neural network-enhanced asymmetric prescribed performance control for multi-task cooperative navigation of USVs via an adaptive formation structure. NEUCOM, 657, 131521. (www)

  @article{ , abstract = {This paper investigates a prescribed performance control algorithm with asymmetric boundary for Unmanned Surface Vehicle (USV) formation to achieve cooperative navigation under marine disturbances. The proposed algorithm consists of a guidance switching mechanism and a robust adaptive control method. In the improved guidance principle, a velocity correction rule is provided to generate accurate reference signals for USVs during path following. Combined with the guidance term, the communication load between controller and actuator is significantly reduced by employing the Dynamic Event-Triggered Mechanism (DETM) with adaptive updating of threshold parameters. By integrating the initial errors into the performance boundary function, the controller can effectively adapt to different initial states. In addition, due to the smooth property of the shifting function, the error oscillation of the system before steady state is effectively suppressed. The Radial Basis Function Neural Networks (RBF-NNs) are utilized to design damping terms, enhancing the anti-interference capability in marine environments and mitigating the effects of nonlinearities in the model. Through the Lyapunov theorem, the Semi-Globally Uniformly Ultimately Bounded (SGUUB) stability of all state variables is guaranteed. Finally, quantitative validation of the algorithm is performed through numerical simulations and comparative analysis. The results demonstrate a control accuracy within 0.5 meters while showing that, compared to Static Event-Triggering Mechanisms (SETM), the DETM reduces control update frequency for surge force and yawing moment by 12.32 % and 18.78 %, respectively.}, archive = {J_NEUCOM}, author = {Guoqing Zhang and Junji Feng and Shilin Yin and Matthew Montebello}, doi = {10.1016/j.neucom.2025.131521}, journal = {Neurocomputing}, month = {12}, pages = {131521}, shortjournal = {Neurocomputing}, title = {Neural network-enhanced asymmetric prescribed performance control for multi-task cooperative navigation of USVs via an adaptive formation structure}, volume = {657}, year = {2025}, }
• (2025). Mine-SSD: Dual-threshold set abstraction and radius-adaptive grouping for 3D object detection in open-pit mines. NEUCOM, 657, 131507. (www)

  @article{ , abstract = {Accurate and efficient 3D object detection is essential for ensuring both safety and operational efficiency in open-pit mines. Due to complex scene structures, broad perception ranges, and significant object size variations, existing point-based 3D detection methods face challenges that limit their applicability in open-pit mines. To address these issues, Mine-SSD, a single-stage 3D object detection method, is developed with dual-threshold set abstraction (DT-SA) and a radius-adaptive grouping mechanism. Specifically, a dual-head self-correlation module is introduced to calculate comprehensive importance scores for each point, enhancing the model’s ability to prioritize key features. Using these importance scores, a dual-threshold self-correlative farthest point sampling (DTSC-FPS) method is applied to retain key non-local information points during downsampling in set abstraction (SA). Additionally, a radius-adaptive grouping mechanism is designed to dynamically adjust the candidate point aggregation radius, capturing critical features of unconventional objects and supporting multi-scale feature processing. Finally, a novel regression loss function is constructed to improve prediction accuracy and balanced performance across objects of different sizes, ensuring reliable performance of Mine-SSD in multi-scale detection. Extensive experiments on an open-pit mine dataset validate the effectiveness of Mine-SSD.}, archive = {J_NEUCOM}, author = {Zhongyu Xie and Yuqian Zhao and Fan Zhang and Biao Luo and Wenliu Hu and Tenghai Qiu}, doi = {10.1016/j.neucom.2025.131507}, journal = {Neurocomputing}, month = {12}, pages = {131507}, shortjournal = {Neurocomputing}, title = {Mine-SSD: Dual-threshold set abstraction and radius-adaptive grouping for 3D object detection in open-pit mines}, volume = {657}, year = {2025}, }
• (2025). SRDA: Self-regulating distribution alignment based on prompt learning for unsupervised domain adaptation. NEUCOM, 657, 131498. (www)

  @article{ , abstract = {Although large-scale pre-trained vision–language models (VLMs) exhibit significant potential for cross-domain visual tasks, existing prompt-learning-based unsupervised domain adaptation (UDA) methods suffer from source domain overfitting and target domain performance degradation. This paper experimentally demonstrates that conventional prompt learning exhibits insufficient cross-domain generalization due to optimization being heavily biased toward the source distribution. To address this challenge, we propose a Self-regulating Distribution Alignment (SRDA) framework. Its core innovation is a dual-branch collaborative optimization mechanism that dynamically balances cross-domain semantic alignment with pre-trained knowledge preservation. Specifically, the self-regulating multimodal prompt branch incorporates three constraints: semantic consistency regularization, dual-domain collaborative contrastive regularization, and text semantic diversity enhancement. This design suppresses prompt overfitting to the source domain while preserving CLIP’s zero-shot generalization capability. The cross-domain alignment branch introduces dynamic dual-domain feature bank and Cross-domain Collaborative Dual Attention module, achieving fine-grained local semantic calibration through moving average prototypes and a dual-layer attention mechanism. Extensive experiments validate SRDA’s effectiveness on downstream UDA tasks. The code is available at https://github.com/QYw12/SRDA .}, archive = {J_NEUCOM}, author = {Yang Qu and Jinlong Shi and Yun Cui and Ao Zhang and Suqin Bai and Ye Lu}, doi = {10.1016/j.neucom.2025.131498}, journal = {Neurocomputing}, month = {12}, pages = {131498}, shortjournal = {Neurocomputing}, title = {SRDA: Self-regulating distribution alignment based on prompt learning for unsupervised domain adaptation}, volume = {657}, year = {2025}, }
• (2025). Fault-tolerant synchronization of drive-response memristive competitive neural networks with multiple actuator failures. NEUCOM, 657, 131489. (www)

  @article{ , abstract = {In this article, the synchronization of drive-response memristive competitive neural networks (MCNNs) under multiple actuator failures is studied through implementing fault-tolerant control scheme. Unlike previous studies, the actuator failures considered in this paper include both bias and effectiveness failures. To address these challenges, a proper mathematical model is first established to capture the impact of actuator failures on control inputs. Subsequently, several sufficient conditions are deduced by designing an appropriate bilayer fault-tolerant controller and constructing a Lyapunov functional to achieve the global exponential synchronization, finite-time synchronization, fixed-time synchronization and predefined-time synchronization respectively. Additionally, the settling time upper bounds for the proposed synchronization methods are determined. In the end, numerical simulations with analysis and comparison are performed to confirm the validity of the proposed results.}, archive = {J_NEUCOM}, author = {Wenjing Duan and Yanli Huang and Quang Dan Le and Tse Chiu Wong}, doi = {10.1016/j.neucom.2025.131489}, journal = {Neurocomputing}, month = {12}, pages = {131489}, shortjournal = {Neurocomputing}, title = {Fault-tolerant synchronization of drive-response memristive competitive neural networks with multiple actuator failures}, volume = {657}, year = {2025}, }
• (2025). Double-boundary awareness of shared categories for source-free universal domain adaptation. NEUCOM, 657, 131473. (www)

  @article{ , abstract = {Source-Free Universal Domain Adaptation (SF-UniDA) aims to adapt a pre-trained source model to an unlabeled target domain without access to source data or prior knowledge of cross-domain category shifts. Existing methods focus on distinguishing target-private unknown samples and assigning pseudo-labels to known samples across the entire label space, including both shared and source-private categories as pseudo-labels, for self-training. However, this data-aware pseudo-labeling approach could mistakenly assign known samples to either source-private or target-private categories, making it sensitive to category shifts and potentially introducing errors or mislabeling. In this paper, we propose Double-boundary Awareness Domain Adaptation (DADA), a class-aware framework that partitions the target domain pseudo-label space into shared, potential source-private, and target-private categories. By labeling target-private samples as unknown and filtering out misassigned source-private samples, DADA enhances the quality of target samples and the reliability of pseudo-labels. To achieve this, we introduce Double-bounded Shared Categories Refinement (DSCR) module, which refines shared classes by identifying both source- and target-private categories based on prior class probabilities and the entropy distribution. Additionally, we incorporate Class-Aware Discriminative Learning (CADL) to enhance discrimination between shared and target-private samples across domains. Experiments on four benchmarks demonstrate the effectiveness of DADA, with overall H-score gains of 8.2 % in the OPDA scenario on Digit dataset and accuracy gains of 10.6 % in the PDA scenario on VisDA dataset. Code is available at: https://github.com/W2Wzj/DADA .}, archive = {J_NEUCOM}, author = {Zhijing Wang and Ji Guo and Xu Sun and Yi Luo and Aiguo Chen}, doi = {10.1016/j.neucom.2025.131473}, journal = {Neurocomputing}, month = {12}, pages = {131473}, shortjournal = {Neurocomputing}, title = {Double-boundary awareness of shared categories for source-free universal domain adaptation}, volume = {657}, year = {2025}, }
• (2025). SGN: Stochastic guidance network for sim-to-real generalization. NEUCOM, 657, 131468. (www)

  @article{ , abstract = {The significant domain differences between synthetic data and real data are a challenging problem for current domain generalized segmentation networks. Therefore, this paper proposes a stochastic guidance network (SGN) for sim-to-real generalization that includes the category reweighting strategy, Multi-scale Feature Fusion Guidance (MSFFG) module and multiple style perturbation modules, which improves the issues caused by the imbalance of the source domain’s sample categories as well as large domain gap. Experimental results show that our SGN can effectively enhance the model’s generalization ability to unseen data. In terms of mean intersection over union (mIoU) metric, compared with SOTA, the SGN improves by 3.86 % and 2.05 % respectively on two real scene-enhanced datasets(Rain_Cityscapes, Foggy_Cityscapes), and an average improvement of 1.48 % on four conventional datasets (BDD100k, Cityscapes, Mapillary, Synthia). Our project can be found at https://githubcom/leo-lab-511/SGN.}, archive = {J_NEUCOM}, author = {Yao Li and Jinlong Shi and Yun Cui and Dan Xu and Wei Teng and Yan Jiang}, doi = {10.1016/j.neucom.2025.131468}, journal = {Neurocomputing}, month = {12}, pages = {131468}, shortjournal = {Neurocomputing}, title = {SGN: Stochastic guidance network for sim-to-real generalization}, volume = {657}, year = {2025}, }
• (2025). AdvOT: Oblivious transfer based on generative adversarial networks against multiple attackers. NEUCOM, 657, 131449. (www)

  @article{ , abstract = {Oblivious Transfer (OT) is crucial in various security protocols, as it serves as a privacy-preserving and secure communication protocol. However, traditional OT protocols often necessitate complex encryption algorithms and involve intricate steps. Given the rapid advancements in artificial intelligence, it is imperative to explore the potential of artificial neural networks in simplifying OT protocols while still meeting stringent security and privacy requirements. To this need, we introduce the Adv ersarial O blivious T ransfer (AdvOT) protocol, which integrates OT with the adversarial learning mechanism of Generative Adversarial Network (GAN). Our approach involves training a neural network model to learn encryption techniques through end-to-end adversarial training, thereby eliminating the reliance on specific algorithms. The AdvOT protocol comprises two phases. Firstly, a Random Oblivious Transfer (ROT) protocol is employed to generate and distribute keys based on the CKKS homomorphic encryption algorithm. Subsequently, neural networks are introduced to replace specific symmetric encryption algorithms and encrypt the messages to be transferred. These neural networks undergo training using the adversarial learning mechanism to develop a symmetric encryption algorithm. Furthermore, to enhance the model, attack networks with varying capabilities are created, resulting in a more secure encryption algorithm capable of withstanding multiple attackers. Experimental results demonstrate that the execution speed of the CKKS-based ROT algorithm is significantly faster compared to the BFV and Paillier algorithms. Moreover, in adversarial network models with multiple attackers, the decryption accuracy for the recipient approaches 100 %, while the accuracy or classification error rate for attackers is approximately 50 %. These findings indicate that the proposed method effectively safeguarded communication between parties from interception.}, archive = {J_NEUCOM}, author = {Yuke Wang and Zhentian Zhong and Ninghao Liu and Xiaohui Li and Junfeng Wang}, doi = {10.1016/j.neucom.2025.131449}, journal = {Neurocomputing}, month = {12}, pages = {131449}, shortjournal = {Neurocomputing}, title = {AdvOT: Oblivious transfer based on generative adversarial networks against multiple attackers}, volume = {657}, year = {2025}, }
• (2025). Challenges and advancements in modeling shock fronts with physics-informed neural networks: A review and benchmarking study. NEUCOM, 657, 131440. (www)

  @article{ , abstract = {Solving partial differential equations (PDEs) with discontinuous solutions—such as shock waves in multiphase viscous flow in porous media—is critical for a wide range of scientific and engineering applications, as they represent sudden changes in physical quantities. Physics-Informed Neural Networks (PINNs), an approach proposed for solving PDEs, encounter significant challenges when applied to such systems. Accurately solving PDEs with discontinuities using PINNs requires specialized techniques to ensure effective solution accuracy and numerical stability. Various methods have been developed to address the challenges of modeling discontinuities within the PINNs framework. This work reviews and benchmarks these approaches across problems of varying complexity, categorizing them into three broad groups, influencing solution accuracy differently. (1) Physics-modification (PM) methods improve accuracy by modifying the system’s physics, such as adding artificial viscosity or enforcing entropy constraints. (2) Loss and training modification (LM) techniques focus on regularizing the loss landscape, often by refining the loss term in high-error regions. (3) Architecture-modification (AM) approaches, on the other hand, propose advanced network designs to handle discontinuities better. A benchmarking study was conducted on two multiphase flow problems in porous media: the classic Buckley-Leverett (BL) problem and a fully coupled system of equations involving shock waves but with varying levels of solution complexity. The findings show that PM and LM approaches can provide accurate solutions for the BL problem by effectively addressing the infinite gradients associated with shock occurrences. In contrast, AM methods failed to effectively resolve the shock waves. When applied to fully coupled PDEs (with more complex loss landscapes), the generalization error in the solutions quickly increased, highlighting the need for ongoing innovation. This study provides a comprehensive review of existing techniques for managing PDE discontinuities using PINNs, offering information on their strengths and limitations. The results underscore the necessity for further research to improve PINNs’ ability to handle complex discontinuities, particularly in more challenging problems with complex loss landscapes. This includes problems involving higher dimensions or multiphysics systems, where current methods often struggle to maintain accuracy and efficiency.}, archive = {J_NEUCOM}, author = {Jassem Abbasi and Ameya D. Jagtap and Ben Moseley and Aksel Hiorth and Pål Østebø Andersen}, doi = {10.1016/j.neucom.2025.131440}, journal = {Neurocomputing}, month = {12}, pages = {131440}, shortjournal = {Neurocomputing}, title = {Challenges and advancements in modeling shock fronts with physics-informed neural networks: A review and benchmarking study}, volume = {657}, year = {2025}, }
• (2025). SparseSCIGaussian: Sparse snapshot compressed images 3D gaussian splatting. NEUCOM, 657, 131422. (www)

  @article{ , abstract = {In this paper, we propose SparseSCIGaussian, a novel method for achieving high-quality novel view synthesis under sparse input conditions. Previous methods often rely heavily on depth or neural priors, which can lead to generalization challenges and significant quality degradation on complex datasets. These limitations arise primarily from the insufficient scene information available in sparse regular images. To overcome these issues, our approach utilizes images captured through Snapshot Compressive Imaging (SCI) as input. SCI-captured images inherently encode richer scene information compared to regular images, thereby substantially improving the quality of novel view synthesis under sparse input conditions. Moreover, SCI images can be conveniently captured using a software-implemented encoder, making them as accessible as traditional images. Experimental results demonstrate that our method improves 2.65 dB (13.04 %) in PSNR compared to previous methods, and further exhibits the inherent advantages of using SCI images for sparse input novel view synthesis.}, archive = {J_NEUCOM}, author = {Haoyuan He and Xuan Wang and Nanning Zheng and Caigui Jiang}, doi = {10.1016/j.neucom.2025.131422}, journal = {Neurocomputing}, month = {12}, pages = {131422}, shortjournal = {Neurocomputing}, title = {SparseSCIGaussian: Sparse snapshot compressed images 3D gaussian splatting}, volume = {657}, year = {2025}, }
• (2025). RehearMixup: Improving rehearsal-based continual learning. NEUCOM, 657, 131404. (www)

  @article{ , abstract = {Neural networks often suffer from catastrophic forgetting when learning new tasks, leading to the loss of previously acquired knowledge. To address this issue, rehearsal-based methods have emerged, which involve storing a subset of data from previous tasks and accessing it during the learning of new tasks. Current rehearsal-based methods focus on selecting representative samples to store in memory. However, there is a considerable lack of exploration of how to exploit the data at hand and consider the correlation between tasks or between past and new knowledge to improve performance. Therefore, we propose a simple yet effective approach named RehearMixup that adapts the Mixup technique into rehearsal-based methods, which synthesizes new samples for learning by interpolating data from past or current tasks. Specifically, we introduce three strategies, namely Cross-Mixup , Intra-Memory-Mixup , and Intra-Current-Mixup , based on the inherent characteristics of rehearsal-based methods - involving the memory and new tasks. Through empirical evaluations under various benchmark scenarios, we compare our approach against different rehearsal-based baselines. The results demonstrate that ours, particularly Intra-Current-Mixup , improves accuracy, backward transfer, forward transfer, and enhances the model’s robustness.}, archive = {J_NEUCOM}, author = {Yan Zhang and Kaiyuan Qi and Dong Wu and Guoqiang Wu and Yilong Yin}, doi = {10.1016/j.neucom.2025.131404}, journal = {Neurocomputing}, month = {12}, pages = {131404}, shortjournal = {Neurocomputing}, title = {RehearMixup: Improving rehearsal-based continual learning}, volume = {657}, year = {2025}, }
• (2025). A zero-shot high-performance fire detection framework based on large language models. NEUCOM, 657, 131403. (www)

  @article{ , abstract = {Fire detection is crucial for minimizing economic damage and safeguarding human lives. Existing methods, including advanced AI and ML techniques, face challenges such as detecting small fires in complex environments and relying on extensive labeled data for training. This paper proposes a novel zero-shot fire detection framework leveraging large language models (LLMs) and contrastive learning-based image–text pre-training models. The framework introduces an enhanced self-attention mechanism for optimizing image embeddings, diverse prompt generation using GPT-3.5 for improved generalization, and a dynamic threshold calculation method based on statistical analysis to enhance detection accuracy and reliability. The proposed method is tested on the public FLAME dataset and a self-collected dataset. Experimental results demonstrate that the proposed method outperforms state-of-the-art models in detecting small fires within complex backgrounds, achieving better detection performance without the need for any training data. This study highlights the potential of zero-shot learning in fire detection and provides a promising solution for real-world fire detection applications.}, archive = {J_NEUCOM}, author = {Hongyang Zhao and Yi Liu and Yuhang Han and Xingdong Li and Yanan Guo and Jing Jin}, doi = {10.1016/j.neucom.2025.131403}, journal = {Neurocomputing}, month = {12}, pages = {131403}, shortjournal = {Neurocomputing}, title = {A zero-shot high-performance fire detection framework based on large language models}, volume = {657}, year = {2025}, }
• (2025). Hetero-MoE by attention: Three-plus tasks learning solver. NEUCOM, 657, 131333. (www)

  @article{ , abstract = {Multi-task learning (MTL) stands as a promising sub-field of machine learning, aiming to simultaneously tackle multiple tasks. By leveraging shared representations and structures across diverse tasks, MTL models often exhibit higher data efficiency compared to single-task models across various domains, including recommender system applications, multi-label classification and other AI applications. However, the efficacy of MTL models is sometimes hindered by the multi-causal task conflict problem. To address this challenge, existing research predominantly focuses on enhancing structural designs and the underlying optimizers. Nevertheless, these approaches often fall short in comprehensively mitigating task conflicts, especially in scenarios involving three or more tasks, such as recommender systems. When shared experts contend with excessive task-related information simultaneously, the effective filtration of potentially harmful knowledge becomes challenging. To this end, we propose a novel Heterogeneous Multi-Expert model with an attention layer, termed HMEA. HMEA introduces Heterogeneous Experts as shared experts to decompose signal connections among three or more tasks. Additionally, it integrates an attention layer to further decouple conflicts among mini-tasks within shared experts. The experiments and ablation studies on various standard and synthetic datasets illustrate the effectiveness of HMEA in alleviating the task conflict problem inherent in three-plus task learning systems.}, archive = {J_NEUCOM}, author = {Dandan Zhang and Guanqi Zeng and Haotian Wu and Hongwen Zhang and Zheng Ye and Yao Yang}, doi = {10.1016/j.neucom.2025.131333}, journal = {Neurocomputing}, month = {12}, pages = {131333}, shortjournal = {Neurocomputing}, title = {Hetero-MoE by attention: Three-plus tasks learning solver}, volume = {657}, year = {2025}, }
• (2025). Predicting nonlinear dynamic systems by causal physics-informed neural networks with ResNet blocks. NEUCOM, 656, 131589. (www)

  @article{ , abstract = {With the continuous advancement of data computational science, the prediction of nonlinear systems has provided effective support for investigating complex problems in the field of natural sciences. Physics-Informed Neural Networks (PINNs) are playing an increasingly prominent role in nonlinear system prediction. Although PINNs have been widely applied across various engineering domains, their utilization in chaotic system prediction remains notably scarce. This paper proposes a novel causal PINNs framework integrated with ResNet blocks. On the one hand, the framework incorporates temporal weighting into the residual loss, utilizing maximum temporal weight as the training termination criterion. Additionally, an annealing strategy is adopted to adaptively adjust the causal parameters, ensuring that the model adheres to physical causality constraints throughout the training process. On the other hand, the framework employs a ResNet-block-based network, which transforms identity mappings into residual mappings. This architectural design significantly enhances training stability when utilizing deep networks. To validate the performance of the proposed method, numerical experiments are conducted on the Lorenz system, Dadras system, and Kuramoto-Sivashinsky equation. The results demonstrate that the causal PINNs with ResNet blocks significantly outperform conventional PINNs in predicting chaotic systems.}, archive = {J_NEUCOM}, author = {Man-Hong Fan and Jun-Hao Zhao and Lin Ding and Xiao-Ying Ma and Rui-Lin Fu}, doi = {10.1016/j.neucom.2025.131589}, journal = {Neurocomputing}, month = {12}, pages = {131589}, shortjournal = {Neurocomputing}, title = {Predicting nonlinear dynamic systems by causal physics-informed neural networks with ResNet blocks}, volume = {656}, year = {2025}, }
• (2025). Frequency-aware fusion for improved video object segmentation. NEUCOM, 656, 131585. (www)

  @article{ , abstract = {Currently, most mainstream memory-based semi-supervised video object segmentation (VOS) methods rely on pixel-level matching to identify target objects. However, the majority of these approaches depend solely on spatial-domain features for representation, which limits their ability to preserve fine-grained details. In addition, they typically adopt a single bottom-up matching strategy, which lacks sufficient global semantic guidance, ultimately leading to suboptimal segmentation performance. To address these issues, we propose a Frequency-Aware Fusion for Improved Video Object Segmentation algorithm (FAFVOS), which incorporates frequency-domain information enhancement and a bidirectional matching mechanism to improve segmentation accuracy. First, we design a Hierarchical Frequency-Aware Encoder (HFAE), which enhances shallow features by leveraging high-frequency components to preserve edge and texture details, and strengthens deep features via low-frequency components to maintain global structural consistency, thereby achieving multi-scale frequency–spatial feature fusion. Second, a frequency-guided bidirectional matching Transformer module is proposed to establish pixel-level and object-level dual-path interactions. By incorporating a cross-attention mechanism, the model effectively facilitates joint reasoning between local pixel-wise details and global object-level semantics. Finally, a high-order moment refinement module is introduced to integrate high-order statistical features, enhancing the model’s ability to capture object deformation and leading to high-quality segmentation results. The proposed method is evaluated on the DAVIS, YouTube-VOS, and MOSE datasets. Experimental results demonstrate that, without relying on complex pretraining strategies or additional datasets, our approach achieves a real-time inference speed of 56 FPS with a J & F score of 88.5 % on the DAVIS 2017 benchmark, surpassing existing representative methods. Moreover, it also achieves consistently superior performance on the more challenging YouTube-VOS and MOSE datasets, further validating the generalization ability and robustness of the proposed approach.}, archive = {J_NEUCOM}, author = {Zhiqiang Hou and Hao Cui and Chenxu Wang and Sugang Ma and Xiaobao Yang and Lei Pu}, doi = {10.1016/j.neucom.2025.131585}, journal = {Neurocomputing}, month = {12}, pages = {131585}, shortjournal = {Neurocomputing}, title = {Frequency-aware fusion for improved video object segmentation}, volume = {656}, year = {2025}, }
• (2025). Utility-driven free tree mining in graph databases. NEUCOM, 656, 131571. (www)

  @article{ , abstract = {Frequent subgraph mining is a fundamental task in data mining, widely applied in various domains such as biological networks, social networks, and computing networks. However, existing methods for frequent subgraph mining often rely solely on support as a single metric, considering subgraphs with higher support as more important. This approach overlooks the intrinsic value of subgraphs, such as in citation networks, where users tend to associate with structures related to their own research areas, not only the frequent ones. To address this limitation, we introduce utility pattern mining into the field of subgraph mining. This mining framework considers both the internal and external values of patterns. Additionally, traditional frequent subgraph mining is hindered by isomorphism calculations, including the computational cost of subgraph isomorphism, which is NP-complete. As a connected acyclic graph, free trees play a significant role in fields such as web mining and biology. Their relatively simple structure can significantly reduce the computational cost of subgraph isomorphism calculations. In this paper, we combine utility pattern mining with frequent free tree mining, defining the problem of frequent high utility free tree mining. We design utility upper bounds that satisfy the downward closure property and propose an algorithm, UFTM (utility free tree miner), for effectively and efficiently mining utility free trees. Furthermore, we collect and test our algorithm on four real-world datasets. The results demonstrate that UFTM can discover more valuable patterns and execute the mining task efficiently.}, archive = {J_NEUCOM}, author = {Zhaoming Chen and Xinyang Chen and Guoting Chen and Wensheng Gan}, doi = {10.1016/j.neucom.2025.131571}, journal = {Neurocomputing}, month = {12}, pages = {131571}, shortjournal = {Neurocomputing}, title = {Utility-driven free tree mining in graph databases}, volume = {656}, year = {2025}, }
• (2025). Automated machine learning based on decomposition, causality and evolutionary multitask optimization for time series forecasting. NEUCOM, 656, 131569. (www)

  @article{ , abstract = {Time series forecasting plays a crucial role in various practical domains, serving as a key tool for planning and control by anticipating patterns and identifying potential future anomalies. This practice enables more effective responses to the complex dynamics of systems, typically relying on the ability to predict observations based on historical data. In this regard, the continuous pursuit of ways to enhance forecasting model accuracy is of paramount importance. Working with time series forecasting demands specific technical expertise. While automation strategies exist for these tasks, the universality of such solutions remains a challenge due to the difficulty of incorporating all available models into a single application. In this context, the present study introduces a new AutoML approach based on Decomposition, Causality, and Evolutionary Multitask Optimization for time series forecasting, called AutoDCE-TS. The AutoDCE-TS is a Multiple Input Single Output system, designed with a four-layer structure that automates the processes of feature extraction and selection, model selection and generation, and prediction. It aims to create more explainable pipelines by constructing sub-pipelines to handle each variable of the multivariate time series, as well as the components produced by the decomposition process. For each variable, a causal feature graph is constructed and used as input to an ensemble model based on regression trees. AutoDCE-TS leverages the similarity among variables to simultaneously select models and their hyperparameters for each sub-pipeline, employing a multitask evolutionary optimization strategy. Experiments conducted on 18 datasets, comparing AutoDCE-TS with 10 other methods, revealed its competitiveness, demonstrating that it is a robust and general-purpose method for time series forecasting.}, archive = {J_NEUCOM}, author = {Patrícia de Oliveira e Lucas and Frederico Gadelha Guimarães and Eduardo M.A.M. Mendes}, doi = {10.1016/j.neucom.2025.131569}, journal = {Neurocomputing}, month = {12}, pages = {131569}, shortjournal = {Neurocomputing}, title = {Automated machine learning based on decomposition, causality and evolutionary multitask optimization for time series forecasting}, volume = {656}, year = {2025}, }
• (2025). A multi-stream diffusion graph convolutional model with adaptive motion-aware attention and self-supervised pretraining for continuous sign language recognition. NEUCOM, 656, 131567. (www)

  @article{ , abstract = {Continuous Sign Language Recognition (CSLR) requires capturing both spatial and temporal dependencies to accurately model sign sequences. To enhance CSLR performance, we propose a Multi-Stream Diffusion Graph Convolution Network (MSD-GCN) from input skeleton data that introduces three key innovations. First, Adaptive Motion-aware Graph Convolution with Bi-level Attention (AMGC-BA) dynamically refines joint connectivity by leveraging semantic motion correlations, improving robustness to signer variations, and enhancing long-term dependencies. Second, signal-enhanced multi-stream representation learning integrates advanced signal processing techniques, including the Adaptive Ridgelet Transform (ART) for pose representation, Variational Mode Decomposition (VMD) for motion decomposition, and Empirical Wavelet Transform (EWT) for contextual feature extraction, ensuring feature robustness, reducing noise, and improving discriminability. Third, self-supervised pretraining leverages contrastive learning, graph reconstruction, and cross-stream feature alignment to mitigate data scarcity, enhance domain adaptation, and improve representation learning. These innovations enable the proposed graph to effectively capture complex motion patterns, distinguish between critical and redundant gestures, and generalize well across diverse signers and datasets. By improving recognition accuracy, robustness, and adaptability, the proposed approach provides a significant advancement in CSLR, addressing the challenges of signer variability, limited labeled data, and the need for fine-grained motion representation. Results on three datasets confirm the superiority of the proposed model compared to 35 comparative models. To the best of our knowledge, this is the first study in CSLR to employ such an extensive range of comparative models for performance evaluation.}, archive = {J_NEUCOM}, author = {Razieh Rastgoo}, doi = {10.1016/j.neucom.2025.131567}, journal = {Neurocomputing}, month = {12}, pages = {131567}, shortjournal = {Neurocomputing}, title = {A multi-stream diffusion graph convolutional model with adaptive motion-aware attention and self-supervised pretraining for continuous sign language recognition}, volume = {656}, year = {2025}, }
• (2025). LGFCNN: A synergistic framework integrating graph-based spatial filter and lightweight CNN for SSVEP recognition. NEUCOM, 656, 131561. (www)

  @article{ , abstract = {Optimizing the feature representation and decoding efficiency of the steady-state visual evoked potentials (SSVEPs) is critical to enhancing the performance of neural signal decoding systems. Current deep learning models often overlook the physical topological information of EEG channels, resulting in suboptimal feature extraction and limited recognition performance. To address these challenges, this study proposes a synergistically designed SSVEP recognition framework to alleviate data insufficiency, improve the feature representation, and enhance decoding efficiency. Specifically, a slicing-and-scaling technique is adopted to improve the model generalization under limited-sample scenarios. A graph-based spatial filter leverages the topological relationships among EEG channels to suppress redundant information and enhance spatial feature quality. A lightweight convolutional neural network (CNN) with fewer parameters is developed to efficiently extract discriminative temporal–spatial features for accurate SSVEP classification. Experimental results on two public benchmark datasets and one self-collected dataset demonstrate that the proposed framework outperforms baseline deep learning models, yielding improvements of at least 6.8 %, 8.5 %, and 0.5 % in peak average classification accuracy, respectively. The maximum average information transfer rates (ITRs) achieved on the three datasets were 221.4 bits/min ,106.7 bits/min , and 133.9 bits/min , respectively. By simultaneously reducing model complexity and improving decoding performance, the proposed framework offers an effective and promising approach for efficient neural signal decoding in SSVEP recognition.}, archive = {J_NEUCOM}, author = {Rui Ma and Yu Cao and Sheng Quan Xie and Mingming Zhang and Jun Li and Zhi-Qiang Zhang}, doi = {10.1016/j.neucom.2025.131561}, journal = {Neurocomputing}, month = {12}, pages = {131561}, shortjournal = {Neurocomputing}, title = {LGFCNN: A synergistic framework integrating graph-based spatial filter and lightweight CNN for SSVEP recognition}, volume = {656}, year = {2025}, }
• (2025). Double missing multi-view multi-label classification via an attention-guided multi-space consistency alignment framework. NEUCOM, 656, 131558. (www)

  @article{ , abstract = {Multi-view multi-label classification (MVMLC) seeks to enhance classification by integrating diverse data views, but its practical use is hindered by missing views and labels, posing the significant challenge of incomplete MVMLC(IMVMLC). Although various IMVMLC approaches have been proposed, most of them handle multiple objectives in a single feature space and thus overlook the conflict between learning consistent common semantics and reconstructing view-specific information. In addition, existing multi-view classification methods mainly consider utilizing the features of each view, while ignoring the inconsistent contributions of each view and usually relying on static average weighting strategies. To this end, we propose our Attention-Guided MultiSpace Consistency Alignment Framework (AMCA). In Stage 1, AMCA introduces multi-space representation learning with dual-level contrastive objectives, explicitly disentangling shared and view-specific semantics to resolve the objective conflict and yield more informative embeddings. In Stage 2, AMCA employs an attention-guided fusion module that dynamically evaluates and integrates multi-view features based on their relevance to the classification task, enabling robust decision-making even with missing data. Extensive experiments validate the effectiveness and superiority of our proposal.}, archive = {J_NEUCOM}, author = {Bingyan Nie and Wulin Xie and Lian Zhao and Jiang Long and Xiaohuan Lu and Yinghao Ye}, doi = {10.1016/j.neucom.2025.131558}, journal = {Neurocomputing}, month = {12}, pages = {131558}, shortjournal = {Neurocomputing}, title = {Double missing multi-view multi-label classification via an attention-guided multi-space consistency alignment framework}, volume = {656}, year = {2025}, }
• (2025). End-to-end transformer-based detection with density-guided query selection for small objects. NEUCOM, 656, 131554. (www)

  @article{ , abstract = {Small object detection remains a persistent challenge in transformer-based detectors due to their limited localization precision and reliance on fixed query mechanisms. In this paper, we propose Hybrid Density-Transformer (HyDeTr), a novel transformer-based object detection framework designed to improve the detection of small and densely packed objects with only a slight trade-off in inference complexity. HyDeTr introduces several key innovations: (1) a Context-Selective Hybrid Attention Encoder (CS-HAE) that distills global context from low-resolution features through efficient kernelized attention while preserving local detail via deformable attention on higher-resolution maps; (2) a Density Map Prediction module that generates a spatial prior highlighting high-object-density regions, facilitating focus on crowded scenes; (3) a Density-Guided Uncertainty-Minimal Query Selection strategy that identifies the most informative query locations based on both classification confidence and predicted density, ensuring that even low-confidence small objects in dense areas are effectively queried; and (4) an improved Query Formulation with dual embeddings, consisting of a content embedding and a 4D anchor box, refined iteratively by the decoder. Our design enables precise, density-aware query initialization and scale adaptation, leading to improved recall and accuracy for small objects. Extensive evaluations demonstrate that HyDeTr outperforms existing methods in detecting small objects, offering significant accuracy gains with only a modest increase in inference complexity, thereby maintaining near real-time performance and full end-to-end trainability.}, archive = {J_NEUCOM}, author = {Nguyen Hoanh and Tran Vu Pham}, doi = {10.1016/j.neucom.2025.131554}, journal = {Neurocomputing}, month = {12}, pages = {131554}, shortjournal = {Neurocomputing}, title = {End-to-end transformer-based detection with density-guided query selection for small objects}, volume = {656}, year = {2025}, }
• (2025). A novel dynamic sparse graph attention framework for cross-domain intelligent diagnosis of rotating machinery. NEUCOM, 656, 131553. (www)

  @article{ , abstract = {In mechanical equipment prognostics, conventional graph neural networks encounter significant limitations when processing high-dimensional dynamic sensor data: inadequate modeling of complex feature interdependencies, insufficient sensitivity to transient fault signatures, and ineffective knowledge transfer in cross-domain applications. To overcome these challenges, we present a DSGA-SAGE, which stands for Dynamic Sparse Graph Attention - SAmpling and aGgrEgation framework. Our approach presents innovations in three aspects: (1) A decentralized graph construction paradigm establishes dynamic associations among multivariate time-series features, enabling precise identification of critical fault patterns through adaptive node-edge interactions. (2) A sparse attention mechanism with trainable topology constraints optimizes the structural weights of graph in the real-time scenarios, achieving 23 % overhead computational reduction while maintaining the accuracy of feature discriminability. (3) A unified cross-domain learning strategy synchronizes multi-condition knowledge transfer through hierarchical loss optimization, ensuring robust generalization across various operational scenarios. Extensive experiments on five industrial datasets demonstrate state-of-the-art performance: achieving the highest accuracy of 96.29 % in fault diagnosis, while realizing 99.87 % Macro-F1 and 99.88 % Micro-F1 scores in cross-domain tasks. Through a comprehensive performance analysis, the superiority of the efficiency and cross-domain adaptability in dynamic sparse graph attention mechanism has been convincingly validated.}, archive = {J_NEUCOM}, author = {Ying Xie and Jixiang Wang and Zhiqiang Xu and Junnan Shen and Lijie Wen and Rongbin Xu and Yun Yang}, doi = {10.1016/j.neucom.2025.131553}, journal = {Neurocomputing}, month = {12}, pages = {131553}, shortjournal = {Neurocomputing}, title = {A novel dynamic sparse graph attention framework for cross-domain intelligent diagnosis of rotating machinery}, volume = {656}, year = {2025}, }
• (2025). Efficient medical image encryption and attack detection using hyperchaotic fibonacci polynomial convolutional neural network in IoT healthcare networks. NEUCOM, 656, 131537. (www)

  @article{ , abstract = {Due to the fast expansion of IoT technology in the health-related field, secure transfer of sensitive medical data in general, and of medical images in particular, has become a major concern. Conventional detection and encryption techniques usually cannot guarantee both high security and computational efficiency under strict real-time constraints. To counter the above problems, this research proposed a novel Hyperchaotic Fibonacci Polynomial Convolutional Neural Network with Crayfish Optimization Algorithm (HFPCNN-COA) for increased medical image security in IoT environments. The medical images are first encrypted with Hyperchaotic System-Fibonacci Q Matrix Encryption (HFQE) on a standard collection. After encryption, they are sent over the network. After transmission, the HAPCNN is applied to detect and classify any potential tampering or transmission attacks. The loss function of HAPCNN is optimized with the Crayfish Optimization Algorithm (COA) to increase the detection accuracy and convergence speed. Lastly, cryptanalysis is conducted to measure the system's resilience to attack. The experimental results prove the workability of the proposed framework, with a Peak Signal-to-Noise Ratio (PSNR) of 53 dB, which corresponds to very good retention of image quality. In addition, the proposed method is quite resistant to data manipulation with a 0.4 % Bit Error Rate (BER) and has reasonable processing times with encryption times of 2.87 ms and decryption times of 2.35 ms. This shows that HFPCNN-COA holds a great deal of promise as an actual and secure means for the transfer of medical image data for IoT-based healthcare systems.}, archive = {J_NEUCOM}, author = {Bhumireddypalli Veerasekharreddy and P. Chinniah and P. Varaprasada Rao and Krishna Prakash Arunachalam}, doi = {10.1016/j.neucom.2025.131537}, journal = {Neurocomputing}, month = {12}, pages = {131537}, shortjournal = {Neurocomputing}, title = {Efficient medical image encryption and attack detection using hyperchaotic fibonacci polynomial convolutional neural network in IoT healthcare networks}, volume = {656}, year = {2025}, }
• (2025). A collaborative decision framework for dynamic control of gear remaining useful life using multi-source information in active health management. NEUCOM, 656, 131536. (www)

  @article{ , abstract = {With the growing demand for intelligent operation and maintenance of vertical mill gearboxes in the cement industry, traditional passive maintenance methods are increasingly inadequate for supporting efficient, proactive management under complex operational conditions. In particular, sudden gear failures often result in unplanned downtime, causing significant economic losses. To address this challenge, this paper proposes a dynamic control approach for gear remaining useful life (RUL) that integrates multi-source information through collaborative decision-making to enable active health management. First, a novel RUL prediction method based on multilevel multi-source domain adaptation (MMDA) is proposed to enhance the generalization capability of the model. By minimizing discrepancies between local and global feature distributions under varying working conditions and aligning the prediction boundaries among predictors, the proposed method achieves accurate RUL predictions. Then, a gear RUL dynamic control method based on multi-information collaborative decision-making is developed. This method dynamically regulates gear RUL using a model-free adaptive control (MFAC) strategy, leveraging multi-source information such as online RUL prediction results, expected usage duration, and real-time working conditions. Finally, a collaborative decision framework for dynamic control of gear RUL is proposed, which enables active gear health management to be implemented, thereby minimizing unscheduled downtime. The effectiveness of the proposed gear RUL dynamic control method is validated on a self-made gear transmission system experimental platform, achieving a 27.6 % reduction in average RMSE compared with state-of-the-art baselines and extending the operational life of gear by approximately 61 h under dynamic control.}, archive = {J_NEUCOM}, author = {Xuegang Li and Yuanyue Pu and Nian Wu and Huajun Cao and Xiaoxi Ding and Wenbin Huang}, doi = {10.1016/j.neucom.2025.131536}, journal = {Neurocomputing}, month = {12}, pages = {131536}, shortjournal = {Neurocomputing}, title = {A collaborative decision framework for dynamic control of gear remaining useful life using multi-source information in active health management}, volume = {656}, year = {2025}, }
• (2025). AttackTracer: Semantic-level adversarial attack location traceability via evidential diffusion model. NEUCOM, 656, 131535. (www)

  @article{ , abstract = {Adversarial attacks pose a significant threat to AI systems, yet existing detection methods mainly focus on image-level threats, limiting fine-grained localization of perturbations. To address this challenge, we propose AttackTracer, the first semantic-level localization framework specifically designed for instance-level adversarial attacks. Instance-level adversarial perturbations are typically sparse and localized, which aligns naturally with the capabilities of diffusion models to progressively reconstruct sparse structures from stochastic noise. Building on this property, AttackTracer models the adversarial mask as a conditional distribution given the adversarial image, allowing iterative refinement and effective recovery of attack regions. To address the inherent instability of diffusion sampling, we introduce the Temporal Evidence Fusion Strategy (TEFS). TEFS integrates Dempster–Shafer theory with a signal-to-noise-ratio (SNR)-guided temporal ensemble, aggregating multi-step predictions to mitigate conflicts and uncertainty, thus achieving robust inference. Furthermore, adversarial perturbations often manifest as subtle high-frequency and edge distortions. To capture these, AttackTracer employs two complementary modules: the Wavelet Frequency Fusion Block (WFFB), which extracts multi-scale frequency features via Discrete Wavelet Transform to enhance sensitivity to sparse perturbations, and the Edge Feature Enhancement Module (EFEM), which models multi-granularity edge structures using parallel branches and FFT to detect boundary distortions. Together, WFFB and EFEM provide complementary views of perturbation patterns. Extensive experiments demonstrate that AttackTracer achieves superior traceability of adversarial regions while maintaining robustness across stochastic sampling and varying scales, highlighting its effectiveness for instance-level attack localization.}, archive = {J_NEUCOM}, author = {Zhentong Zhang and Xinde Li and Pengfei Zhang and Kui Wang and Tianrong Gao and Tao Shen}, doi = {10.1016/j.neucom.2025.131535}, journal = {Neurocomputing}, month = {12}, pages = {131535}, shortjournal = {Neurocomputing}, title = {AttackTracer: Semantic-level adversarial attack location traceability via evidential diffusion model}, volume = {656}, year = {2025}, }
• (2025). A prototype-guided federated learning based fault diagnosis method of mechanical transmission system under label distribution skew. NEUCOM, 656, 131532. (www)

  @article{ , abstract = {To address the label distribution skew in federated learning based mechanical fault diagnosis, a federated learning based diagnosis framework combining prototypes and hybrid classifier is proposed. Firstly, prototypes are constructed based on sample feature means, and an exponential moving average strategy is introduced to smooth the aggregation of prototypes across rounds, while the prototype constraint loss function is constructed to guide the convergence of client features to the global prototype and compress the distance between similar samples. Secondly, a hybrid classifier architecture combining a local classifier with a global prototype classifier is proposed to learn local feature and global class prototypes through a two-branch structure, and a dynamic weighting strategy is used to achieve the output fusion. Finally, a prototype separation strategy is introduced on the server side, which detects pairs of confused class prototypes by Euclidean distance, increases the distance between similar prototypes, and avoids the prototype overlapping issue. In order to verify the effectiveness of the proposed method, nine kinds of faults of bearings, rotors and gears in mechanical transmission system are fabricated, and four types of fault diagnosis experiments with different degrees of label skew are designed, and the results show that the proposed method can effectively identify all the fault classes, and it still achieves an accuracy of 91.00 % in the extreme distribution skew task, which is significantly better than the other comparative methods, which provides a new feasible way for the distributed data driven federated learning based intelligent mechanical fault diagnosis.}, archive = {J_NEUCOM}, author = {Hongwei Fan and Shenglin Liu and Xiangang Cao and Xuhui Zhang}, doi = {10.1016/j.neucom.2025.131532}, journal = {Neurocomputing}, month = {12}, pages = {131532}, shortjournal = {Neurocomputing}, title = {A prototype-guided federated learning based fault diagnosis method of mechanical transmission system under label distribution skew}, volume = {656}, year = {2025}, }
• (2025). A predefined-time double-integral zeroing neural network model for linear equations flows and its application on dynamic position. NEUCOM, 656, 131531. (www)

  @article{ , abstract = {Solving time-varying linear equation flows presents a significant challenge in dynamic systems due to the continuously evolving coefficients, which undermine the effectiveness of traditional numerical methods. Moreover, the presence of external noise further exacerbates the difficulty of obtaining accurate solutions. To address these issues, this paper proposes a predefined-time double-integral zeroing neural network (PTDIZNN) model, inspired by the enhanced robustness of the conventional DIZNN framework. Specifically, a novel time-based gain is incorporated into the design of the DIZNN, ensuring predefined-time convergence of the proposed PTDIZNN model. A comprehensive theoretical analysis is conducted to verify its stability, convergence, and robustness properties. Furthermore, comparative simulations demonstrate that the PTDIZNN outperforms existing models in terms of solution accuracy and robustness under both column-full-rank and square-array coefficient scenarios. Finally, the effectiveness of the PTDIZNN is verified through its successful application in dynamic target positioning, highlighting its potential for broader real-time applications.}, archive = {J_NEUCOM}, author = {Jialiang Chen and Linju Li and Lin Xiao}, doi = {10.1016/j.neucom.2025.131531}, journal = {Neurocomputing}, month = {12}, pages = {131531}, shortjournal = {Neurocomputing}, title = {A predefined-time double-integral zeroing neural network model for linear equations flows and its application on dynamic position}, volume = {656}, year = {2025}, }
• (2025). SAFE-TAP: Semantic-aware and fused embedding for TAP rule security detection. NEUCOM, 656, 131529. (www)

  @article{ , abstract = {Trigger-Action Programming (TAP) has emerged as a widely adopted paradigm for enabling automated interoperability among IoT devices. Despite its convenience, TAP introduces significant security vulnerabilities. To address this issue, we propose SAFE-TAP, a novel framework for detecting malicious TAP rules that integrates global semantic understanding with temporal feature analysis. To further enhance the detection performance, we introduce an innovative data augmentation strategy that leverages Large Language Models (LLMs) to generate semantically consistent rule variations. This approach improves data set balance and enhances the generalizability of the model. Experimental results demonstrate that SAFE-TAP outperforms baseline methods, and the incorporation of LLM-based data augmentation significantly improves detection performance under imbalanced data scenarios.}, archive = {J_NEUCOM}, author = {Zhejun Kuang and Yusheng Zhu and Dawen Sun and Jian Zhao and Yongheng Xing and Feng Wang and Lei Sun}, doi = {10.1016/j.neucom.2025.131529}, journal = {Neurocomputing}, month = {12}, pages = {131529}, shortjournal = {Neurocomputing}, title = {SAFE-TAP: Semantic-aware and fused embedding for TAP rule security detection}, volume = {656}, year = {2025}, }
• (2025). Bipartite contrastive multi-view clustering with singular value modulation. NEUCOM, 656, 131528. (www)

  @article{ , abstract = {Contrastive multi-view clustering (CMvC) has attracted increasing attention for its semantic mining capacity. However, existing CMvC methods often process pairwise views to explore consistency, inevitably ignoring the joint information and inherent redundancy among multiple views. In this paper, we propose a novel Bipartite Contrastive Multi-view Clustering with Singular Value Modulation (BCMVC) that reformulates contrastive learning as a binary classification problem. Specifically, unlike existing pairwise-view sequential processing methods, we construct a correlation learning module that simultaneously mines consistent information across multiple views. This module effectively explores joint information at both the instance level and category level, with each level equipped with a dedicated correlation learner. By leveraging the concat and random shuffle strategy to encapsulate the positive and negative sample sets, the level-specific correlation learner is effectively optimized to enhance the discrimination of samples. Meanwhile, a deep singular value weighting module is introduced to refine the learned representations through a weighted singular value reconstruction strategy, mitigating the adverse effects of noisy information. Extensive experiments on seven benchmark datasets demonstrate that our method achieves substantial advancements compared with other state-of-the-art approaches. The code is available at https://github.com/zhangt-make/BCMVC .}, archive = {J_NEUCOM}, author = {Teng Zhang and Pengyuan Li and Zisen Kong and Dongxia Chang and Yao Zhao}, doi = {10.1016/j.neucom.2025.131528}, journal = {Neurocomputing}, month = {12}, pages = {131528}, shortjournal = {Neurocomputing}, title = {Bipartite contrastive multi-view clustering with singular value modulation}, volume = {656}, year = {2025}, }
• (2025). The framework and memristive circuit design for attention-regulated working memory. NEUCOM, 656, 131525. (www)

  @article{ , abstract = {Adaptive behavior and decision-making depend on constantly selecting relevant information from the external environment and internal states. Inspired by the working memory structure and the top-down and bottom-up attention mechanisms in cognitive neuroscience, this work proposes an attention-regulated working memory model. This model provides a brain-inspired approach to integrate perception, long-term memory, and action. It processes current external multisensory stimuli and retrieves stored knowledge from internal reinforcement simultaneously, leading to adaptive and rapid executive actions. On this basis, a memristive circuit is designed to realize rich cognitive functions in an online in-situ learning and in-memory computing manner. The designed circuit consists of four main components: (1) the phonological loop and visuospatial sketchpad consider different audio-visual input patterns and varying stimulus salience, realizing the filtration, synchronization, and encoding of multimodal signals; (2) the attention control module captures and maintains attention driven by multisensory stimulation; (3) the episodic buffer achieves reward reinforcement, forming or resetting the top-down attentional bias signal; (4) the central executive control module regulates the relationships between the two attentional pathways, thus transforming the random exploration process into a learnable final action. Finally, simulation results in LTSPICE demonstrate that our circuit can be adaptively applied to the cognitive control and execution system of robots within complicated circumstances.}, archive = {J_NEUCOM}, author = {Jihong Zhang and Xiaoping Wang and Zhanfei Chen and Chao Yang}, doi = {10.1016/j.neucom.2025.131525}, journal = {Neurocomputing}, month = {12}, pages = {131525}, shortjournal = {Neurocomputing}, title = {The framework and memristive circuit design for attention-regulated working memory}, volume = {656}, year = {2025}, }
• (2025). A temporally coded multilayer spiking neural network and its memristor-based hardware implementation. NEUCOM, 656, 131523. (www)

  @article{ , abstract = {Artificial neural networks (ANNs) have demonstrated remarkable progress in various domains. However, ANNs suffer from enormous time and energy consumption during training and inference processes. Brain-inspired spiking neural networks (SNNs) have recently attracted more attention due to their higher biological plausibility and potential cost-efficient properties. However, most existing SNNs significantly degrade in performance and efficiency when simulated on conventional CPU/GPU hardware. Therefore, a novel temporally coded multilayer SNN (TMSNN) is proposed in this study. It is a typical event-driven model, which encodes information in the relative timing of spikes rather than in firing rates and uses the leaky integrate-and-fire neuron as the basic unit to pursue high biological plausibility. Its multilayer architecture enables the model to solve complicated problems effectively. On the other hand, the proposed TMSNN can be implemented on memristor-based hardware, which uses customized weight quantization and sharing techniques to mitigate the size restrictions of the memristor crossbars. After refining the weights using the simulated annealing algorithm, the hardware implementation of TMSNN can achieve very competitive performance on benchmark datasets, outperforming state-of-the-art temporally coded SNNs in our experiments. The source code of TMSNN is available at https://github.com/jhc050998/Memristor-Crossbar-Based-SNN .}, archive = {J_NEUCOM}, author = {Haochang Jin and Xiuzhi Yang and Shuangbao Song and Zhenyu Song and Junkai Ji}, doi = {10.1016/j.neucom.2025.131523}, journal = {Neurocomputing}, month = {12}, pages = {131523}, shortjournal = {Neurocomputing}, title = {A temporally coded multilayer spiking neural network and its memristor-based hardware implementation}, volume = {656}, year = {2025}, }
• (2025). Exponential extended dissipative synchronization control of Cohen–Grossberg neural networks with four kinds of time-varying delays. NEUCOM, 656, 131522. (www)

  @article{ , abstract = {This paper studies the exponential extended dissipative synchronization control problem of Cohen–Grossberg neural networks (CGNNs) with four kinds of time-varying delays. The types of delays involve time-varying leakage, neutral, distributed and transmission delays. Due to the increasing complexity of control requirements and time delays in practice, some performance analysis approaches and techniques cannot be directly applied, or are faced with the problem of high computational complexity. To this end, a more general and computationally efficient novel method is proposed. Firstly, a sufficient condition to guarantee the existence and uniqueness of the solution of CGNN is presented by defining a new norm, and a representation of the unique solution is first put forward. Then, the state-feedback controller and novel system solutions-based inequality are constructed to obtain exponential extended dissipative synchronization criteria. This proposed approach overcomes the difficulty of constructing a suitable Lyapunov–Krasovskii functional (LKF) under complex time delays and control requirements, and reduces computational complexity. Furthermore, to solve the nonlinear terms in the obtained criteria, an algorithm is designed. Finally, the derived results are validated for feasibility by three numerical examples, and their potential applications in image processing are showcased.}, archive = {J_NEUCOM}, author = {Kairong Tu and Yu Xue}, doi = {10.1016/j.neucom.2025.131522}, journal = {Neurocomputing}, month = {12}, pages = {131522}, shortjournal = {Neurocomputing}, title = {Exponential extended dissipative synchronization control of Cohen–Grossberg neural networks with four kinds of time-varying delays}, volume = {656}, year = {2025}, }
• (2025). A rate-dependent coreset selector for continual learning on time-varying data distributions. NEUCOM, 656, 131519. (www)

  @article{ , abstract = {In this paper we review the concept of “phase” defined in Class-Incremental Learning (CIL), i.e., learning new classes while not forgetting old ones. Due to this design, classic CIL algorithms are mostly offline or can handle only intensive data distribution shifts across the phases. However, real-world data streams are often online, usually with uncertain or untraceable changes in their data distributions. To this end, we design the per-step distribution shifts by modeling the class sampling weights using bell-shaped curves. Such a design respects the rise-and-fall nature and presents realistic but underexplored challenges for CIL: 1) The data non-stationarity across steps requires the models to identify the recent dynamics and adopt an appropriate learning strategy for knowledge memorization and adaptation. 2) Over all steps, the proposed streams exhibit various class-imbalance patterns , with different majority classes and time-varying imbalance ratios. To address the challenges, we propose a novel Rate-Dependent Coreset Selector (RDCS), which essentially presents an adaptive and robust sample selection criterion when constructing memory for replay. We conduct extensive experiments by generating the proposed data streams on multiple image benchmarks and implementing RDCS in an efficient approximation, showing its superior performance.}, archive = {J_NEUCOM}, author = {Zilin Luo and Zichen Tian and Yaoyao Liu and Qianru Sun}, doi = {10.1016/j.neucom.2025.131519}, journal = {Neurocomputing}, month = {12}, pages = {131519}, shortjournal = {Neurocomputing}, title = {A rate-dependent coreset selector for continual learning on time-varying data distributions}, volume = {656}, year = {2025}, }
• (2025). MLFork: Bearing fault diagnosis via mamba-powered few-shot learning model with multi-level architecture enhanced by spatial-wise and channel-wise local vector attention. NEUCOM, 656, 131518. (www)

  @article{ , abstract = {Bearings are one of the most important components of electrical machines and devices; however, they are prone to damage, leading to the lack of safety and the malfunction of machines. Some methods including deep-learning ones can be used for bearing fault diagnosis; however, in reality, the models have to adapt to the shortage of training data from clients while still maintaining good performance. To overcome this issue, the novel “MLFork” model is proposed, following the Few-shot algorithm for limited training with improvements in the feature extraction and the pre-classification steps. For feature extraction, a new Bi-Context Visual State Space Block is introduced, which excellently learns the global context of the sample in multiple ways. Before the Multi-Level classification module, separate routes for spatial-wise and channel-wise local vector attention are used to highlight the important details of the local descriptor. To evaluate the performance of the model, various experiments were done on the Case Western Reserve University dataset (CWRU) and the Paderborn University dataset (PU), where the “MLFork" model showed promising results. The code for this model will be available at: https://github.com/thzhere/MLFork .}, archive = {J_NEUCOM}, author = {Duy-Thai Nguyen and Van-Quoc-Viet Nguyen and Thi-Thao Tran and Van-Truong Pham}, doi = {10.1016/j.neucom.2025.131518}, journal = {Neurocomputing}, month = {12}, pages = {131518}, shortjournal = {Neurocomputing}, title = {MLFork: Bearing fault diagnosis via mamba-powered few-shot learning model with multi-level architecture enhanced by spatial-wise and channel-wise local vector attention}, volume = {656}, year = {2025}, }
• (2025). All-in-one image restoration via diffusion models with degradation perception and semantic enhancement. NEUCOM, 656, 131517. (www)

  @article{ , abstract = {Image restoration is a fundamental task in computer vision. However, most existing methods are tailored for single-degradation scenarios, limiting their applicability in real-world conditions where multiple degradations often co-occur. To address this issue, we propose a degradation-aware image restoration framework. A bidirectional Mamba module is introduced to process fused spatial-frequency features, enabling accurate identification of degradation types via a multi-degradation encoding strategy. Based on the predicted degradation, a fine-tuned CLIP model with an attention mechanism is employed to extract semantic features. These features are then integrated with degradation representations and fed into a conditional denoising diffusion model to progressively reconstruct high-quality images. To facilitate evaluation, we construct the Multi-Degradation Perception Dataset (MDPD), specifically designed for complex degradation scenarios. Experimental results demonstrate that our method achieves over 98 % classification accuracy in identifying degradation types. On the MDPD dataset, it achieves a PSNR of 36.25 dB and improves SSIM by 0.01 to 0.04 across various degradation combinations.}, archive = {J_NEUCOM}, author = {Jiangang Jiang and Zhe Chen and Yuxin Su and Pancheng Zhang and Yihui Hu}, doi = {10.1016/j.neucom.2025.131517}, journal = {Neurocomputing}, month = {12}, pages = {131517}, shortjournal = {Neurocomputing}, title = {All-in-one image restoration via diffusion models with degradation perception and semantic enhancement}, volume = {656}, year = {2025}, }
• (2025). Neural operator-based composite learning adaptive backstepping control for linear 2×2 hyperbolic PDE systems. NEUCOM, 656, 131516. (www)

  @article{ , abstract = {This paper proposes a novel neural operator (NO)-based composite learning adaptive backstepping control scheme for stabilizing uncertain linear 2 × 2 hyperbolic PDE systems. This method addresses key challenges arising from complex PDE dynamics, model uncertainties, and high computational costs, within a backstepping design framework. Our approach integrates two main components: 1) A composite learning adaptive controller, which combines both historical and real-time data to construct informative matrices for parameter updates. This strategy enables accurate and exponential parameter convergence under finite excitation (FE) conditions, thereby improving transient performance and guaranteeing exponential system stability. 2) An efficient NO-based approximation method, where a deep operator network (DeepONet) is trained to approximate the nonlinear mapping from composite parameter estimates to backstepping kernel gains. The controller is constructed using the approximate kernels, which eliminates the need to repeatedly solve kernel PDE online, significantly improving the computational efficiency and accelerating real-time control. Furthermore, theoretical analysis proves closed-loop boundedness and exponential stability under the proposed scheme. Numerical simulations verify its effectiveness and superiority.}, archive = {J_NEUCOM}, author = {Xianhe Zhang and Yu Xiao and Xiaodong Xu and Biao Luo and Weihua Gui and Tingwen Huang}, doi = {10.1016/j.neucom.2025.131516}, journal = {Neurocomputing}, month = {12}, pages = {131516}, shortjournal = {Neurocomputing}, title = {Neural operator-based composite learning adaptive backstepping control for linear 2×2 hyperbolic PDE systems}, volume = {656}, year = {2025}, }
• (2025). RAG-LER: Ranking adapted generation with language-model enabled regulation. NEUCOM, 656, 131514. (www)

  @article{ , abstract = {Large language models (LLMs) have demonstrated impressive capabilities across diverse NLP tasks, yet they still struggle with hallucination due to limited parametric knowledge. Retrieval Augmented Generation (RAG) addresses this issue by integrating non-parametric data stores. However, straightforward integration of information retrieval or end-to-end training of these components often leads to suboptimal results or computational inefficiency. In this work, we introduce RAG-LER, a framework that enhances an LM’s context understanding and improves the quality and accuracy of provided passages through an LM-supervised re-ranker. RAG-LER fine-tunes a pre-trained LM to follow instructions and discriminately use provided information. It then leverages this fine-tuned LM to generate ranking scores, which serve as supervised labels for training the re-ranker. We also introduce a confidence-weighted objective that filters unreliable LLM supervision signals while preserving the original re-ranker capabilities. By harnessing LLMs’ strong capabilities, our approach eliminates the need for manual human labeling in re-ranker training while achieving improved performance. Experiments demonstrate that RAG-LER outperforms existing retrieval-augmented LMs on open-domain QA and fact-checking tasks, while exhibiting consistently improved performance when applied to different retrieval methods, highlighting its versatility and effectiveness.}, archive = {J_NEUCOM}, author = {Fengwen Zhai and Wenyang Tang and Jing Jin}, doi = {10.1016/j.neucom.2025.131514}, journal = {Neurocomputing}, month = {12}, pages = {131514}, shortjournal = {Neurocomputing}, title = {RAG-LER: Ranking adapted generation with language-model enabled regulation}, volume = {656}, year = {2025}, }
• (2025). TalkingAvatar: Learning 3D talking human avatar via NeRF. NEUCOM, 656, 131513. (www)

  @article{ , abstract = {Despite the remarkable progress in 3D talking head generation, directly generating expressive upper-body 3D talking avatars that integrate realistic talking heads and lifelike bodies remains challenging. To address these issues, a novel TalkingAvatar is proposed to reconstruct vivid avatars from dynamic monocular videos while endowing them with expressive animation capabilities. Given audio and SMPL-X pose sequences, TalkingAvatar can animate 3D human avatars with lip-synced mouth movements and complex hand gestures. Specifically, TalkingAvatar utilizes a fast deformer for articulated neural fields (Fast-SNARF) to optimize human volumetric representations in a canonical T-pose. It also incorporates a novel Hybrid Modulation Attention Module (HMAM) that focuses on capturing dynamic movements. Additionally, a UV feature map is used for detailed static texture learning. Notably, HMAM leverages a modulation mechanism to exploit comprehensive dynamic motion information from both body pose and speech audio, ensuring expressive control of mouth and hands. Moreover, to minimize interference between different body parts, TalkingAvatar adopts a part-aware learning strategy that employs multiple regional grids to model the head, hands, and body areas separately. This strategy significantly enhances the fidelity of small-scale body regions. Experiments demonstrate that our method outperforms the state-of-the-art in both avatar reconstruction and animation of mouth and hand movements, generating high-fidelity co-speech gesture videos, while requiring significantly less training data.}, archive = {J_NEUCOM}, author = {Lingyun Yu and Chuang Chen and Chuanbin Liu and Wu Liu and Quanwei Yang and Yizhi Liu and Meng Shao}, doi = {10.1016/j.neucom.2025.131513}, journal = {Neurocomputing}, month = {12}, pages = {131513}, shortjournal = {Neurocomputing}, title = {TalkingAvatar: Learning 3D talking human avatar via NeRF}, volume = {656}, year = {2025}, }
• (2025). Incremental algorithms on low-rank approximation of large-scale kernel matrices based on perturbation of invariant subspaces. NEUCOM, 656, 131512. (www)

  @article{ , abstract = {Kernel methods are one of the most commonly used techniques in machine learning. In Mitz and Shkolnisky (2022) [27] , a framework of perturbation-based kernel matrix approximation is proposed, which is based on the tool of matrix perturbation analysis. However, there are two shortcomings in this framework. First, it requires that some dominant eigenvalues of kernel matrices are distinct in theory. However, in practical applications, when using a randomly sampled dataset, some kernel matrices generated by certain kernel functions are prone to having multiple eigenvalues due to data distribution or parameter settings. Second, from the algorithmic perspective, one has to know the error matrix of the kernel matrix in advance, which is unrealistic for real-world applications. Thus, the most common situation in practical applications is to pay attention to the case of multiple eigenvalues, and it is interesting to generalize the original perturbation-based kernel approximation framework to the scenario where there are multiple eigenvalues. In this work, we present a perturbation result on eigenvalues and eigenspaces of a kernel matrix whose dominant eigenvalues can be multiple. Based on this result, we propose a low-rank approximation to kernel matrix. On the other hand, as far as we are aware, efficient algorithms are still lacking for updating large-scale kernel matrices, and there are few algorithms addressing batch-incremental kernel methods. Based on our proposed truncated formula, we consider the incremental problem of large-scale kernel matrices and propose two incremental algorithms for updating large-scale kernel matrices. Numerical experiments demonstrate the efficiency of the proposed algorithms for solving incremental data problems and incremental kernel ridge regression.}, archive = {J_NEUCOM}, author = {Xiaxin Li and Gang Wu}, doi = {10.1016/j.neucom.2025.131512}, journal = {Neurocomputing}, month = {12}, pages = {131512}, shortjournal = {Neurocomputing}, title = {Incremental algorithms on low-rank approximation of large-scale kernel matrices based on perturbation of invariant subspaces}, volume = {656}, year = {2025}, }
• (2025). FMSF: Future-preference modeling with similar-user features for next POI recommendation. NEUCOM, 656, 131511. (www)

  @article{ , abstract = {Abundant user check-in records in location-based social networks enhance the development of point-of-interest (POI) recommendation systems. The existing studies attempt to learn users’ past, current, and future preferences from their own sequential behaviors. Various approaches have been explored to model user visiting behaviors for the prediction of future preferences and have achieved considerable performance. However, most previous work ignores the impact of other users’ preferences on the prediction of current users’ future preferences. Thus, this work proposes a novel Future-preference Modeling with Similar-user Features (FMSF) model for next POI recommendation. It integrates the preferences of a user and those of other users to accurately model his/her multi-step future preferences. Specifically, it adopts a dynamically-updated similarity matrix to extract the information of similar users. Then, it incorporates an attention mechanism to assign distinct attention weights to the characteristics of both the current and similar users, which promotes the prediction of the future preferences of the current users. Therefore, the method proposed in this paper can offer users more precise recommendation results. Extensive experiments are conducted on three real-world datasets, which demonstrate the advantages of the proposed method.}, archive = {J_NEUCOM}, author = {Wenjing Luan and Zhichao Feng and Liang Qi and Xiaoyu Sean Lu}, doi = {10.1016/j.neucom.2025.131511}, journal = {Neurocomputing}, month = {12}, pages = {131511}, shortjournal = {Neurocomputing}, title = {FMSF: Future-preference modeling with similar-user features for next POI recommendation}, volume = {656}, year = {2025}, }
• (2025). Permutation XOR cellular automata and direct stable periodic orbits. NEUCOM, 656, 131510. (www)

  @article{ , abstract = {This paper presents a permutation XOR cellular automaton (PXCA), a simple three-layer discrete-time dynamical system. The input-to-hidden layer corresponds to an elementary cellular automaton of the XOR rule and the hidden-to-output layer is the shift-type one-to-one permutation connection. The dynamics are described by an autonomous difference equation of binary state variables. Depending on the permutation connection, the PXCA generates a variety of direct stable periodic orbits (DBPOs) characterized by strong stability and fast transient phenomena. As a main result, we provide theoretical evidence that clarifies the number, period, and stability of DBPOs for general odd-dimensional PXCAs. Performing a precise numerical analysis, we have clarified that, depending on the dimension and a parameter, the period of DBPOs varies complicatedly and can become very long. Applications of the DBPOs include time series approximation and switching circuit control. As a fundamental step toward the applications, we present a simple FPGA based hardware prototype and have confirmed typical DBPOs experimentally.}, archive = {J_NEUCOM}, author = {Mikito Onuki and Yosuke Suzuki and Toshimichi Saito}, doi = {10.1016/j.neucom.2025.131510}, journal = {Neurocomputing}, month = {12}, pages = {131510}, shortjournal = {Neurocomputing}, title = {Permutation XOR cellular automata and direct stable periodic orbits}, volume = {656}, year = {2025}, }
• (2025). Multi-modal dynamic brain graph representation learning for brain disorder diagnosis via temporal sequence model. NEUCOM, 656, 131509. (www)

  @article{ , abstract = {Brain network representation learning leverages graph-based algorithms to enhance understanding of functional brain organization. Recently, deep learning approaches based on graph neural network (GNN) have shown promising results in various brain network analysis tasks. Nevertheless, despite significant achievements in brain graph learning, early models still exhibit limitations in dynamic modeling and multi-modal network fusion. Dynamic modeling of brain networks entails learning sequential spatial interactions across time. Inspired by recent advances in large language model architectures, particularly RWKV, which combines the strengths of recurrent neural networks (RNNs) and Transformers. We propose an e fficient t emporal m ulti-modal g raph n eural n etwork (ET_MGNN), that captures complex temporal dependencies while integrating dynamic functional connectivity (DFC) and structural connectivity (SC) into a unified brain network representation. The proposed model demonstrates competitive performance in brain disorder classification on three datasets, outperforming several strong baselines. For instance, ET_MGNN an average classification accuracy improvement of 11.8 % on autism spectrum disorder (ASD) vs healthy controls, 32.9 % on Alzheimer's disease (AD) vs. mild cognitive impairment (MCI), compared to the well-suited STAGIN model. Furthermore, we introduce an interpretable graph reading mechanism that can identify disorder-relevant brain regions. In summary, ET_MGNN combines large-scale language sequence modeling with dynamic brain graph representation learning to improve the accuracy of brain disease diagnosis, providing insightful findings for dynamic brain network modeling.}, archive = {J_NEUCOM}, author = {Jinwei Lang and Li-Zhuang Yang and Hai Li}, doi = {10.1016/j.neucom.2025.131509}, journal = {Neurocomputing}, month = {12}, pages = {131509}, shortjournal = {Neurocomputing}, title = {Multi-modal dynamic brain graph representation learning for brain disorder diagnosis via temporal sequence model}, volume = {656}, year = {2025}, }
• (2025). PVFT-net: A point-voxel fusion method for self-supervised scene flow estimation with transformer. NEUCOM, 656, 131508. (www)

  @article{ , abstract = {Scene flow estimation is a computer vision task that aims to estimate the 3D motion field of points from two consecutive frames of point clouds, and has a wide range of applications in various fields such as robotics and autonomous driving. Most of the existing methods estimate scene flow through point-based models, but ignore the irregularity of point clouds and the inefficiency of point-level computation. And voxel-based methods can hardly avoid the loss of detailed information. Therefore, we propose a point-voxel fusion method that contains a point branch and a voxel branch. The voxel branch projects the point cloud to regular local grids and captures coarse-grained local features from non-empty voxels through Sparse Grid Attention (SGA) with the shift window strategy. And the point branch captures fine-grained global features through dual attention consisting of Deformable Global Attention (DGA) and Channel Self-Attention (CSA), while compensating for the information loss in the voxel branch. Considering that it is difficult to directly describe the local geometric structure of complex objects in the scene with the shape of 3D objects potentially learned only through xyz coordinates, we explicitly encode the local surface information of the point cloud through the Umbrella Surface Feature Extraction (USFE) module. In addition, we introduce Density Sensitive Metric(DSM) loss to reduce the impact of outliers and density distribution mismatch problems. We validate the effectiveness of our method by performing experiments on the Flyingthings3D and KITTI datasets. Our method outperforms all other self-supervised methods and achieves highly competitive results compared to fully supervised methods. We achieve improvements in all metrics, especially EPE, which is decreased by 8.51 % on the KITTI o dataset and 15.79 % on the KITTI s dataset.}, archive = {J_NEUCOM}, author = {Xuezhi Xiang and Xi Wang and Xiaoheng Li and Xiankun Zhou and Lei Zhang and Xiantong Zhen}, doi = {10.1016/j.neucom.2025.131508}, journal = {Neurocomputing}, month = {12}, pages = {131508}, shortjournal = {Neurocomputing}, title = {PVFT-net: A point-voxel fusion method for self-supervised scene flow estimation with transformer}, volume = {656}, year = {2025}, }
• (2025). Attacking the spike: On the security of spiking neural networks to adversarial examples. NEUCOM, 656, 131506. (www)

  @article{ , abstract = {Spiking neural networks (SNNs) have attracted much attention for their high energy efficiency and for recent advances in their classification performance. However, unlike traditional deep learning approaches, the analysis and study of the robustness of SNNs to adversarial examples remain relatively underdeveloped. In this work, we focus on advancing the adversarial attack side of SNNs and make three major contributions. First, we show that successful white-box adversarial attacks on SNNs are highly dependent on the underlying surrogate gradient estimation technique, even in the case of adversarially trained SNNs. Second, using the best single surrogate gradient estimation technique, we analyze the transferability of adversarial attacks on SNNs and other state-of-the-art architectures like Vision Transformers (ViTs), as well as CNNs. Our analyzes reveal two key areas where SNN adversarial attacks can be enhanced: no white-box attack effectively exploits the use of multiple surrogate gradient estimators for SNNs, and no single model attack is effective at generating adversarial examples misclassified by both SNNs and non-SNN models simultaneously. For our third contribution, we develop a new attack, the Mixed Dynamic Spiking Estimation (MDSE) attack to address these issues. MDSE utilizes a dynamic gradient estimation scheme to fully exploit multiple surrogate gradient estimator functions. In addition, our novel attack generates adversarial examples capable of fooling both SNN and non-SNN models simultaneously. The MDSE attack is as much as 91.4 % more effective on SNN/ViT model ensembles and provides a 3 × boost in attack effectiveness on adversarially trained SNN ensembles, compared to conventional white-box attacks like Auto-PGD. Our experiments are broad and rigorous, covering three datasets (CIFAR-10, CIFAR-100 and ImageNet) and nineteen classifier models (seven for each CIFAR dataset and five models for ImageNet). We will release a fully publicly available code repository for the models and attacks upon publication.}, archive = {J_NEUCOM}, author = {Nuo Xu and Kaleel Mahmood and Haowen Fang and Ethan Rathbun and Caiwen Ding and Wujie Wen}, doi = {10.1016/j.neucom.2025.131506}, journal = {Neurocomputing}, month = {12}, pages = {131506}, shortjournal = {Neurocomputing}, title = {Attacking the spike: On the security of spiking neural networks to adversarial examples}, volume = {656}, year = {2025}, }
• (2025). Text-to-image person re-identification via collaborating pre-trained diffusion and discriminative models. NEUCOM, 656, 131505. (www)

  @article{ , abstract = {Text-to-image person re-identification (TIReID) aims to retrieve pedestrian images from a database that match given text queries. Currently, the most advanced methods involve transferring powerful multi-modal knowledge from the contrastive language-image pretraining (CLIP) model to perform cross-modal matching. However, CLIP primarily focuses on coarse-grained global contextual modeling of single image-text pairs, neglecting fine-grained compositional matching of complex visual-textual concepts. This makes it challenging to ensure fine-grained cross-modal matching between pedestrians and text queries. To address this issue, a novel framework, Collaborating Pre-trained Diffusion and Discriminative Models (CPDD), is proposed in this work. The CPDD comprises three modules: a fine-grained features learning (FFL) module, a semantic consistency alignment (SCA) module, and a masked-text interactive modeling (MIM) module. Firstly, the FFL learns feature representations containing fine-grained matching information between images and text through the reverse denoising process of a diffusion model. Next, a semantic consistency loss is designed in the SCA, which ensures the semantic consistency between the fine-grained matching information and the input image and text information. Then, the MIM propagates fine-grained matching information into the visual- textual context by a cross-modal interactive encoder, achieving fine-grained matching between images and text and enabling fine-grained cross-modal matching. Extensive experiments on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets show that the proposed method achieves significant performance improvements compared to current research results, achieving Rank-1 accuracy of 74.87 %, 63.31 %, and 61.26 %, respectively.}, archive = {J_NEUCOM}, author = {Wenjing Zhang and Chenyue Xu and Huajing Wu and Quange Tan and Qianli Zhou and Rong Wang}, doi = {10.1016/j.neucom.2025.131505}, journal = {Neurocomputing}, month = {12}, pages = {131505}, shortjournal = {Neurocomputing}, title = {Text-to-image person re-identification via collaborating pre-trained diffusion and discriminative models}, volume = {656}, year = {2025}, }
• (2025). Layer-wise contrastive learning BERT for sentence representation of GitHub. NEUCOM, 656, 131504. (www)

  @article{ , abstract = {Every day, on GitHub, end-users submit a large number of issues that must be addressed to ensure the success of software projects. Though BERT-based pre-trained language models achieve high performance on many downstream tasks, the sentence representation of [CLS] from the top layer of BERT has a limited ability to capture the semantic meaning of sentences. GitHub issue reports often include code snippets and user-generated terms not found in standard vocabularies. Therefore, the classification predictions of BERT are affected. To generate better sentence semantic representations of BERT for GitHub, we propose a layer-wise Contrastive Learning BERT (CLBERT), which uses contrastive learning to enhance the representation ability by contrasting the layer-by-layer representation. Further, to obtain as comprehensive information as possible, representations of each layer are extracted and learned by an attention mechanism as the final classification features. Finally, experiments conducted on two GitHub data sets show that our proposed model significantly improves classification performance.}, archive = {J_NEUCOM}, author = {Daoquan Chen and Wei Zhang and Shengyu Lu and Yuanguo Lin and Xinyu Gu and Xiuze Zhou}, doi = {10.1016/j.neucom.2025.131504}, journal = {Neurocomputing}, month = {12}, pages = {131504}, shortjournal = {Neurocomputing}, title = {Layer-wise contrastive learning BERT for sentence representation of GitHub}, volume = {656}, year = {2025}, }
• (2025). Attacking all tasks at once using adversarial examples in multi-task learning. NEUCOM, 656, 131503. (www)

  @article{ , abstract = {Visual content understanding frequently relies on multi-task models to extract robust representations of a single visual input for multiple downstream tasks. However, in comparison to extensively studied single-task models, the adversarial robustness of multi-task models has received significantly less attention and many questions remain unclear: (1) How robust are multi-task models to single task adversarial attacks, (2) Can adversarial attacks be designed to simultaneously attack all tasks in a multi-task model, and (3) How does parameter sharing across tasks affect multi-task model robustness to adversarial attacks? This paper aims to answer these questions through careful analysis and rigorous experimentation. First, we analyze the inherent drawbacks of two commonly-used adaptations of single-task white-box attacks in attacking multi-task models. We then propose a novel attack framework, Dynamic Gradient Balancing Attack (DGBA). Our framework poses the problem of attacking all tasks in a multi-task model as an optimization problem that can be efficiently solved through integer linear programming. Extensive evaluation on two popular MTL benchmarks, NYUv2 and Tiny-Taxonomy, demonstrates the effectiveness of DGBA compared to baselines in attacking both clean and adversarially trained multi-task models. Our results also reveal a fundamental trade-off between improving task accuracy via parameter sharing across tasks and undermining model robustness due to increased attack transferability from parameter sharing.}, archive = {J_NEUCOM}, author = {Lijun Zhang and Xiao Liu and Kaleel Mahmood and Caiwen Ding and Hui Guan}, doi = {10.1016/j.neucom.2025.131503}, journal = {Neurocomputing}, month = {12}, pages = {131503}, shortjournal = {Neurocomputing}, title = {Attacking all tasks at once using adversarial examples in multi-task learning}, volume = {656}, year = {2025}, }
• (2025). Convolutional neural network combined with edge complexity and channel attention mechanism for unsupervised superpixel segmentation. NEUCOM, 656, 131502. (www)

  @article{ , abstract = {Superpixel segmentation is crucial for enhancing image processing efficiency and accuracy. To address the challenges of decreased accuracy and insufficient stability in adaptive superpixel generation faced by existing algorithms in complex image segmentation, we propose ECN, an unsupervised superpixel segmentation algorithm based on convolutional neural networks (CNN) integrating edge complexity and channel attention mechanisms. The ECN algorithm first calculates edge complexity using the Sobel operator, which guides the sequential network in determining the number of feature channels and the kernel size of the fast 1D convolution. Subsequently, low-level features with positional information are transformed into deep features through the sequential network, dynamically adjusting the weights of each feature channel using the channel attention mechanism. Finally, the target function is minimized during inference, enabling unsupervised superpixel generation. We validate ECN's applicability by combining it with Linear Discriminant Analysis (LDA) and Locality Fisher Discriminant Analysis (LFDA) to develop Superpixel Unsupervised Linear Discriminant Analysis (SULDA). Experimental results on BSDS500 and NYUv2 datasets show ECN outperforms existing methods, producing stable and higher-quality superpixel segmentation. Application tests on Indian Pines and Pavia University scenes confirm ECN's significant practical utility.}, archive = {J_NEUCOM}, author = {Fugui Luo and Shihua Li and Minghui Chang and Yuting Liu and Kaitong Liu}, doi = {10.1016/j.neucom.2025.131502}, journal = {Neurocomputing}, month = {12}, pages = {131502}, shortjournal = {Neurocomputing}, title = {Convolutional neural network combined with edge complexity and channel attention mechanism for unsupervised superpixel segmentation}, volume = {656}, year = {2025}, }
• (2025). A knowledge distillation-based object detection model focused on road scene perception and localization. NEUCOM, 656, 131501. (www)

  @article{ , abstract = {Object detection is crucial for unmanned systems, as it enables real-time classification and localization of objects in road scenes. Besides detection accuracy, which remains robust to variations in object scales, an effective object detection algorithm also demands superior performance in processing time. To address these issues, this paper proposes a knowledge distillation-based object detection model, PLE-RepPoints-Lite, to compromise the performance of detection accuracy and speed for unmanned systems. We also design perception and localization enhancement (PLE) strategies, which consist of parallel dynamic attention, multi-scale composite localization confidence, and a feedback closed-loop structure, to enhance the capabilities of perception and localization in complex road environments. To improve the real-time performance, a hybrid lightweight approach for road scenes is designed. Experimental results on the Cityscapes and BDD100K datasets show that our approach achieves state-of-the-art results with average precision (AP) of 34.6 and 40.1, respectively. Furthermore, it operates at 34.2 frames per second (FPS) at a 1280 × 640 resolution, satisfying real-time requirements.}, archive = {J_NEUCOM}, author = {Yufei Xie and Ying Shi and Changjun Xie and Qin Hu and Yue Liu and Chaojun Lin}, doi = {10.1016/j.neucom.2025.131501}, journal = {Neurocomputing}, month = {12}, pages = {131501}, shortjournal = {Neurocomputing}, title = {A knowledge distillation-based object detection model focused on road scene perception and localization}, volume = {656}, year = {2025}, }
• (2025). Confident neural network regression with bootstrapped deep ensembles. NEUCOM, 656, 131500. (www)

  @article{ , abstract = {With the rise in the popularity and usage of neural networks, trustworthy uncertainty estimation is becoming increasingly essential. One of the most prominent uncertainty estimation methods is Deep Ensembles [20]. A classical parametric model has uncertainty in the parameters due to the fact that the data on which the model is built is a random sample. A modern neural network has an additional uncertainty component since the optimization of the network is random. Lakshminarayanan et al. [20] noted that Deep Ensembles do not incorporate the classical uncertainty induced by the effect of finite data. In this paper, we present a computationally cheap extension of Deep Ensembles for the regression setting, called Bootstrapped Deep Ensembles , that explicitly takes this classical effect of finite data into account using a modified version of the parametric bootstrap. We demonstrate through an experimental study that our method significantly improves upon standard Deep Ensembles. The resulting confidence intervals demonstrate superior coverage without sacrificing accuracy.}, archive = {J_NEUCOM}, author = {Laurens Sluijterman and Eric Cator and Tom Heskes}, doi = {10.1016/j.neucom.2025.131500}, journal = {Neurocomputing}, month = {12}, pages = {131500}, shortjournal = {Neurocomputing}, title = {Confident neural network regression with bootstrapped deep ensembles}, volume = {656}, year = {2025}, }
• (2025). Parameter-free multi-view clustering via refined tensor learning. NEUCOM, 656, 131497. (www)

  @article{ , abstract = {As multi-view data becomes more prevalent in real-world applications, multi-view clustering (MVC) has emerged as a powerful technique for unsupervised representation learning. To uncover the intrinsic structure, it is crucial to consider information from different spaces. Focusing solely on the sample space limits the method’s ability to effectively model multi-view data, as the informative patterns embedded in the feature space are often overlooked. Furthermore, to integrate high-order correlations, tensor-based MVC methods have been widely adopted to preserve the low rank structure of multi-view data. Traditional tensors can not achieve selective tensor rank minimization as they lack an explicit mechanism to model the retention of singular values based on their individual information contributions. Additionally, existing methods rely on hyper-parameters, undermining generalizability across different datasets. In response to these limitations, we propose a novel Parameter-free Multi-view Clustering via Refined Tensor Learning (PRTL), which is based on bidirectional regression matrices to perform data reconstruction and extract salient features. To further achieve an adaptive low-rank tensor structure, we propose a Quadratic Decay Tensor (QDT) regularization as a non-convex alternative to conventional rank minimization, which selectively retains salient information while filtering out noise dynamically, resulting in a more expressive joint representation. Meanwhile, we incorporate the hyper-Laplace graph to capture richer relationships than those modeled by conventional pairwise graphs. Notably, PRTL eliminates the need for hyper-parameters, making it more practical and robust. Experiments on diverse datasets demonstrate that PRTL consistently surpasses existing state-of-the-art clustering methods. Our code is available at https://github.com/jiaxinyang04/PRTL .}, archive = {J_NEUCOM}, author = {Jiaxin Yang and Qian Liu and Yuemeng Huang and Chunyan Yang and Wengeng Chen and Yu Lu and Jiale Wang and Wenzhe Liu and Huibing Wang}, doi = {10.1016/j.neucom.2025.131497}, journal = {Neurocomputing}, month = {12}, pages = {131497}, shortjournal = {Neurocomputing}, title = {Parameter-free multi-view clustering via refined tensor learning}, volume = {656}, year = {2025}, }
• (2025). An extensive multi-variant deep neural network approach to enhance genomic prediction of endometriosis. NEUCOM, 656, 131496. (www)

  @article{ , abstract = {Deep neural networks have shown significant advancements in modelling complex non-linear relationships in high-dimensional biomedical data. Understanding the interplay between genetic variants and disease susceptibility is still a considerable challenge that prevents certain genomic diseases to be predicted accurately for clinical interventions. In this study, we introduce the Extensive Multi-Variant Deep Neural Network (EMV-DNN), an innovative deep learning methodology designed to enhance polygenic risk prediction. Unlike conventional polygenic risk score methods, EMV-DNN incorporates single nucleotide polymorphisms (SNPs) alongside structural variants including insertions and deletions (indels), short tandem repeats (STRs), and copy number variants (CNVs) using variant-specific subnetworks to extract informative embeddings which capture a richer and holistic genomic context. Evaluated on real-world cohorts from the UK Biobank and All of Us, EMV-DNN outperforms conventional PRS methods and classic machine learning algorithms across binary and multi-class prediction tasks. Beyond predictive performance, SHapley Additive exPlanations (SHAP) analysis revealed biologically plausible variant–gene–disease associations, highlighting pathways related to endometrial cell proliferation, fibrosis, and immune regulation. Our findings underscore the value of multi-variant integration and non-linear approaches to capture the intricate genetic architecture of complex genomic diseases. Despite challenges such as dataset limitations and the complexity of diseases with multiple contributing factors, the EMV-DNN methodology presents a promising avenue for enhancing the predictive accuracy of PRS, thereby facilitating personalized healthcare interventions and advancing our understanding of genetic predispositions to disease.}, archive = {J_NEUCOM}, author = {Zelia Soo and Hua Lin and Yue Yang and Mark Grosser and Mengjia Wu and Yi Zhang and Jie Lu}, doi = {10.1016/j.neucom.2025.131496}, journal = {Neurocomputing}, month = {12}, pages = {131496}, shortjournal = {Neurocomputing}, title = {An extensive multi-variant deep neural network approach to enhance genomic prediction of endometriosis}, volume = {656}, year = {2025}, }
• (2025). A self-supervised facial expression restoration and recognition method based on improved masked auto-encoder. NEUCOM, 656, 131495. (www)

  @article{ , abstract = {This study proposes a method for reconstructing occluded facial expressions. Firstly, a self-supervised learning Masked Auto-Encoder based facial expression recognition (MAE-FER) method is introduced, which effectively reduces the computational cost and parameter count by enhancing the multi-scale local-global self-attention interaction encoder, thereby improving the training efficiency and generalization capability of the model. Secondly, to address the problem of facial expression occlusion in real-world scenarios, a MAE-based occlusion detector is designed to detect occluded parts of the face, providing effective support for subsequent reconstruction tasks. Subsequently, the Dynamic Weight Allocation Generative Adversarial Network (DWA-GAN) for facial expression occlusion recovery is proposed, which achieves precise occlusion recovery by dynamically allocating weights to reference image blocks, significantly improving the accuracy of reconstruction. Finally, feature fusion is performed on the reconstructed results and applied to the FER task to further enhance classification accuracy and stability. Utilizing the pre-trained MAE-FER model, key hidden vectors are extracted from facial expression images, containing important feature information related to expression recognition. Through this step, closely related features to expression recognition are selected while irrelevant details are discarded, optimizing the inter-class distance issue of facial expressions. Next, to address the performance degradation caused by label ambiguity, an improved Rotate Erasing Attention Consistency (REAC) method is adopted, which effectively mitigates the negative impact of label ambiguity, further improving the accuracy and stability of FER. Experimental results demonstrate that the method achieves the best performance on the RAF-DB dataset.}, archive = {J_NEUCOM}, author = {Chaolong Zhang and Yuanping Xu and Zhijie Xu and Rongqiang Gou and Weiye Wang and Jin Jin and Jian Huang}, doi = {10.1016/j.neucom.2025.131495}, journal = {Neurocomputing}, month = {12}, pages = {131495}, shortjournal = {Neurocomputing}, title = {A self-supervised facial expression restoration and recognition method based on improved masked auto-encoder}, volume = {656}, year = {2025}, }
• (2025). TemPrompt: Multi-task prompt learning for temporal relation extraction in RAG-based crowdsourcing systems. NEUCOM, 656, 131494. (www)

  @article{ , abstract = {Temporal relation extraction (TRE) aims to grasp the evolution of events or actions, and thus shape the workflow of associated tasks, so it is recognized as a pivotal technology for facilitating the rational scheduling and efficient execution of crowdsourcing tasks. However, existing methods still struggle with limited and unevenly distributed annotated data. Inspired by the abundant global knowledge stored within pre-trained language models (PLMs), some studies have explored using prompts to guide PLMs in completing TRE, but their improvements are still unsatisfactory, as the model treats all the tokens equally, resulting in a limited understanding of temporal order. In this paper, we propose a multi-task prompt learning framework for TRE (TemPrompt), incorporating prompt tuning and contrastive learning to tackle these issues. In the framework, we design temporal event reasoning in the form of masked language modeling as auxiliary tasks to enable the PLM to distinguish tokens essential for temporal reasoning from those serving general contextual purposes, thereby fostering the model’s comprehension of temporal knowledge. Additionally, to elicit more effective prompts for PLMs, we introduce a task-oriented prompt construction approach that thoroughly takes the myriad factors of TRE into consideration for the automatic generation of high-quality and easy-to-interpret prompts. Contrastive learning is employed to further mitigate data issues by extracting more distinctive sample representations. Experimental results demonstrate that TemPrompt outperforms all baseline methods across most metrics and exhibits strong generalization to unseen events. Two case studies—one on designing and manufacturing printed circuit boards and the other on developing defect detection systems—are provided to validate its feasibility and effectiveness in crowdsourcing scenarios.}, archive = {J_NEUCOM}, author = {Jing Yang and Yu Zhao and Linyao Yang and Xiao Wang and Long Chen and Fei-Yue Wang}, doi = {10.1016/j.neucom.2025.131494}, journal = {Neurocomputing}, month = {12}, pages = {131494}, shortjournal = {Neurocomputing}, title = {TemPrompt: Multi-task prompt learning for temporal relation extraction in RAG-based crowdsourcing systems}, volume = {656}, year = {2025}, }
• (2025). EHGFL: Contrastive distillation for efficient heterogeneous graph few-shot learning. NEUCOM, 656, 131493. (www)

  @article{ , abstract = {Heterogeneous graphs (HGs), as a general modeling paradigm for multi-typed entities and complex interactions in ubiquitous real-world networks, have attracted extensive research enthusiasm. While self-supervised learning (SSL) on heterogeneous graph neural networks (HGNNs) demonstrates promising performance, existing SSL approaches face critical limitations: (1) The inherent multiple relation types and meta-path aggregations in HGNNs create prohibitive training and inference costs that restrict scalability; (2) The local message-passing paradigm confines information propagation to immediate neighborhoods, limiting the model’s ability to capture long-range dependencies and global structural patterns essential for complex reasoning tasks; (3) Task specialization necessitates costly fine-tuning of the HGNN backbones. To address these issues, we propose Efficient Heterogeneous Graph Few-shot Learning (EHGFL) to improve HGNNs’ scalability and global-structure modeling capabilities. Specifically, EHGFL first employs instance discrimination contrastive learning for self-supervised pretraining of HGNNs. To enhance efficiency, we introduce a novel cross-model contrastive distillation mechanism that transfers HGNNs’ heterogeneous structure modeling ability to a concise, globally-structure-aware multilayer perceptron. This feature-space distillation process preserves heterogeneous structure representations while avoiding expensive neighborhood aggregation and enhancing global feature awareness. Furthermore, to bridge the gap between pretraining objectives and downstream tasks, we adopt prompt tuning techniques specifically designed for the student model, enabling effective adaptation with limited labeled data. Extensive experiments on two real-world HG datasets demonstrate that the proposed EHGFL framework substantially accelerates training and inference while achieving superior few-shot node classification accuracy compared to state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Ning Ruan and Yufei Zeng and Huan Liu and Dong Liu and Pengfei Jiao}, doi = {10.1016/j.neucom.2025.131493}, journal = {Neurocomputing}, month = {12}, pages = {131493}, shortjournal = {Neurocomputing}, title = {EHGFL: Contrastive distillation for efficient heterogeneous graph few-shot learning}, volume = {656}, year = {2025}, }
• (2025). Scattered data augmentation for generalization in visual reinforcement learning. NEUCOM, 656, 131492. (www)

  @article{ , abstract = {Data augmentation (DA) has shown a significant potential to enhance generalization performance in visual reinforcement learning (VRL). However, existing research on DA-based methods is predominantly empirical, and the mechanism for why DA enhances generalization remains theoretically under-explored. To bridge this gap, we derive a generalization error upper bound for VRL from the perspective of data distribution distance. Based on this bound, we provide a theoretical explanation of the mechanism by which DA improves generalization: we find that DA that satisfies certain conditions can reduce the distance between the training and test distributions, thus making the training and test samples closer. In addition, we conditionally prove that training data with higher variance can provide a higher generalization performance. Motivated by our analysis, we propose Scattered Data Augmentation (ScDA) framework. ScDA constructs a data transformation system with the agent serving as the discriminator, aiming to provide more diverse training data for agent training. Experiments are conducted across various tasks and numerous test modes in DeepMind Control Generalization Benchmark2 (DMC-GB2) and robotic tasks. Results demonstrate that our ScDA framework can be integrated with different baseline algorithms and significantly enhance policy generalization, outperforming the current state-of-the-art methods in the DMC-GB2 tests, confirming the effectiveness of the theoretical analysis in this work. The code for this work can be found at: https://github.com/scdadev/scdadev .}, archive = {J_NEUCOM}, author = {Hao Lei and Yu Zhao and Yi Xin and Zhang Shaonan and Ke Liangjun}, doi = {10.1016/j.neucom.2025.131492}, journal = {Neurocomputing}, month = {12}, pages = {131492}, shortjournal = {Neurocomputing}, title = {Scattered data augmentation for generalization in visual reinforcement learning}, volume = {656}, year = {2025}, }
• (2025). Monotonic convergence of adaptive caputo fractional gradient descent for temporal convolutional networks. NEUCOM, 656, 131491. (www)

  @article{ , abstract = {Fractional derivatives have gained prominence in optimization for their inherent non-locality and memory-dependent properties, effectively capturing historical dependencies. This work introduces an Adaptive Fractional-order Gradient Descent (AFGD) algorithm based on Caputo fractional derivatives, with deep integration into Temporal Convolutional Networks (TCNs). Unlike conventional fixed-order methods, AFGD employs an adaptive fractional-order mechanism to enhance optimization. Theoretically, we establish rigorous proofs for AFGD’s monotonic convergence in loss function minimization, supported by numerical simulations of its convergence behavior. Evaluated on the MIT-BIH arrhythmia five-class classification benchmark, TCNs optimized with AFGD achieve superior accuracy over established methods, demonstrating the efficacy of the proposed gradient scheme for deep learning optimization.}, archive = {J_NEUCOM}, author = {Zhiwei Xiao and Jiejie Chen and Xuewen Zhou and Bin Wei and Ping Jiang and Zhigang Zeng}, doi = {10.1016/j.neucom.2025.131491}, journal = {Neurocomputing}, month = {12}, pages = {131491}, shortjournal = {Neurocomputing}, title = {Monotonic convergence of adaptive caputo fractional gradient descent for temporal convolutional networks}, volume = {656}, year = {2025}, }
• (2025). An adaptive PID-guided tensor wheel decomposition model for dynamic weighted network representation. NEUCOM, 656, 131490. (www)

  @article{ , abstract = {Dynamic Weighted Networks (DWN) usually appear in various big data-related complex systems and can describe real-time interactions between a large number of entities. As the number of entities increases dramatically, it is impossible for each entity to have complete interaction with each other, which results in such a DWN being High-Dimensional and Incomplete (HDI). Tensor Wheel Decomposition (TWD), as a novel tensor network, has powerful representation capabilities, but existing TWD-based methods require additional computational and storage costs to process an HDI DWN. To address these challenges, we propose an Adaptive integral separation PID–guided Tensor Wheel Decomposition (APTWD) model that: 1) employs a data density-oriented loss function, ensuring the representation learning is focused on the existing information in the target network to obtain more accurate low-rank embedding; and 2) develops a parameter learning scheme with error control feedback based on the integral separation PID controller to minimize the convergence iteration process. Experiments on six real-world DWN datasets demonstrate that APTWD consistently outperforms state-of-the-art methods, delivering higher representation accuracy and significantly reduced computational cost.}, archive = {J_NEUCOM}, author = {Jiqiu Chen and Qu Wang and Hao Wu}, doi = {10.1016/j.neucom.2025.131490}, journal = {Neurocomputing}, month = {12}, pages = {131490}, shortjournal = {Neurocomputing}, title = {An adaptive PID-guided tensor wheel decomposition model for dynamic weighted network representation}, volume = {656}, year = {2025}, }
• (2025). Multi-view unsupervised feature selection based on graph discrepancy learning. NEUCOM, 656, 131487. (www)

  @article{ , abstract = {In multi-view learning, unsupervised feature selection plays a vital role in reducing dimensionality while preserving discriminative information distributed across diverse data modalities. Despite notable progress, existing approaches frequently exhibit two key limitations: they often overlook the complementary benefits of integrating global and local structural information, and they inadequately model complex nonlinear relationships or align structural representations across views. To address these challenges, we propose a novel framework, termed Multi-view unsupervised feature selection based on graph discrepancy learning (GDFS). The proposed method jointly constructs global graph structures in a projected low-dimensional space and local graphs in a nonlinear kernel-induced space, effectively capturing both high-level semantic structures and fine-grained neighborhood dependencies. A graph discrepancy term is introduced to explicitly reduce structural discrepancies between global and local representations, thus enhancing consistency and robustness. In addition, a low-rank tensor constraint is applied to the stack of global graphs to uncover high-order correlations across views. A consensus clustering matrix is further learned to provide pseudo-label supervision, which guides the selection of discriminative features. Extensive experiments on six benchmark multi-view datasets demonstrate that GDFS consistently surpasses state-of-the-art methods in terms of clustering performance, thereby confirming its effectiveness, scalability, and generalizability. The code is available at https://github.com/xyw0111/2025-GDFS .}, archive = {J_NEUCOM}, author = {Yiwan Xu and Xijiong Xie and Xianliang Jiang and Yujie Xiong}, doi = {10.1016/j.neucom.2025.131487}, journal = {Neurocomputing}, month = {12}, pages = {131487}, shortjournal = {Neurocomputing}, title = {Multi-view unsupervised feature selection based on graph discrepancy learning}, volume = {656}, year = {2025}, }
• (2025). A novel mixed-attribute data anomaly detection method based on granular-ball multi-kernel fuzzy rough sets. NEUCOM, 656, 131486. (www)

  @article{ , abstract = {Fuzzy rough sets have established a novel approach to anomaly detection through uncertainty handling. Nevertheless, traditional approaches are susceptible to noise. Existing kernelized methods improve feature representation via kernel transformations. However, they are typically restricted to a single-kernel framework, which limits the capacity to model the heterogeneous nature of mixed-attribute data. To address this issue, this study proposes a granular-ball generation algorithm tailored to the characteristics of mixed-attribute data. Multiple kernel functions are employed to effectively integrate the fuzzy relations of various attribute types. By integrating fuzzy rough set theory, granular-ball computing, and multi-kernel methods, a granular-ball multi-kernel fuzzy rough set model is proposed. Besides, a novel unsupervised anomaly detection method is proposed to effectively process mixed-attribute data. This method integrates kernelized fuzzy relations across various attribute types, constructs kernelized fuzzy information granules, and computes anomaly scores based on multiple granular-ball kernelized fuzzy information granules. Finally, an anomaly factor is introduced to quantify the anomaly degree of data objects. Comparative experiments were conducted on 16 public datasets. The novel approach consistently outperformed current methodologies in AUC metrics while demonstrating superior robustness across diverse data samples.}, archive = {J_NEUCOM}, author = {Cong Gao and Qiu Wang and Yanping Chen and Qingqi Pei and Zhongmin Wang}, doi = {10.1016/j.neucom.2025.131486}, journal = {Neurocomputing}, month = {12}, pages = {131486}, shortjournal = {Neurocomputing}, title = {A novel mixed-attribute data anomaly detection method based on granular-ball multi-kernel fuzzy rough sets}, volume = {656}, year = {2025}, }
• (2025). Knowledge-guided encoder-decoder network for soft sensor of copper matte grade in flash smelting process. NEUCOM, 656, 131485. (www)

  @article{ , abstract = {Accurate real-time detection of copper matte grade is critical for state identification and optimization control in flash smelting, yet remains challenging due to the complex and harsh industrial environment. To address this issue, this study proposes a knowledge-guided encoder-decoder network. In this method, Bidirectional Gated Recurrent Unit serves as the backbone architecture for both the encoding and decoding processes, enabling nonlinear dynamic modeling in the temporal domain. The encoder integrates a composite variable attention mechanism, which leverages process knowledge to prioritize key variables based on their importance. A temporal decay attention mechanism is added to the decoder, endowing the model with the ability to simulate the temporal dependency between copper matte grade and process variables through prior knowledge. These knowledge-guided designs strengthen the ability of model to capture process-specific relationships between input variables and copper matte grade. Industrial experiments based on real production data from a smelting plant in China, show that the proposed model achieves optimal performance, with 96 % absolute errors not exceeding 0.5 %. It demonstrates that the proposed model not only provides accurate and real-time copper matte grade estimation but also maintains robustness in industrial environments, verifying its potential for practical application in flash smelting process.}, archive = {J_NEUCOM}, author = {Zhou Zou and Can Zhou and Chunhua Yang}, doi = {10.1016/j.neucom.2025.131485}, journal = {Neurocomputing}, month = {12}, pages = {131485}, shortjournal = {Neurocomputing}, title = {Knowledge-guided encoder-decoder network for soft sensor of copper matte grade in flash smelting process}, volume = {656}, year = {2025}, }
• (2025). Multi-resolution QP-adaptive generative face video compression using multi-level generator. NEUCOM, 656, 131484. (www)

  @article{ , abstract = {In this paper, we propose a multi-resolution quantization parameter (QP)-adaptive generative face video compression (GFVC) framework to realize face video communication at an ultra-low bitrate. By leveraging deep generative models and semantic feature representation, the proposed framework achieves high perceptual quality while significantly reducing bitrate. The proposed framework dynamically adjusts feature granularity based on QP values and integrates modules such as multi-level multi-DConv head transposed attention (MDTA) and multi-level spatially adaptive denormalization (SPADE) to enhance both spatial fidelity and temporal consistency. To ensure adaptability and standardization, we further extend the proposed framework to support multi-resolution inputs and incorporate feature encoding based on Supplemental Enhancement Information (SEI) in VVC. Specifically, we introduce the flag gfv_enhancement_matrix_flag to transmit an optional 8 × 8 enhancement matrix, enabling precise refinement of inter-frame reconstruction in compliance with VVC. A multi-reference frame buffer mechanism is also implemented to improve long-term temporal coherence through attention-guided reference selection. Experimental results demonstrate that the proposed GFVC framework achieves average BD-rate gains of 63.87 % in DISTS and 61.99 % in LPIPS on benchmark datasets compared to the VVC anchor (VTM-22.2 LDB mode). Without retraining, the proposed framework operates smoothly even on face videos with a resolution of 512 × 512 , achieving 23.40 % BD-rate gain in DISTS and indicating strong scalability. These results validate the practical feasibility of the proposed GFVC framework in real-world video conferencing and telepresence scenarios, especially under ultra-low bandwidth conditions.}, archive = {J_NEUCOM}, author = {Wenbo Kang and Lu Liu and Cheolkon Jung}, doi = {10.1016/j.neucom.2025.131484}, journal = {Neurocomputing}, month = {12}, pages = {131484}, shortjournal = {Neurocomputing}, title = {Multi-resolution QP-adaptive generative face video compression using multi-level generator}, volume = {656}, year = {2025}, }
• (2025). ST-CML: A contrastive meta learning framework for spatio-temporal graph few-shot learning with cross-city transfer. NEUCOM, 656, 131483. (www)

  @article{ , abstract = {Spatio-temporal graph learning is a critical methodology for addressing smart city computational tasks, including traffic speed prediction, vehicle trajectory analysis, and air quality forecasting. However, the scarcity of available data in many cities poses a significant constraint due to the challenges and costs associated with large-scale data collection. To overcome this limitation, leveraging data from cities with abundant data resources and enhancing model performance through knowledge transfer techniques is essential. In this study, we propose ST-CML. Our approach utilizes data from multiple source cities to train a meta-learner, which generates parameters for downstream spatio-temporal models. The meta-learner is then fine-tuned to adapt to target cities with limited data. We introduce a graph contrastive loss to improve the learning of spatio-temporal structures during meta-learning. This loss function guides the model in mitigating spatio-temporal feature deviations across cities during knowledge transfer. Experimental evaluations on real-world traffic datasets demonstrate that our framework surpasses existing methods and achieves significant effectiveness in cross-city few-shot learning scenarios.}, archive = {J_NEUCOM}, author = {Haichen Lyu and Chun Wang}, doi = {10.1016/j.neucom.2025.131483}, journal = {Neurocomputing}, month = {12}, pages = {131483}, shortjournal = {Neurocomputing}, title = {ST-CML: A contrastive meta learning framework for spatio-temporal graph few-shot learning with cross-city transfer}, volume = {656}, year = {2025}, }
• (2025). Filter bank convolutional network with dual channel attention for multi-class fine joint motor imagery decoding. NEUCOM, 656, 131480. (www)

  @article{ , abstract = {Objective Motor imagery brain-computer interface (MI-BCI) is a representative BCI system. Recent studies in MI-BCI focus on Fine Joint MI (FJMI) decoding that recognizes the motor intention of different joints from one upper limb. However, due to the low spatial difference between EEG patterns of FJMI, achieving optimal performance remains a challenge of multi-class FJMI decoding studies. Methods: We proposed a novel approach named Filter Bank Convolutional Network with Dual Channel Attention (FB-DCANet) that enables feature extraction and selection in MI-EEG across multi-class FJMI tasks. This network features a combined filter bank in frequency and time domain that simultaneously extracts spatio-temporal information from four frequency bands (alpha, beta, theta, and low gamma), accompanied with temporal convolutional modules for additional temporal information extraction. Moreover, a feature selection method based on Dual Channel Attention was proposed combining preliminary intra-band feature selection via Residual Channel Self-Attention (RCSA) and further inter-band feature selection from different frequency bands by Efficient Channel Attention (ECA). Results: We performed experiments using FJMI-EEG data from the unilateral upper limb, and FB-DCANet achieved an accuracy of 59.34 % in a 4-class classification scenario (hand MI, elbow MI, shoulder MI, and resting state), and interpretability of FB-DCANet was analyzed by visualization of Class Activation Map (CAM) and attention values. Conclusion and Significance: This work presents a novel approach with a time-frequency filter bank and Dual Channel Attention-based feature selection for multi-class FJMI decoding, which can be utilized to develop a rehabilitation system based on FJMI-BCI.}, archive = {J_NEUCOM}, author = {Jiaming Chen and Yueqi Zhang and Kaide Liu and Xinkang Hu and Meng Xu and Dan Wang and Weibo Yi}, doi = {10.1016/j.neucom.2025.131480}, journal = {Neurocomputing}, month = {12}, pages = {131480}, shortjournal = {Neurocomputing}, title = {Filter bank convolutional network with dual channel attention for multi-class fine joint motor imagery decoding}, volume = {656}, year = {2025}, }
• (2025). Output-feedback synchronization of semi-markov jump two-time-scale neural networks: Dual event-triggered scheme. NEUCOM, 656, 131479. (www)

  @article{ , abstract = {This paper addresses the problem of synchronization control for semi-Markov jump two-time-scale neural networks, in which the output-feedback mechanism is adopted and a dual event-triggered scheme is employed using a double-rate sampling method to balance system performance and communication efficiency. First, considering the two-time-scale property of the semi-Markov jump neural networks, the dual-rate sampling strategy is adopted such that two independent event-triggered conditions for different time scales can be designed, which ensure efficient resource utilization while maintaining system performance. Then, a Lyapunov–Krasovskii functional with the singular perturbation parameter is constructed to deduce sufficient conditions ensuring that the synchronization error system is stochastically stable and satisfies a given H ∞ performance index. Moreover, the solution for obtaining the controller gains is presented to guarantee synchronization of the considered system under a dual event-triggered scheme. Finally, the feasibility of the methods is demonstrated by two examples, including a numerical example and an image encryption. They show that this event-triggered mechanism provides an efficient new synchronization control scheme for semi-Markov jump two-time-scale neural network systems while reducing the network burden.}, archive = {J_NEUCOM}, author = {Wenyan Zuo and Ya-Nan Wang and Feng Li and Sangmoon Lee}, doi = {10.1016/j.neucom.2025.131479}, journal = {Neurocomputing}, month = {12}, pages = {131479}, shortjournal = {Neurocomputing}, title = {Output-feedback synchronization of semi-markov jump two-time-scale neural networks: Dual event-triggered scheme}, volume = {656}, year = {2025}, }
• (2025). Efficient multi-view graph condensation via gradient-flow induced graph convolutional networks. NEUCOM, 656, 131478. (www)

  @article{ , abstract = {Multi-view learning based on graph convolutional networks boosts performance by incorporating diverse perspectives, leading to significant achievements and successful applications across various academic and practical fields. However, multi-view graph convolutional networks suffer from substantial computational challenges on large-scale graphs. To address this limitation, graph condensation has emerged as a promising direction by creating a smaller composite graph that allows for efficient network training while preserving performance. Furthermore, previous studies have demonstrated that encouraging performance in graph learning is achieved via graph compression. To this end, we attempt to introduce graph condensation into the multi-view learning for computation acceleration. This approach not only reduces training costs significantly but also achieves sub-linear time complexity and memory consumption during network training. Further, we propose a gradient flow induced graph convolutional network from partial differential equations, which offers theoretical guarantees and potential new insights for the graph-related network architecture construction with transparent model interpretability. Extensive experiments on seven real-world multi-view datasets demonstrate that the proposed method sharply decreases model training time while ensuring competitive multi-view semi-supervised classification.}, archive = {J_NEUCOM}, author = {Lu Liu and Yang Huang and Yueyang Pi and Zhicheng Wei and Jinbo Li and Shiping Wang}, doi = {10.1016/j.neucom.2025.131478}, journal = {Neurocomputing}, month = {12}, pages = {131478}, shortjournal = {Neurocomputing}, title = {Efficient multi-view graph condensation via gradient-flow induced graph convolutional networks}, volume = {656}, year = {2025}, }
• (2025). Optimal event-triggered control for multi-agent systems with hierarchical framework. NEUCOM, 656, 131477. (www)

  @article{ , abstract = {This study investigates the event-triggered optimal control problem for a class of linear second-order multi-agent systems (MASs) with external disturbances. A hierarchical framework is proposed to address the challenges that arise from the information of the coupled neighbors and external disturbances, integrating the communication, learning, and control layers. Specifically, the communication layer utilizes event-triggered mechanisms (ETMs) to transmit neighbor information, facilitating virtual consensus. The learning layer connects the communication and control layers, employing reinforcement learning (RL) to optimize tracking control with ETMs. The control layer achieves real consensus by aligning the agent states with the processed information from the communication layer. Moreover, this framework effectively mitigates the effects of coupled neighbor information on the controller and suppresses the transmission of external disturbances through the communication network. Finally, two simulation examples are used to verify the anti-interference of the hierarchical framework i.e., it’s still possible to achieve consensus after being disturbed and the effectiveness of considering the reinforcement learning layer via event-triggered mechanism which reduces the communication and learning burden to achieve optimal control.}, archive = {J_NEUCOM}, author = {Denghao Pang and Yechen Guo and Jinde Cao and Boxiang Li and Xiao-Wen Zhao}, doi = {10.1016/j.neucom.2025.131477}, journal = {Neurocomputing}, month = {12}, pages = {131477}, shortjournal = {Neurocomputing}, title = {Optimal event-triggered control for multi-agent systems with hierarchical framework}, volume = {656}, year = {2025}, }
• (2025). Optimal bayesian regression with a non-informative prior: Order-recursive and incremental learning. NEUCOM, 656, 131476. (www)

  @article{ , abstract = {Optimal Bayesian regression (OBR) for data generated from a multidimensional vector autoregressive process of order p , denoted as VAR ( p ) , has a closed-form analytic expression that has been previously obtained. Despite the closed-form expressions to compute the “OBR-VAR”, in certain practical scenarios the computational cost involved in training OBR-VAR is a bottleneck. From a computational perspective, two common scenarios that incur excessive computational cost are: 1) given a set of training data, estimating the unknown model order p generally entails computing the OBR-VAR from scratch for every p in a range from 1 to a maximum value; and 2) in dynamic environments where data arrives sequentially, currently one must recompute OBR-VAR from scratch for every new upcoming observation. To address the first issue, in this paper, an order-recursive OBR-VAR regressor using QR decomposition is proposed. This method efficiently updates the regressor without recalculating it from scratch for each p , significantly reducing computational complexity while preserving model accuracy. Analytical results demonstrate that the proposed order-recursive method achieves a computational complexity reduction by a factor proportional to p , making it scalable to larger datasets and higher model orders. To address the second issue, an incremental version of the OBR-VAR algorithm is developed for real-time data processing. This method updates the regressor incrementally as new data points arrive, maintaining accuracy without the need for costly recomputation of key matrices. Its capability makes it well-suited for continuous-time data acquisition and streaming applications, where timely and accurate responses are critical. In all cases we assume an improper non-informative prior to model the case of having no prior knowledge about the problem. Theoretical analysis and empirical evaluations using synthetic and real datasets demonstrate that both methods significantly outperform the standard OBR-VAR algorithm in terms of computational complexity while preserving accuracy.}, archive = {J_NEUCOM}, author = {Samira Reihanian and Amin Zollanvari and Siamac Fazli}, doi = {10.1016/j.neucom.2025.131476}, journal = {Neurocomputing}, month = {12}, pages = {131476}, shortjournal = {Neurocomputing}, title = {Optimal bayesian regression with a non-informative prior: Order-recursive and incremental learning}, volume = {656}, year = {2025}, }
• (2025). A dual-stream GCN-based action recognition framework using trustworthy fusion decision from different skeleton descriptors. NEUCOM, 656, 131475. (www)

  @article{ , abstract = {In complex motion processes, different human skeleton descriptors can characterize skeletal features across various dimensions. The frequency of spatiotemporal changes in different joints is largely influenced by the type of action. This paper presents a dual-stream GCN-based action recognition framework, which involves Slow-stream and Fast-stream networks to process skeletal features of different spatiotemporal change characteristics. In the parallel processing architecture of graph convolutional layers, adaptive adjacency matrices that strengthen spatial and temporal feature extraction are proposed to learn the implicit relationships between skeletal joints. Furthermore, different skeletal features have significantly varying impacts on the accuracy of action recognition. The Dirichlet distribution and an optimized Dempster combination rule are introduced for trustworthy decision when fusing multi-branch opinions obtained from different skeleton descriptors. Extensive experiments on three authoritative datasets demonstrate that the proposed method achieves state-of-the-art performance while reducing uncertainty in action recognition.}, archive = {J_NEUCOM}, author = {Wenrui Zhu and Donghui Shi and Junqi Yu}, doi = {10.1016/j.neucom.2025.131475}, journal = {Neurocomputing}, month = {12}, pages = {131475}, shortjournal = {Neurocomputing}, title = {A dual-stream GCN-based action recognition framework using trustworthy fusion decision from different skeleton descriptors}, volume = {656}, year = {2025}, }
• (2025). Bipartite tracking consensus for fractional-order nonlinear multiagent systems with sampled-data and input saturation. NEUCOM, 656, 131472. (www)

  @article{ , abstract = {This study focuses on the sampled-data bipartite tracking consensus (BTC) issue for a type of fractional-order (FO) nonlinear multiagent systems (FOMASs) subject to input saturation. In the network under consideration, agents exhibit both competitive (CM) and cooperative (CO) interactions simultaneously. By employing Lyapunov stability theory, the FO Halanay-type Inequality , and the linear matrix inequality (LMI) approach, several criteria are derived to guarantee that the considered MASs can attain the BTC. Moreover, by utilizing the matrix decomposition (MD) approach, the dimensions of matrix inequalities are significantly reduced, which helps alleviate computational complexity. As a result, the derived results can be effectively applied to large-scale FOMASs. Also, the controller gain matrix is clearly represented based on the solutions of a series of matrix inequalities. Besides, we present a method for estimating the maximum attraction region of BTC. Ultimately, numerical simulation is employed to substantiate our theoretical analysis.}, archive = {J_NEUCOM}, author = {Zhi Qiao and Luyang Yu and Yuman Li and Yurong Liu}, doi = {10.1016/j.neucom.2025.131472}, journal = {Neurocomputing}, month = {12}, pages = {131472}, shortjournal = {Neurocomputing}, title = {Bipartite tracking consensus for fractional-order nonlinear multiagent systems with sampled-data and input saturation}, volume = {656}, year = {2025}, }
• (2025). Heterogeneous federated semantic segmentation. NEUCOM, 656, 131470. (www)

  @article{ , abstract = {Federated Learning (FL) offers a promising solution for semantic segmentation in scenarios involving data distribution across isolated clients. Despite recent advances, federated semantic segmentation (FSS) continues to face key challenges. One major issue is the shift from centralized to decentralized training, where diverse and limited local data hinder consistent pixel-level representation learning. Another challenge is data heterogeneity from imbalanced class distributions across clients, which weakens feature consistency and degrades global performance. These limitations often lead to inconsistent feature learning and degraded global performance. To address the challenges of class heterogeneity and insufficient pixel-level representation learning in FSS, we propose a novel pixel-aware FSS framework that improves local adaptation and semantic consistency. Specifically, we design a fine-tuning strategy that initializes each client with a lightweight pre-trained model and performs local updates over multiple epochs. This improves model adaptability to local distributions while reducing communication overhead. To further enhance semantic consistency across heterogeneous clients, we introduce a client clustering strategy based on pixel-level semantic features. Clients with similar class distributions are grouped to encourage consistent feature learning within clusters. Cluster-level training and aggregation are then followed by a global aggregation step, promoting more robust and aligned semantic understanding. Empirical evaluation across multiple benchmark datasets confirms that our method achieves consistently high segmentation precision and enhanced model adaptability in highly heterogeneous federated scenarios.}, archive = {J_NEUCOM}, author = {Chen Zhang and Jiarui Wang and Yu Xie and Xinlei Wang and Bin Yu}, doi = {10.1016/j.neucom.2025.131470}, journal = {Neurocomputing}, month = {12}, pages = {131470}, shortjournal = {Neurocomputing}, title = {Heterogeneous federated semantic segmentation}, volume = {656}, year = {2025}, }
• (2025). Generalized containment control for delayed fractional-order nonlinear multi-agent systems with unknown disturbances. NEUCOM, 656, 131469. (www)

  @article{ , abstract = {In this article, the issue of generalized containment control is explored for a class of delayed fractional-order nonlinear multi-agent systems (MASs) with unknown disturbances. It is assumed that the MAS under consideration has multiple dynamic leaders, and its dynamics are governed by fractional-order differential equations, and suffers from the unknown but norm-bounded external disturbances. Also, the directed graph of the MAS is assumed to have a united directed spanning tree. Furthermore, for the sake of saving communication resources, an event-triggered mechanism is introduced to regulate the signal transmission. In the presence of the external disturbances, the generalized containment control is analyzed by means of Lyapunov stability theory, algebraic graph theory, Halanay-type inequality, etc., and sufficient conditions are established to ensure that all followers ultimately enter a certain neighborhood of the convex hull formed by the leaders. In the meanwhile, it is also proven that the Zeno phenomenon can be excluded for the concerned MAS. Finally, a numerical simulation is presented to further illustrate the effectiveness of the theoretical results.}, archive = {J_NEUCOM}, author = {Zhi Qiao and Luyang Yu and Hong Lin and Yurong Liu}, doi = {10.1016/j.neucom.2025.131469}, journal = {Neurocomputing}, month = {12}, pages = {131469}, shortjournal = {Neurocomputing}, title = {Generalized containment control for delayed fractional-order nonlinear multi-agent systems with unknown disturbances}, volume = {656}, year = {2025}, }
• (2025). CBH-YOLO: A steel surface defect detection algorithm based on cross-stage mamba enhancement and hierarchical semantic graph fusion. NEUCOM, 656, 131467. (www)

  @article{ , abstract = {The detection of surface defects in steel materials plays a pivotal role in ensuring industrial production quality and operational safety. However, existing deep learning-based detection methods face significant challenges in steel surface defect detection, including limited receptive field coverage, inadequate multi-scale feature integration, and insufficient feature discrimination under complex backgrounds. To address these limitations, this work introduces CBH-YOLO, a novel steel surface defect detection algorithm. The proposed framework incorporates three fundamental innovation modules: (1) Cross-stage Mamba-Enhanced Multi-scale Convolution (CMMC) module, which synergistically combines the advantages of state space models with large kernel convolutions alongside adaptive attention mechanisms, substantially expanding receptive field coverage while enhancing multi-scale feature extraction capabilities; (2) Binary Amplification Matrix (BAM) module, which innovatively integrates FlexWave (FXW) dynamic activation functions with OmniScale (OSC) multi-scale perception mechanisms to achieve adaptive nonlinear feature mapping and refined representation; (3) Hierarchical Semantic Graph Fusion Network (HSGFN), which models high-order correlations among features through hypergraph structures combined with adaptive feature fusion mechanisms, enabling more effective multi-scale feature integration. Comprehensive experimental validation on NEU-DET and GC10-DET benchmark datasets demonstrates that CBH-YOLO achieves improvements of 2.7 % and 3.2 % in mAP@50 metrics compared to the baseline YOLOv11 model, while maintaining exceptional computational efficiency. This research provides a high-precision, high-efficiency solution for steel surface defect detection, offering significant theoretical value and practical application prospects.}, archive = {J_NEUCOM}, author = {Bo Gao and Jingcheng Tong and RongRong Fu and ZhenZhen Zhang and YiLin Yuan}, doi = {10.1016/j.neucom.2025.131467}, journal = {Neurocomputing}, month = {12}, pages = {131467}, shortjournal = {Neurocomputing}, title = {CBH-YOLO: A steel surface defect detection algorithm based on cross-stage mamba enhancement and hierarchical semantic graph fusion}, volume = {656}, year = {2025}, }
• (2025). DMNet: Incomplete multi-modal alzheimer’s disease diagnosis with dual memory network. NEUCOM, 656, 131465. (www)

  @article{ , abstract = {Due to the challenge of missing modality, existing multi-modal learning methods become impractical and missing modality is a serious impediment to a good multi-modal learning performance. Meanwhile, we note that the existing methods for addressing missing modality issue tend to explore complete information by using either cross-generative approaches via simply filling in missing modality data, and do not consider the specific information, resulting in a sub-optimal performance for Alzheimer’s Disease diagnosis with multi-modal data. To address this problem, we propose a novel Dual Memory Network (DMNet) that comprises the Tabular Alignment Memory bank (TAM) and Dynamic Re-optimizing Memory bank (DRM) to complement the missing modality information in multi-modal learning for Alzheimer’s disease diagnosis. More specifically, TAM stores the information aligned with clinical tabular data, and maintains the feature distribution alignment between clinical tabular data and imaging modalities. Besides, TAM is updated by a memory aligning strategy. Then, DRM stores modality specific information from complete modalities, and we design a memory optimizing strategy that incorporates Feature Consistency (FC) loss and Memory Correspondence (MC) loss to update the memory items in DRM to effectively represent specific information of modalities. This novel dual memory network enhances model performance and improves model usability in multi-modal learning with missing modality, providing a more informative feature distribution to complement the missing modality. Extensive experiments, including quantitative and qualitative analyses, as well as various ablation studies, demonstrate that our proposed method achieves state-of-the-art performance in the classification task on the ADNI dataset.}, archive = {J_NEUCOM}, author = {Nana Jia and Zhiao Zhang and Tong Jia}, doi = {10.1016/j.neucom.2025.131465}, journal = {Neurocomputing}, month = {12}, pages = {131465}, shortjournal = {Neurocomputing}, title = {DMNet: Incomplete multi-modal alzheimer’s disease diagnosis with dual memory network}, volume = {656}, year = {2025}, }
• (2025). WDWorm: A runtime-efficient and user-friendly GUI-based toolbox for experimenting with the nerve net of c. elegans. NEUCOM, 656, 131464. (www)

  @article{ , abstract = {Nerve net simulators of C. elegans heavily support research on its nerve net functionality by offering the possibility to conduct digital experiments instead of real ones. However, current software tools are complex and difficult to use for non-programmers. With WDWorm, we offer a user-friendly toolbox with graphical user interface for simulating and experimenting with C. elegans’ nerve net. It does not require an installation and allows for several modifications of the nerve net, including parameter changes for each neuron and connection or the deactivation of individual neurons. Furthermore, a comparison with other software tools highlights that WDWorm is currently the most runtime-efficient approach for simulating and digitally experimenting with C. elegans . To invite other developers and researchers, we provide the source code in an open-access format under a CC-BY 4.0 Creative Commons license. The code is publicly available at https://github.com/dsacri/WDWorm .}, archive = {J_NEUCOM}, author = {Sebastian Jenderny and Daniel Sacristán and Philipp Hövel and Christian Albers and Isabella Beyer and Karlheinz Ochs}, doi = {10.1016/j.neucom.2025.131464}, journal = {Neurocomputing}, month = {12}, pages = {131464}, shortjournal = {Neurocomputing}, title = {WDWorm: A runtime-efficient and user-friendly GUI-based toolbox for experimenting with the nerve net of c. elegans}, volume = {656}, year = {2025}, }
• (2025). Curriculum learning-based slimmable cross-component prediction for video coding. NEUCOM, 656, 131463. (www)

  @article{ , abstract = {Cross-component prediction plays an important role in video coding, which aims to eliminate redundancy between color components under the guidance of luma information. Recently, learning-based cross-component prediction has made significant strides in performance. However, current cross-component prediction methods typically train models directly on a dataset with different types of data, which generally results in overfitting for the flat textured data and underfitting for the complex textured data. To improve coding performance without excessively increasing the complexity, a cost-effective attention-based slimmable cross-component prediction network (SCCPN) is proposed. Although trained as a single model, SCCPN is capable of being executed at different levels of capacity, resulting in varying prediction results tailored to data with different characteristics. With the goal of further improving the generalization capability and prediction accuracy of the network, a curriculum learning strategy combined with slimmable convolutions is then designed, which employs the classification of prediction difficulty to represent whether the texture is flat or complex, and fits complex data with a small number of additional parameters. An adaptive search strategy is also introduced to speed up the selection of channels for slimmable convolutions. Experimental results demonstrate that when integrated into H.266/Versatile Video Coding (VVC), SCCPN achieves up to −0.62 %/−3.34 %/−2.68 % BD-rate reductions on Y/Cb/Cr components, respectively, over the H.266/VVC anchor. The performance gain outperforms the state-of-the-art learning-based cross-component prediction methods, while the increased complexity in both encoding and decoding is lower than the other compared cross-component prediction methods using neural networks. Moreover, performance gain can also be observed when SCCPN is integrated into the latest reference software of Beyond VVC, with BD-rate reductions of −0.17 %/−1.00 %/−1.02 % on Y/Cb/Cr components respectively.}, archive = {J_NEUCOM}, author = {Chengyi Zou and Shuai Wan and Marc Gorriz Blanch and Luka Murn and Juil Sock and Fei Yang and Luis Herranz}, doi = {10.1016/j.neucom.2025.131463}, journal = {Neurocomputing}, month = {12}, pages = {131463}, shortjournal = {Neurocomputing}, title = {Curriculum learning-based slimmable cross-component prediction for video coding}, volume = {656}, year = {2025}, }
• (2025). Representative negative sampling for graph positive-unlabeled learning. NEUCOM, 656, 131462. (www)

  @article{ , abstract = {Graph positive-unlabeled (GPU) learning is an important task that aims to develop binary classifiers using only positive and unlabeled nodes, which are commonly encountered in real-life applications. Although selecting reliable negative samples is a highly promising approach, it typically only selects high-confidence examples, which lack representativeness and fail to fully capture the diversity of the negative example space. To address this gap, our key insight, inspired by galactic dynamics, is to model the positive prototype center as a continuously evolving gravitational center maintained via a momentum moving average, just like the stars in the universe are always moving forward rather than remaining still. This dynamic anchor allows us to robustly define a reliable negative region—its “gravitational field”—for sampling representative “planets” (negative examples). We propose StarHunter-PU (SH-PU), a framework that operationalizes this insight by unifying graph representation learning with our dynamic, prototype-guided representative sampling algorithm. This ensures the sampled negatives are both diverse and informative, providing accurate information for training a robust binary classification model. Experimental results on real-world datasets show that our StarHunter-PU method significantly outperforms existing methods and even achieves competitive performance compared to fully labeled methods.}, archive = {J_NEUCOM}, author = {Luyue Wang and Xinyuan Feng and Rui Mao and Yin Li and Chunquan Liang}, doi = {10.1016/j.neucom.2025.131462}, journal = {Neurocomputing}, month = {12}, pages = {131462}, shortjournal = {Neurocomputing}, title = {Representative negative sampling for graph positive-unlabeled learning}, volume = {656}, year = {2025}, }
• (2025). Performance and interpretability analysis of code generation large language models. NEUCOM, 656, 131461. (www)

  @article{ , abstract = {As Large Language Models (LLMs) are increasingly getting integrated into software development workflows, understanding their reliability, error patterns and interpretability in real-world development scenarios is crucial for establishing their practical utility. This study evaluates and interprets the performance of 15 open-source LLM models, including Code LLaMa, Granite Code, DeepSeek-Coder-V2, and Yi-Coder on code translation and generation from requirements using the Rosetta Code dataset across diverse programming languages and tasks. Syntactic correctness and code quality are quantified using metrics such as CodeBLEU, chrF, and METEOR. Interpretability is explored through Feature Ablation and Shapley Value Sampling to elucidate prompt processing mechanisms. Results indicate high syntactic correctness and quality scores for models such as DeepSeek-Coder-V2 and Yi-Coder, alongside observed sensitivities to specific prompt components. This research provides quantitative and qualitative insights into the capabilities and limitations of open-source code-generating LLMs, informing model selection and the understanding of LLM-generated code.}, archive = {J_NEUCOM}, author = {Vishnu S. Pendyala and Neha B. Thakur}, doi = {10.1016/j.neucom.2025.131461}, journal = {Neurocomputing}, month = {12}, pages = {131461}, shortjournal = {Neurocomputing}, title = {Performance and interpretability analysis of code generation large language models}, volume = {656}, year = {2025}, }
• (2025). Enhancing long-term memory in federated class continual learning with lightweight adapters. NEUCOM, 656, 131459. (www)

  @article{ , abstract = {Federated Learning (FL) collaboratively trains a global model by aggregating local model parameters rather than raw data. Traditional FL frameworks assume that data classes are predefined and static. However, clients often encounter continuous data streams with dynamically emerging classes in practical applications, leading to a phenomenon known as catastrophic forgetting. Federated Class-Continual Learning (FCCL) has been introduced to address this challenge but still suffers from significant performance deterioration in scenarios with expanding task scales, particularly for tasks learned in the distant past. We propose a novel FCCL framework leveraging lightweight adapters to mitigate catastrophic forgetting as the number of tasks scales. To tackle the challenge of long-term memory decline, we developed task-specific adapters for clients to enhance memory retention. Additionally, we developed an image generation method tailored for lightweight adapters and trained task discriminators using the generated images. This enables the automatic loading of lightweight modules during inference, reducing human intervention. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet achieve significant performance improvements ranging from 1.73 to 4.07 times compared to baseline methods, effectively mitigating catastrophic forgetting in class-scaling scenarios. The complete implementation is available at https://github.com/notaerfa/FCLORA .}, archive = {J_NEUCOM}, author = {Pan Wang and Ji Wang and Zhengyi Zhong and Weidong Bao and Yaohong Zhang and Jianguo Chen}, doi = {10.1016/j.neucom.2025.131459}, journal = {Neurocomputing}, month = {12}, pages = {131459}, shortjournal = {Neurocomputing}, title = {Enhancing long-term memory in federated class continual learning with lightweight adapters}, volume = {656}, year = {2025}, }
• (2025). Adaptive granular-ball based density peak clustering. NEUCOM, 656, 131458. (www)

  @article{ , abstract = {In the field of data processing, the Granular-ball (GB) provides a coarse-grained data representation, offering a novel approach to improving clustering efficiency. The Fast Density Peak Clustering Algorithm based on Granular-balls (GB-DP) reduces computational granularity, which not only decreases operation time in large-scale data processing but also produces good clustering results. However, the GB-DP algorithm faces two main issues: sensitivity to the threshold parameter for generating GB and the requirement for manually selecting clustering centers, both of which affect the algorithm's efficiency and stability. To address these challenges, this paper proposes an Adaptive Granular-ball based Density Peak Clustering Algorithm (AGB-DP). First, a weighted Distribution Measure (DM) is used to dynamically generate GB. In contrast to the fixed threshold strategy used in GB-DP, this method effectively captures the data's distribution characteristics, mitigating the problem of parameter sensitivity. Second, by integrating two factors—data volume and geometric compactness—the density of GB is redefined, enhancing the accuracy of density calculations. Finally, an automatic screening strategy is employed to select GB as clustering centers, eliminating the instability introduced by manual intervention. Experimental results on both synthetic and real-world datasets demonstrate that AGB-DP, requiring only the number of clusters to be specified, achieves superior clustering results on most datasets compared to classical clustering algorithms and recent DP-based methods and shows greater robustness and stability.}, archive = {J_NEUCOM}, author = {Xingguo Zhang and Li Xu and Weikuan Jia}, doi = {10.1016/j.neucom.2025.131458}, journal = {Neurocomputing}, month = {12}, pages = {131458}, shortjournal = {Neurocomputing}, title = {Adaptive granular-ball based density peak clustering}, volume = {656}, year = {2025}, }
• (2025). Exploiting the connections between images and deep feature vectors in model inversion attacks. NEUCOM, 656, 131457. (www)

  @article{ , abstract = {Model inversion attack aims to reconstruct private samples from given deep neural networks. As the connections between the images and their corresponding deep feature vectors are unknown, it is difficult to utilize the information in the feature vectors during inversion. In this paper, connections between the images and their deep convolutional feature vectors are investigated. The directions of the vectors are used to represent the structures of both image vectors and feature vectors. Cosine similarity is further used to measure the structural similarity between different vectors. For a given target feature extractor, we find that the structures of the images and their feature vectors are highly correlated. Using this-property, Aug-MIA is proposed to perform model inversion with a few leaked feature vectors. In Aug-MIA, the feature vectors are first augmented by the proposed Structure Augmentation Algorithm. Then, a reconstruction model is trained using these augmented feature vectors to reconstruct images. Various experiments are performed on different datasets to validate our ideas. The results show that Aug-MIA performs better when fewer feature vectors are available. Specifically, when only 1 feature vector per class is leaked, it can improve the reconstruction rate by about 10.7 % on FaceScrub and about 4.2 % on MNIST, respectively.}, archive = {J_NEUCOM}, author = {Zeping Zhang and Jie Huang}, doi = {10.1016/j.neucom.2025.131457}, journal = {Neurocomputing}, month = {12}, pages = {131457}, shortjournal = {Neurocomputing}, title = {Exploiting the connections between images and deep feature vectors in model inversion attacks}, volume = {656}, year = {2025}, }
• (2025). Latent-space diffusion models for stealthy and transferable adversarial attacks on object detection. NEUCOM, 656, 131456. (www)

  @article{ , abstract = {Tons of prior works have leveraged Generative Adversarial Networks (GANs) to synthesize adversarial examples that exhibit visual fidelity. Nonetheless, the intricacy of GANs’ latent space complicates the generation of imperceptible adversarial noises, rendering the process difficult to control. The emergence of diffusion models, which iteratively refine images through a progressive denoising mechanism, offers a more tractable and interpretable solution for a controllable generation. Inspired by this, we propose a latent-space-based covert adversarial attack framework (LSDM) grounded in diffusion models to craft adversarial examples that are both visually natural and highly effective against object detection models. Central to our approach is the Latent Space Perceptual Consistency Constraint, which ensures visual-consistency by embedding perturbations into the latent space for each denoising step, while utilizing the original image as a condition guider during the de-noising pass. Moreover, to balance attack performance and the risk of overfitting, we also propose a Stepwise Prediction and Adaptive Optimization strategy, which dynamically modulates the perturbations at the current time step and determines optimal number of diffusion time steps based on the transferability of the attack against diverse black-box models. To further enhance the framework’s attacking transferability, we introduce a novel Multi-box Translation Attack strategy, which augments the spatial location diversities for each bounding box. Extensive experiments demonstrate that, compared with state-of-the-art methods, LSDM further reduces the average black-box detection mAP by 1.52 %, while improving image quality scores by 1.71 % on object detection datasets such as COCO and VOC, showcasing superior attack effectiveness and visual fidelity. The source code is publicly available at https://github.com/LSDM .}, archive = {J_NEUCOM}, author = {Wenxuan Wang and Huihui Qi and Zhixiang Huang and Bangjie Yin and Peng Wang}, doi = {10.1016/j.neucom.2025.131456}, journal = {Neurocomputing}, month = {12}, pages = {131456}, shortjournal = {Neurocomputing}, title = {Latent-space diffusion models for stealthy and transferable adversarial attacks on object detection}, volume = {656}, year = {2025}, }
• (2025). User intent disentanglement for multi-behavior recommendation via information bottleneck principle. NEUCOM, 656, 131454. (www)

  @article{ , abstract = {Multi-behavior recommendation systems have advanced rapidly by leveraging users’ diverse auxiliary behavior interactions to improve recommendations for the target behavior (a.k.a. purchase). While existing methods have made strides by integrating auxiliary behaviors with purchase histories to deliver high-quality recommendations, they often fail to identify spurious correlation intents within auxiliary behaviors that conflict with users’ target intents. Indiscriminately incorporating such correlations into the prediction of target intents may lead to performance degradation. To address this issue, we propose a Multi-Behavior Intent Disentanglement framework (MBID) for multi-behavior recommendation, which focuses on disentangling spurious correlation intents via the Information Bottleneck (IB) principle. In particular, we first design a time-sensitive spurious correlation coefficient to quantify spurious correlation intents and guide the subsequent multi-intent learning. Then, to disentangle spurious correlation intents, we propose a projection-based intent extraction method to decompose the genuine and spurious correlation intents within auxiliary behaviors. Based on this, we conceive an IB-based multi-intent learning task to disentangle the spurious correlation intents and transfer the genuine correlation intents from auxiliary behaviors into the target behavior, thereby obtaining high-quality representations of the target intent. Extensive experiments on three real-world datasets demonstrate that MBID significantly outperforms the state-of-the-art baselines by effectively disentangling the spurious correlation intents. We release our model implementation at: https://github.com/LokHsu/MBID .}, archive = {J_NEUCOM}, author = {Chenzhong Bin and Tongxin Xu and Feng Zhang}, doi = {10.1016/j.neucom.2025.131454}, journal = {Neurocomputing}, month = {12}, pages = {131454}, shortjournal = {Neurocomputing}, title = {User intent disentanglement for multi-behavior recommendation via information bottleneck principle}, volume = {656}, year = {2025}, }
• (2025). PolyModNet: Advanced positional encodings and ethical bias mitigation in adaptive multimodal fusion for multilingual language understanding. NEUCOM, 656, 131450. (www)

  @article{ , abstract = {Natural Language Understanding (NLU) plays a crucial role in Natural Language Processing (NLP), enabling machines to interpret and process human language across various applications. Despite advancements, challenges remain, including variations in data types, inconsistencies in labeling, computational demands, and biases in training datasets. These challenges emphasize the need for ethical and effective NLU solutions. To address these issues, the proposed PolyModNet combines techniques from NLP and computer vision to improve both text and image understanding. The model enhances data representation and compensates for limited training data using advanced augmentation methods such as mixup, gridmask, and positional encoding, optimized for Vision Transformer. By integrating RoBERTa-BERT and Vision Transformer, PolyModNet ensures accurate alignment of text and image features through Transformer-based encoding, specialized transformations, and structured positional encodings. Additionally, it employs a universal multilingual framework that enables language-independent retrieval and flexible task adaptation. Ethical concerns are addressed through bias detection and adversarial training, ensuring fairness in multimodal analysis. Extensive evaluations demonstrate the model’s effectiveness across multiple NLP tasks, achieving 85.71 % accuracy in sentiment analysis, strong text classification performance (CoLA: 64.1 %, SST2: 96.4 %), and high accuracy in text-image retrieval (R@1: 72.00, R@5: 89.25, R@10: 92.10). The model also delivers competitive results in multimodal translation (BLEU: 45.36, METEOR: 55.62) and cross-modal retrieval (text-to-image: R@1: 67.4, image-to-text: R@1: 82.3).}, archive = {J_NEUCOM}, author = {Shaharyar Alam Ansari and Mohd Anas Wajid and Mohd Arif and Mohammad Saif Wajid}, doi = {10.1016/j.neucom.2025.131450}, journal = {Neurocomputing}, month = {12}, pages = {131450}, shortjournal = {Neurocomputing}, title = {PolyModNet: Advanced positional encodings and ethical bias mitigation in adaptive multimodal fusion for multilingual language understanding}, volume = {656}, year = {2025}, }
• (2025). Wavelet attention is all you need in multimodal medical image fusion. NEUCOM, 656, 131448. (www)

  @article{ , abstract = {In the multimodal medical image fusion, the fusion method based on frequency domain features is a research hotspot. However, “how to effectively enhance the high frequency and low frequency information in frequency domain?”, “how to fully interact the cross-modal spatial features in feature fusion?” are the keys to improve the fusion performance. To solve this problem, this paper proposes a multimodal medical image fusion network (WTA-Net) based on Wavelet Attention. The main innovations are as follows: Firstly, the Encoder-Decoder fusion network WTA-Net with dual-encoder and single-decoder is proposed. The network effectively capture the frequency domain features in different modal images and enhance the ability of information flow between modalities. Secondly, a Wavelet Attention(WA) is designed in the encoder, which effectively enhance the high frequency and low frequency information of the lesion. Thirdly, the Cross Modal Information Fusion Module(CMIFM) is designed in the fusion stage, which fully interactive cross-modal spatial features. Finally, experiments are performed on the Whole Brain Atlas dataset and the PET-CT lung dataset. In brain MRI images and PET images comparison experiment, IE, AG and EN evaluation indexes are improved by 18.92 %, 14 % and 18.25 %, respectively. In CT mediastinal window images and PET images comparison experiment, IE and SF evaluation indexes are improved by 12.08 % and 49.4 %, respectively, WTA-Net highlight the lession information, which has positive significance for computer-aided diagnosis.}, archive = {J_NEUCOM}, author = {Tao Zhou and Mingzhe Zhang and Zhe Zhang and Jiaqi Wang and Yang Liu and Huiling Lu}, doi = {10.1016/j.neucom.2025.131448}, journal = {Neurocomputing}, month = {12}, pages = {131448}, shortjournal = {Neurocomputing}, title = {Wavelet attention is all you need in multimodal medical image fusion}, volume = {656}, year = {2025}, }
• (2025). A dual-domain mutual compensation network for multi-modality image fusion. NEUCOM, 656, 131443. (www)

  @article{ , abstract = {Existing research has demonstrated that fusing both spatial and frequency domain information from images can enhance fusion model performance, particularly in terms of saliency preservation and texture enhancement. However, designing effective fusion strategies to coordinate complementary information from both domains, while maximizing the unique characteristics and advantages of each and avoiding information conflict or redundancy, remains a challenge that requires further exploration and optimization. To address this issue, we propose a spatial–frequency Dual-Domain Mutual Compensation Network, termed D2Fusion. In our approach, the Mamba module serves as the core component of the spatial branch, capturing long-range dependencies to enhance the focus on the global spatial structure of input features. Simultaneously, the frequency branch utilizes fast Fourier Transform and convolutional neural networks to extract local texture details from the phase and magnitude components of the input features. Unlike traditional dual-branch networks, we introduce a novel phase fusion operation within the frequency branch, which combines phase information from different modalities to generate salient target features that complement and enhance the spatial features. Furthermore, to maximize the exchange of complementary characteristics between spatial, frequency, and salient target features, we design a Mutual Compensation Block (MCB) that accounts for feature differences and a decomposition loss function based on discrete cosine distance. The MCB facilitates compensatory fusion, while the decomposition loss function reduces feature similarity prior to compensation, maximizing the complementary information between domain features. Extensive experiments validate the superiority of our method, demonstrating that D2Fusion outperforms existing state-of-the-art approaches in both multi-modal image fusion and downstream task performance. The code for this framework is available at https://github.com/hz777xx/D2Fusion .}, archive = {J_NEUCOM}, author = {Jiwei Hu and Zhen Hu and Ping Lou and Kin-Man Lam and Qiwen Jin}, doi = {10.1016/j.neucom.2025.131443}, journal = {Neurocomputing}, month = {12}, pages = {131443}, shortjournal = {Neurocomputing}, title = {A dual-domain mutual compensation network for multi-modality image fusion}, volume = {656}, year = {2025}, }
• (2025). Anchor-aware representation learning for multi-view clustering. NEUCOM, 656, 131441. (www)

  @article{ , abstract = {Anchor-based multi-view clustering has garnered considerable attention in recent years owing to its ability to reduce computational overhead and enable efficient processing of large-scale datasets. However, existing anchor-based multi-view clustering models still present limitation: while orthogonality constraints are imposed on anchors to enhance their discriminative properties, the inherent relationships among anchors are neglected. To address this limitation, a novel Anchor-Aware Representation Learning for Multi-view Clustering (AARLMC) model is proposed. Specifically, anchor-wise self-representation learning is implemented, with orthogonality constraints applied to the anchor self-representation matrices to uncover intrinsic relationships among anchors. Furthermore, enhanced anchor representations are generated through this process. The anchor graphs are stacked into a third-order tensor with tensor nuclear norm constraint to explore the high-order correlations among multi-view data. Anchor-wise self-representation learning, enhanced anchor representations, and tensor representation are integrated into a unified framework. An optimization algorithm is developed to solve the proposed model. Comparative experiments on twelve benchmark datasets against thirteen state-of-the-art multi-view clustering methods demonstrate that the proposed model achieves superior performance. The source code is available on https://github.com/guowei1314/AARLMC .}, archive = {J_NEUCOM}, author = {Haotian Zhang and Wei Guo and Ruiyin Liu and Qiang Yang and Xuefei Xiao and Jilin Li and Gang Lei and Gang Chen}, doi = {10.1016/j.neucom.2025.131441}, journal = {Neurocomputing}, month = {12}, pages = {131441}, shortjournal = {Neurocomputing}, title = {Anchor-aware representation learning for multi-view clustering}, volume = {656}, year = {2025}, }
• (2025). Soft-label generator based on classifier weights. NEUCOM, 656, 131436. (www)

  @article{ , abstract = {Soft labels provide rich information between classes. Classification models obtain better generalization ability when soft labels are used as training targets in addition to hard ground-truth labels. In this paper, we propose a new approach named TarSamp to derive effective soft-targets only with the model’s classifier layer. This approach offers a simple, generic, and low-cost solution for soft label generation by fully leveraging the class-level semantics captured by the classifier layer and uncertainty injection with random sampling. We apply TarSamp to both teacher-free and teacher-available scenarios by using the classifier layer from the online model and a pre-trained teacher model, respectively. Extensive experiments on five standard image datasets are provided to evaluate the proposed approach for classifier training. TarSamp achieves more than 8 % accuracy on average for the teacher-free setting with ResNet-18, and gives on par performance by getting rid of querying to the teacher model in each forward pass during distillation for the teacher-available situation. Our results demonstrate that the proposed approach makes as a fundamental yet competitive baseline for a wide range of soft label based supervised learning.}, archive = {J_NEUCOM}, author = {Xinkai Chu and Jian-Ping Mei and Hang Zhou and Jie Chen and Rui Yan and Jing Fan}, doi = {10.1016/j.neucom.2025.131436}, journal = {Neurocomputing}, month = {12}, pages = {131436}, shortjournal = {Neurocomputing}, title = {Soft-label generator based on classifier weights}, volume = {656}, year = {2025}, }
• (2025). CESDet: Hard pedestrian object detection architecture based on the cognitive experience of structure of human body. NEUCOM, 656, 131429. (www)

  @article{ , abstract = {Conventional deep learning pedestrian detection methods usually only use the information of the current image itself. Incorrect results that go against common sense are prone to occur when dealing with hard objects with small size, unusual pose, or occlusions. Recent approaches try to enhance the hard objects by constructing and utilizing empirical information. However, due to an insufficient understanding of human body structure, the constructed experience exhibits poor generalization ability to diverse poses. Furthermore, when leveraging experiential features to enhance the features of hard objects, the process is susceptible to interference from occlusions and background. Inspired by human vision, we propose CESDet, a novel pedestrian detection network that constructs and utilizes Cognition Experience of Structure of human body in an unsupervised manner. The key technical innovations are three folds: (1) an unsupervised Cognition Experience of Structure construction module that addresses pose generalization by automatically forming decoupled body parts and pose semantics, (2) a part-level fine-grained verification and feature enhancement module that addresses the interference of occlusions and background with the guidance of Cognition Experience of Structure, and (3) an end-to-end pedestrian detection network for hard objects based on the two proposed modules. Experiments comparing with seven methods on three datasets demonstrate that CESDet achieves state-of-the-art performance, with highest AP on the training dataset, and lowest degradation of AP on a novel unseen dataset. The proposed framework advances the detection of hard objects by exploiting the automatically constructed Cognition Experience of Structure with decoupled part-level appearance and pose.}, archive = {J_NEUCOM}, author = {Yanglin Pu and Xiaohui Hao and Shan Yang and Hangyuan Yang and Shengxin Dai}, doi = {10.1016/j.neucom.2025.131429}, journal = {Neurocomputing}, month = {12}, pages = {131429}, shortjournal = {Neurocomputing}, title = {CESDet: Hard pedestrian object detection architecture based on the cognitive experience of structure of human body}, volume = {656}, year = {2025}, }
• (2025). Event-related potential extraction workflow based on kernel density estimation. NEUCOM, 656, 131425. (www)

  @article{ , abstract = {Event-related potentials (ERPs) are a critical neuroscientific tool for investigating brain responses to external stimuli and serve as a key linking mechanism in brain–computer interface systems. Traditional ERP extraction methods rely on threshold-based trial rejection and time-locked averaging techniques, which often have limited ability to handle outlier data and are susceptible to random artifacts. To address this, we propose a novel ERP extraction workflow based on kernel density estimation. We construct trial-wise datasets at the sampling-point granularity and model the probability distribution of each trial using Gaussian kernel density estimation, effectively reducing outlier influence while preserving all trial data. The fitted probability density function serves as the objective function for ERP extraction, enabling active reconstruction of optimal ERP waveforms by incorporating inherent EEG temporal dependencies. Specifically targeting uneven noise distribution across multiple channels, we introduce an adaptively steering kernel dynamically generated from electrode covariance matrices, which optimizes the adaptive matching of inter-channel noise structures to ensure more precise density function fitting. Using two real datasets and simulated datasets, our comparative analyses of baseline root mean square error, component-level statistical metrics, and residual correlations demonstrate that, compared with the traditional trial rejection and time-locked averaging methods, our approach exhibits outstanding effectiveness in isolating ERP components from raw signals and significantly reduces the impact of outlier contamination.}, archive = {J_NEUCOM}, author = {Weizhuang Kong and Zihao Zhang and Jing Zhu and Yizhou Li and Xiaowei Li and Bin Hu}, doi = {10.1016/j.neucom.2025.131425}, journal = {Neurocomputing}, month = {12}, pages = {131425}, shortjournal = {Neurocomputing}, title = {Event-related potential extraction workflow based on kernel density estimation}, volume = {656}, year = {2025}, }
• (2025). Multi-view optimization and refinement for high-fidelity 4D gaussian splatting. NEUCOM, 656, 131424. (www)

  @article{ , abstract = {Reconstructing dynamic 3D scenes from 2D images and synthesizing temporally diverse views is challenging due to the interplay between scene complexity and temporal dynamics. While 3D Gaussian Splatting offers an efficient solution for static scene modeling, extending it to dynamic scenes faces significant challenges in motion representation and texture fidelity. We propose a novel framework based on multi-view interpolation and joint optimization to address these challenges in sparse-view dynamic scenes. This framework combines linear and spherical interpolation strategies to generate novel views, producing high-quality interpolated images from multiple fitted viewpoints. Additionally, it incorporates consistency constraints to optimize texture representation, enhancing the reconstruction performance for dynamic scenes. Experimental results demonstrate that the proposed framework significantly improves detail fidelity and motion representation in dynamic scene reconstruction.}, archive = {J_NEUCOM}, author = {Jinhui Lin and Zhenyang Wei and Silei Shen and Fang Zhou and Xiaobin Zhu and Xu-Cheng Yin}, doi = {10.1016/j.neucom.2025.131424}, journal = {Neurocomputing}, month = {12}, pages = {131424}, shortjournal = {Neurocomputing}, title = {Multi-view optimization and refinement for high-fidelity 4D gaussian splatting}, volume = {656}, year = {2025}, }
• (2025). KD-KI: Knowledge distillation with knowledge infusion for anomaly detection and localization. NEUCOM, 656, 131423. (www)

  @article{ , abstract = {Anomaly detection and localization are crucial for improving product reliability in industrial quality inspection. Existing knowledge distillation methods often cause student networks to merely mimic features of teacher networks, which makes it difficult to achieve stable and generalized detection performance. This paper introduces the KD-KI framework, which uses a knowledge infusion mechanism to transfer structured hierarchical knowledge from the teacher network to the student network. This guides the student to learn more robust representations of normal samples. Additionally, a feature bias loss is used to optimize the similarity of shallow-layer features, improving detection accuracy and localization precision. KD-KI can be deployed with standard convolutional networks and is suitable for real-time industrial inspection systems. Experimental results demonstrate that the proposed KD-KI model can yield improved performance in anomaly detection and localization compared to other competing models.}, archive = {J_NEUCOM}, author = {Wei Huang and Zhaonan Xu and Rongchun Wan and Xuhua Yang and Bingyang Zhang}, doi = {10.1016/j.neucom.2025.131423}, journal = {Neurocomputing}, month = {12}, pages = {131423}, shortjournal = {Neurocomputing}, title = {KD-KI: Knowledge distillation with knowledge infusion for anomaly detection and localization}, volume = {656}, year = {2025}, }
• (2025). DFCon: Dominant frequency enhanced ultra-long time series contrastive forecasting. NEUCOM, 656, 131418. (www)

  @article{ , abstract = {Ultra-long time series forecasting (ULTSF) is crucial for fields like energy management, traffic planning, and climate prediction. However, as the forecast horizon increases, concept drift becomes a major challenge, as a fixed-length historical window struggles to generalize ultra-long temporal patterns. Extending the input series length increases computational costs and demands a higher model capacity for capturing longer temporal dependencies. To address these issues, we propose DFCon, a dominant frequency enhanced contrastive learning framework for ULTSF. DFCon combines dilated convolutions for feature extraction and multi-layer perceptrons for forecasting, with a dual contrastive loss based on dominant frequency enhancement. We introduce Temporal DFCon, which enhances the model’s sensitivity to these frequency-domain features during training, thereby improving its ability to model global temporal dependencies in the input series. Furthermore, cross-window Autocorrelated DFCon is proposed, which mitigates concept drift by constructing autocorrelated relative positive and negative samples without introducing noisy data. Experiments on five benchmark datasets show that DFCon outperforms existing methods, demonstrating its effectiveness in ULTSF. The code for this work is publicly available at: https://github.com/coding4qq/DFCon .}, archive = {J_NEUCOM}, author = {Qiaoqiao Liu and Hui Liu and MingJie Yang and Yuheng Wei and Junzhao Du}, doi = {10.1016/j.neucom.2025.131418}, journal = {Neurocomputing}, month = {12}, pages = {131418}, shortjournal = {Neurocomputing}, title = {DFCon: Dominant frequency enhanced ultra-long time series contrastive forecasting}, volume = {656}, year = {2025}, }
• (2025). Breaking the alignment barrier: A spatiotemporal alignment-free RGBT tracking approach. NEUCOM, 656, 131382. (www)

  @article{ , abstract = {RGB-Thermal (RGBT) tracking leverages complementary information from visible and infrared modalities to improve tracking robustness in complex environments. However, its practical deployment remains constrained by the stringent requirement for precise spatiotemporal alignment between heterogeneous modalities—a condition rarely satisfied in real-world applications. To overcome this limitation, we present SAFNet, a novel Spatiotemporal Alignment-Free Network that eliminates the need for precise cross-modal alignment through innovative architectural designs. Our framework develops a spatiotemporal interaction query module incorporating cross-modal temporal attention, which re-establishes inter-modal temporal correlations for unregistered inputs by leveraging similarity learning across asynchronous data streams. For spatial discrepancy mitigation, we propose a dual-branch pre-tracking network employing deep cross-correlation analysis, combined with an adaptive feature fusion strategy under the guidance of joint response distribution. Furthermore, we devise an innovative dynamic template update mechanism that adaptively adjusts modal update rates to maintain temporal consistency across heterogeneous data streams. Comprehensive evaluations validate SAFNet’s state-of-the-art performance across four benchmark datasets (GTOT, RGBT210, RGBT234, LasHeR), demonstrating significant improvements in tracking accuracy. The proposed architecture represents a significant advancement toward practical deployment of robust RGBT tracking systems in real-world environments with asynchronous multimodal inputs.}, archive = {J_NEUCOM}, author = {Xiaodong Liu and Meibo Lv and Daming Zhou and Lingyu Si and Ruiheng Zhang and Hui Xu}, doi = {10.1016/j.neucom.2025.131382}, journal = {Neurocomputing}, month = {12}, pages = {131382}, shortjournal = {Neurocomputing}, title = {Breaking the alignment barrier: A spatiotemporal alignment-free RGBT tracking approach}, volume = {656}, year = {2025}, }
• (2025). The implicit regularization of gradient flow on separable datasets in ReLU networks. NEUCOM, 656, 131367. (www)

  @article{ , abstract = {This work aims to investigate the implicit regularization of gradient flow when training non-linearly separable multi-class datasets in ReLU networks. We prove that the gradient flow leads the directions of each layer to converge. Besides, we show that the convergent directions of different layers collectively form a Karush-Kuhn-Tucker (KKT) point of the max normalized margin problem in parameter space. In particular, for any l ∈ [ L ] , we prove that the convergent direction of the l -th layer is the local maximum of the max normalized margin problem when other directions are fixed. This indicates that the convergent directions maximize the margin for each layer separately. Moreover, we show the growth rate of weights across different layers and demonstrate their layer-balanced properties. Furthermore, we prove that the cross-entropy loss converges and we give the tight bound of the convergence rate. In addition, we present a generalization bound for ReLU networks in multi-class tasks based on Rademacher complexity. Our results demonstrate that the generalization bound is primarily influenced by the normalized margin and the number of training samples. Finally, we present a series of numerical experiments to verify our theory.}, archive = {J_NEUCOM}, author = {Xiangyun Hui and Xiaoxuan Ma and Yixuan Yang and Song Li}, doi = {10.1016/j.neucom.2025.131367}, journal = {Neurocomputing}, month = {12}, pages = {131367}, shortjournal = {Neurocomputing}, title = {The implicit regularization of gradient flow on separable datasets in ReLU networks}, volume = {656}, year = {2025}, }
• (2025). Exploring cycle cover variants: A dataless neural networks approach. NEUCOM, 656, 131361. (www)

  @article{ , abstract = {Cycle Cover, a fundamental concept in graph theory, plays a critical role in various applications, including network design, transportation optimization, and bioinformatics. A Cycle Cover is a collection of cycles covering all the vertices of a given graph, ensuring that each vertex belongs to exactly one cycle. In this paper, we explore various aspects of Cycle Cover variants. We employ the dataless neural networks framework to establish single differentiable functions for each of these variants. Recent research has demonstrated the capability of the dataless neural networks framework in representing a host of combinatorial optimization problems. Motivated by these findings, we propose dataless neural networks tailored for the Cycle Cover variants. Additionally, we present a rigorous proof of the correctness of our approach.}, archive = {J_NEUCOM}, author = {Sangram K. Jena and K. Subramani and Alvaro Velasquez}, doi = {10.1016/j.neucom.2025.131361}, journal = {Neurocomputing}, month = {12}, pages = {131361}, shortjournal = {Neurocomputing}, title = {Exploring cycle cover variants: A dataless neural networks approach}, volume = {656}, year = {2025}, }
• (2025). Lightweight deep learning for visual perception: A survey of models, compression strategies, and edge deployment challenges. NEUCOM, 656, 131357. (www)

  @article{ , abstract = {The increasing demand for the deployment of deep neural networks (DNNs) in edge devices has led to the development of lightweight deep learning (LDL) models designed to operate efficiently under resource constraints. Although DNNs have achieved remarkable success in various applications, their high computational requirements often limit their deployment on devices with restricted memory and processing power. This challenge has motivated researchers to develop optimized LDL models that balance accuracy, speed, and efficiency while maintaining competitive performance. Despite existing surveys covering specific aspects of LDL models, a comprehensive review encompassing image classification, object detection, and segmentation remains limited. This proposed survey systematically explores recent advancements in LDL models, highlighting their architectures, optimization techniques, and real-world applications. This survey conducts an empirical evaluation by testing latest state-of-the-art LDL models on the Jetson Orin edge device using benchmark datasets: ImageNet for classification, VisDrone for object detection, and COCO for segmentation. The experimental analysis focuses on key performance metrics, including inference speed, model accuracy, and computational efficiency, while comparing LDL models with their conventional counterparts. This study provides a holistic understanding of the role of LDL models in edge computing, providing insight into emerging trends, challenges, and future research directions in the field.}, archive = {J_NEUCOM}, author = {Syed Muhammad Raza and Syed Murtaza Hussain Abidi and Md Masuduzzaman and Soo Young Shin}, doi = {10.1016/j.neucom.2025.131357}, journal = {Neurocomputing}, month = {12}, pages = {131357}, shortjournal = {Neurocomputing}, title = {Lightweight deep learning for visual perception: A survey of models, compression strategies, and edge deployment challenges}, volume = {656}, year = {2025}, }
• (2025). LaSNN: Layer-wise ANN-to-SNN distillation for effective and efficient training in deep spiking neural networks. NEUCOM, 656, 131351. (www)

  @article{ , abstract = {Spiking Neural Networks (SNNs) are biologically realistic and practically promising in low-power computation because of their event-driven mechanism. Usually, the training of SNNs suffers from accuracy loss on various tasks, yielding an inferior performance compared with ANNs. A conversion scheme is proposed to obtain competitive accuracy by mapping trained ANNs’ parameters to SNNs with the same structures. However, an enormous number of time steps are required for these converted SNNs, thus losing the energy-efficient benefit. Utilizing both the accuracy advantages of ANNs and the computing efficiency of SNNs, a novel SNN training framework is proposed, namely layer-wise ANN-to-SNN knowledge distillation (LaSNN). In order to achieve competitive accuracy and reduced inference latency, LaSNN transfers the learning from a well-trained ANN to a small SNN by distilling the knowledge rather than converting the parameters of ANN. The information gap between heterogeneous ANN and SNN is bridged by introducing the attention scheme. The knowledge in an ANN is effectively compressed and then efficiently transferred by utilizing our layer-wise distillation paradigm. We conduct detailed experiments to demonstrate the effectiveness, efficacy, and scalability of LaSNN on three benchmark data sets (CIFAR-10, CIFAR-100, and Tiny ImageNet). We achieve competitive top-1 accuracy compared to ANNs and faster inference than converted SNNs with similar performance. More importantly, LaSNN is dexterous and extensible that can be effortlessly developed for SNNs with different architectures/depths and input encoding methods, contributing to their potential development.}, archive = {J_NEUCOM}, author = {Di Hong and Yu Qi and Yueming Wang}, doi = {10.1016/j.neucom.2025.131351}, journal = {Neurocomputing}, month = {12}, pages = {131351}, shortjournal = {Neurocomputing}, title = {LaSNN: Layer-wise ANN-to-SNN distillation for effective and efficient training in deep spiking neural networks}, volume = {656}, year = {2025}, }
• (2025). Fusing spatio-temporal information using supervised local low-rank correlation embedding for depression recognition. NEUCOM, 656, 131350. (www)

  @article{ , abstract = {Electroencephalography (EEG) signals contain rich spatio-temporal information that reflects the brain’s dynamic activity, making it widely used in depression recognition. However, effectively integrating this information to capture discriminative and complementary features remains a key challenge. To address this issue, we propose a novel Discriminative Local Low-Rank Correlation Embedding (DLLCE) to fuse spatio-temporal information of EEG. DLLCE integrates shared low-rank representation, local invariance, discriminative constraints, and enhanced correlation analysis into a unified framework. Specifically, the shared low-rank representation is used to capture the common structural patterns, while the correlation analysis aims to reduce redundancy among feature sets. In addition, the Laplacian regularization is applied to the shared representation to preserve the local geometric structure of the original data. To further enhance discriminative capability, a discriminant graph embedding term is incorporated to exploit label information. Experimental results on EEG datasets demonstrate that DLLCE achieves superior performance compared to existing methods. This work provides new insights into EEG-based mental health assessment and holds promise for early depression diagnosis and clinical decision support.}, archive = {J_NEUCOM}, author = {Lu Zhang and Peng Xu and Zhijun Yao and Xinyan Zhang and Juan Wang and Bin Hu and Gang Feng and Hong Peng}, doi = {10.1016/j.neucom.2025.131350}, journal = {Neurocomputing}, month = {12}, pages = {131350}, shortjournal = {Neurocomputing}, title = {Fusing spatio-temporal information using supervised local low-rank correlation embedding for depression recognition}, volume = {656}, year = {2025}, }
• (2025). Knowledge-driven baseline for few-shot fine-grained visual recognition. NEUCOM, 656, 131302. (www)

  @article{ , abstract = {Few-shot fine-grained visual recognition (FS-FGVR), a practical yet challenging task, aims to break the dilemma of having scarce training examples for new fine-grained tasks. Meta-learning-based methods target this issue by employing the learning-to-learn strategy to train a well-generalized meta-learner from seen fine-grained tasks for unseen fine-grained tasks. However, most existing works rely too much on small-scale fine-grained training tasks. Specifically, these works demand large amounts of fine-grained data to sample these training tasks, and they are unable to generalize well to new tasks. Consequently, model capacity can be highly restricted when limited training references are available. This paper presents a novel coarse-to-fine framework named Knowledge-Driven baseline for FS-FGVR by transferring knowledge from large-scale and coarse-grained datasets. This framework departs the meta-training phase into the coarse-grained meta-pretraining and fine-grained meta-finetuning phases. First, off-the-shelf coarse-grained data is introduced to build the initialization correlations as prior knowledge. Then, we use prior knowledge to infer the representational interactions and correlations of the fine-grained representations. Extensive experiments show that our method outperforms the current methods on the public few-shot fine-grained benchmarks. We also develop extensive studies to extend our method to few-shot texture visual recognition scenarios.}, archive = {J_NEUCOM}, author = {Jieqi Sun and Jian Li and Yafeng Li}, doi = {10.1016/j.neucom.2025.131302}, journal = {Neurocomputing}, month = {12}, pages = {131302}, shortjournal = {Neurocomputing}, title = {Knowledge-driven baseline for few-shot fine-grained visual recognition}, volume = {656}, year = {2025}, }
• (2025). BMTM-net: A rotating machinery fault diagnosis network based on 2D-1D fusion with bidirectional multi-granularity transformer-mamba. NEUCOM, 656, 131293. (www)

  @article{ , abstract = {Rotating machinery plays a critical role in industrial operations, yet existing diagnostic methods often struggle with missing correlations between sensor data, weak noise immunity, and insufficient long-range feature extraction. To address these challenges, this paper proposes BMTM-Net, a fault diagnosis network based on 2D-1D fusion with Bidirectional Multi-granularity Transformer-Mamba (BMTM). The network consists of two main components: a 2D sequential information interaction network and a 1D temporal information extraction network. The 2D network captures inter-sensor sequence relationships and temporal dependencies using a Bidirectional Multi-granularity Transformer (BM-Transformer) and an embedded Sequential-Temporal Attention Module (ST-Attention), while the 1D network enhances feature completeness and extracts temporal information through a Bidirectional Multi-granularity Mamba (BM-Mamba) network, integrated with a Channel Attention-based Fusion Module (CAFM) for adaptive feature fusion. To evaluate BMTM-Net’s effectiveness and stability, experiments were conducted on datasets from Southeast University and a Self-Built bogie integrated test stand, with various levels of noise introduced to assess noise immunity. The results demonstrate that BMTM-Net achieves over 99 % accuracy across all four datasets and maintains high accuracy even under severe noise interference (SNR = −10 dB), outperforming other state-of-the-art methods with accuracy rates of 99.60 %, 99.40 %, 98.54 %, and 94.38 %, respectively. Additionally, the model exhibits low complexity, further confirming its robustness and effectiveness in noisy environments.}, archive = {J_NEUCOM}, author = {E. Xia and Yirong Liu and Jinyang Gong and Xunhua Dai and Tongyang Pan and Shiyi Wang}, doi = {10.1016/j.neucom.2025.131293}, journal = {Neurocomputing}, month = {12}, pages = {131293}, shortjournal = {Neurocomputing}, title = {BMTM-net: A rotating machinery fault diagnosis network based on 2D-1D fusion with bidirectional multi-granularity transformer-mamba}, volume = {656}, year = {2025}, }
• (2025). Advancements in radiomics: A comprehensive survey of feature types and their correlation on modalities and regions. NEUCOM, 656, 131192. (www)

  @article{ , abstract = {This survey paper provides an overview of different feature types used in radiomics research and their applications across various medical imaging modalities and disease domains. The paper delves into the key aspects of the radiomics workflow, including data engineering techniques for image acquisition, preprocessing, fusion, and segmentation. It then presents a comprehensive review of the most commonly employed feature categories in radiomics, such as shape-based, first-order statistical, second-order texture, and transform-based features. The paper also discusses the emerging role of deep learning features extracted using convolutional neural networks, recurrent neural networks, and transformers. The analysis of feature usage trends across different anatomical regions and imaging modalities offers valuable insights that can guide the optimization of feature engineering strategies in future radiomics research. The survey concludes by highlighting several opportunities for further advancement in the field, including the need for larger multi-center datasets, multi-modal data fusion, self-supervised learning, and the development of efficient embedded models for on-device deployment.}, archive = {J_NEUCOM}, author = {Luca Zedda and Andrea Loddo and Cecilia Di Ruberto}, doi = {10.1016/j.neucom.2025.131192}, journal = {Neurocomputing}, month = {12}, pages = {131192}, shortjournal = {Neurocomputing}, title = {Advancements in radiomics: A comprehensive survey of feature types and their correlation on modalities and regions}, volume = {656}, year = {2025}, }
• (2025). Progressive low-confidence pseudolabeling for semisupervised node classification. NEUCOM, 656, 131166. (www)

  @article{ , abstract = {Graph neural networks (GNNs) have demonstrated remarkable achievements in handling graph-structured data. However, the performance of GNNs is typically limited by the lack of sufficient labeled data, which are time-consuming to obtain in real-world scenarios. Pseudolabeling has been applied to GNNs by augmenting the training set data with unlabeled data. Most pseudolabeling methods on graphs assign pseudolabels to nodes based on high-confidence thresholds. However, nodes near labeled ones generally obtain high confidence scores during training. This results in an increasing number of similar nodes being assigned pseudolabels during training, which potentially leads to a distribution shift between the labeled dataset and the augmented dataset. The distribution of the augmented dataset diverges significantly from that of the entire graph data, causing the GNNs to perform poorly on test data. In this paper, we propose a progressive low-confidence pseudolabeling (PLCP) method to progressively leverage the low-confidence data. Specifically, pseudolabels are assigned to nodes within a predefined confidence-based ranking range. To alleviate distribution shift, we keep this range constant throughout the training process to prevent excessive nodes from being assigned pseudolabels. The range is designed to be sufficiently wide to leverage low-confidence nodes. Low-confidence nodes from the range propagate information to their neighbors, which helps the model capture patterns in uncertain regions. To alleviate the impact of noisy pseudolabels, a validation-based reassignment scheme is proposed to utilize validation metrics to assign more reliable pseudolabels. Numerous experiments are conducted to demonstrate that our proposed PLCP improves the performance of state-of-the-art GNNs on graph datasets in comparison with several established methods.}, archive = {J_NEUCOM}, author = {Tao Zhu and Hua Mao and Hui Liu and Jie Chen}, doi = {10.1016/j.neucom.2025.131166}, journal = {Neurocomputing}, month = {12}, pages = {131166}, shortjournal = {Neurocomputing}, title = {Progressive low-confidence pseudolabeling for semisupervised node classification}, volume = {656}, year = {2025}, }
• (2025). Deep-DFVAR: Dynamic factor vector autoregression with deep learning for regional house price index forecasting. NEUCOM, 656, 131103. (www)

  @article{ , abstract = {House Price Index (HPI) is an indicator that reflects changes in residential house prices over time. Predicting HPI is crucial for homebuyers to determine the right time to purchase and for policymakers to formulate housing policies. Recent studies have reported that neural network approaches outperform classical methods in HPI forecasting. However, challenges remain due to limited monthly HPI data and its time-varying statistical properties. As a result, state-of-the-art time series forecasting models often respond slowly to abrupt changes and lack economic interpretability. To address these issues, we propose Deep-DFVAR, a hybrid framework that decomposes regional HPI into shared (common trends) and idiosyncratic (regional variations) components. The shared component is predicted with Vector Autoregression (VAR) based on Granger causality, which improves interpretability and responds faster to changes. The idiosyncratic component is modeled with our deep learning model, which benefits from reduced distribution shift (train–test gap). We evaluate Deep-DFVAR on South Korean and United States datasets, empirically demonstrating that our framework outperforms traditional and recent time series forecasting models. All data and code are publicly available at: https://github.com/YeoJiSu/House-Price-Index-Prediction .}, archive = {J_NEUCOM}, author = {Jisu Yeo and Artyom Stitsyuk and Jaesik Choi}, doi = {10.1016/j.neucom.2025.131103}, journal = {Neurocomputing}, month = {12}, pages = {131103}, shortjournal = {Neurocomputing}, title = {Deep-DFVAR: Dynamic factor vector autoregression with deep learning for regional house price index forecasting}, volume = {656}, year = {2025}, }
• (2025). Event-based triple asynchronous switching control for fuzzy neural networks: A full-information dependent lyapunov approach. NEUCOM, 656, 131097. (www)

  @article{ , abstract = {This paper employs the full-information dependent Lyapunov-Krasovskii functional (LKF) analysis approach to investigate the multiple weighting adaptive event-triggered triple asynchronous switching control problem for Takagi-Sugeno fuzzy neural networks with semi-Markov jump parameters. Considering the influence of factors such as network delays and disturbances, there may be asynchronous premise variables and modes among the system, event generator and controller. Therefore, a triple asynchronous switching control framework under the multiple weighting adaptive event-triggered scheme is developed. Under this control framework, a novel full-information dependent LKF analysis approach is proposed to analyze the stability of neural networks, which fully considers the system information, such as the membership functions (MFs) information, modes information and state information. Meanwhile, a new MFs dependent optimal H ∞ performance index is introduced to achieve better disturbance attenuation ability. The proposed analysis approach is helpful in determining the controller gains and reducing the conservatism. Ultimately, four simulation examples are provided to show the effectiveness and superiority of proposed approach.}, archive = {J_NEUCOM}, author = {Yiteng Zhang and Linchuang Zhang and Yonghui Sun and Wen Yang}, doi = {10.1016/j.neucom.2025.131097}, journal = {Neurocomputing}, month = {12}, pages = {131097}, shortjournal = {Neurocomputing}, title = {Event-based triple asynchronous switching control for fuzzy neural networks: A full-information dependent lyapunov approach}, volume = {656}, year = {2025}, }
• (2025). DCHF_T: A multi-dimensional adaptive compression approach for transformer-based models. NEUCOM, 656, 131071. (www)

  @article{ , abstract = {In recent years, pre-trained language models based on the Transformer architecture have achieved significant results in many natural language processing tasks. However, the high computational cost limits their application in real-world scenarios. Previous Transformer compression methods typically focus on single-dimensional compression, which may cause over-compression and consequently degrade model performance. Additionally, these methods lack targeted optimization for specific downstream tasks. In this paper, we propose DCHF_T, a multi-dimensional adaptive compression approach that compresses Transformer models through token compression, attention head pruning, and a lightweight FFN. This approach selects the most informative tokens during training, prunes unimportant tokens, and retains their information in a compressed form, allowing the model to focus more on task-relevant inputs. Furthermore, DCHF_T combines attention head pruning and a lightweight FFN to reduce computation and parameter size across multiple dimensions. We employ multi-objective evolutionary search to optimize the trade-off between accuracy and efficiency under various computational budgets. Experimental results on the GLUE benchmark demonstrate that DCHF_T achieves the best compression–performance trade-off. While maintaining the highest accuracy, DCHF_T achieves a reduction of 3.7 × and 3.6 × in FLOPs on BERT-base and RoBERTa-base, respectively. By implementing adaptive multi-dimensional compression, DCHF_T provides a systematic solution for deploying Transformer models in resource-constrained scenarios.}, archive = {J_NEUCOM}, author = {Yaoyao Yan and Da Wang and Jing Ye and Hui Yu and Dianjie Lu and Yuang Zhang and Weizhi Xu and Fang’ai Liu}, doi = {10.1016/j.neucom.2025.131071}, journal = {Neurocomputing}, month = {12}, pages = {131071}, shortjournal = {Neurocomputing}, title = {DCHF_T: A multi-dimensional adaptive compression approach for transformer-based models}, volume = {656}, year = {2025}, }
• (2025). Key-concept thinking prompting for improved reasoning in large language models. NEUCOM, 656, 130986. (www)

  @article{ , abstract = {Recent advancements in large language models (LLMs) have significantly accelerated the development of natural language processing (NLP) research, demonstrating remarkable capabilities in understanding and generating human-like text. However, these models face limitations when it comes to system 2 tasks, which require slow, multi-step, and conscious reasoning. To address these limitations, we introduce a method called Key-Concept Thinking (KCT), which enhances the model’s reasoning ability by directing it to identify and prioritize key concepts within a problem. Building on the Chain-of-Thought (CoT) prompting method, KCT anchors its approach in core ideas, allowing the model to form a deeper understanding of the problem’s structure and purpose. This targeted approach aims to improve both the accuracy and efficiency of the model’s reasoning, making it better equipped to handle tasks that require precision and deep understanding. We evaluate our proposed prompting strategies using 24 reasoning tasks across four categories: arithmetic, commonsense, symbolic, and other logical reasoning tasks, with three prominent large models: ChatGLM4, Baichuan2, and DeepSeek, respectively. The results show that the Zero-shot-KCT and Zero-shot-CoT-KCT strategies outperform traditional zero-shot and few-shot prompting strategies, highlighting the effectiveness of incorporating key concept thinking into the reasoning processes of LLMs. Our findings have implications for the development of more effective prompting strategies for LLMs that can handle complex reasoning tasks with higher accuracy and coherence.}, archive = {J_NEUCOM}, author = {Minghua Tang and Chen Bian and Liming Yang and Xueling Zhong}, doi = {10.1016/j.neucom.2025.130986}, journal = {Neurocomputing}, month = {12}, pages = {130986}, shortjournal = {Neurocomputing}, title = {Key-concept thinking prompting for improved reasoning in large language models}, volume = {656}, year = {2025}, }
• (2025). A first look at temporal bias in industrial image datasets. NEUCOM, 655, 131488. (www)

  @article{ , abstract = {Deep learning-based machine vision methods have emerged as a crucial means to enhance industrial production efficiency and quality. Vision models exhibit superior performance upon initial deployment, while their performance deteriorates significantly as the input images bias over time. This bias is termed temporal bias, which is often overlooked in vision industry applications. Therefore, this paper investigates temporal bias in industrial image datasets, using tires and high-reflectivity parts in car interiors as representative examples of industrial products. Firstly, Name that Dataset experiments based on Convolutional Neural Networks (CNN) are designed to demonstrate the existence of temporal bias for two industrial images respectively. This method successfully mines the image data distribution bias in datasets that are difficult to observe. Additionally, the designed cross-dataset generalization experiments exhibit the hazards of temporal bias on model inference by simulating real-world industrial situations. While the proposed image feature dimensionality reduction method enables the visualization of this temporal bias, with the results demonstrating that the temporal bias in defect images is more pronounced than in normal images. Finally, this paper proposes three possible solutions on how to minimize the model performance loss due to temporal bias. The comparison reveals that the effectiveness of different solutions for mitigating temporal bias is related to conditions such as the size of the dataset.We hope that these studies and observations will provide valuable insights to guide future research on de-bias methods for industrial datasets.}, archive = {J_NEUCOM}, author = {You Zhou and Jinhua Zhang and Qianqian Cao and Jun Hu}, doi = {10.1016/j.neucom.2025.131488}, journal = {Neurocomputing}, month = {11}, pages = {131488}, shortjournal = {Neurocomputing}, title = {A first look at temporal bias in industrial image datasets}, volume = {655}, year = {2025}, }
• (2025). XMIX: Graph-based temporal credit assignment and attention-augmented value decomposition for multi-agent cooperative reinforcement learning. NEUCOM, 655, 131471. (www)

  @article{ , abstract = {Multi-agent reinforcement Learning (MARL) needs to balance individual flexibility and global goal consistency in dynamic collaboration. However, the performance of traditional Value Decomposition (VD) methods remains constrained by ambiguous credit assignment, which manifests in three aspects: the difficulty of explicitly modeling complex inter-agent dependencies, the bias introduced by partial observability, and the inability to distinguish heterogeneous individual contributions under a shared global reward. To address these limitations, we propose a comprehensive framework based on the Graph Isomorphism Network (GIN), termed XMIX. XMIX leverages GIN to explicitly model the collaborative topology among agents and dynamically capture dependencies between neighboring agents, thereby improving the precision of credit assignment. Furthermore, it incorporates a multi-head attention mechanism that enables agents to learn latent behavioral patterns from the graph structure, thus enhancing credit assignment in complex graph spaces. Finally, XMIX introduces a gradient-based contribution quantification strategy to resolve credit ambiguity at the temporal level and mitigate the identity-related bias caused by centralized training. Experiments show that our algorithm outperforms existing Graph Neural Network (GNN)-based MARL algorithms as well as classical MARL algorithms on StarCraft SMAC and SMACv2. Moreover, it achieves a significantly higher win rate compared to baseline algorithms in heterogeneous scenarios of super-hard difficulty (e.g., 3 s _ v s _ 5 z and 3 s 5 z _ v s _ 3 s 6 z ), which demonstrates the excellent performance of XMIX in complex tasks. Our code is available at: https://github.com/jiwenhao66/XMIX .}, archive = {J_NEUCOM}, author = {Shudong Wang and Wenhao Ji and Haiyuan Gui and Kuijie Zhang and Luqi Wang and Shanchen Pang}, doi = {10.1016/j.neucom.2025.131471}, journal = {Neurocomputing}, month = {11}, pages = {131471}, shortjournal = {Neurocomputing}, title = {XMIX: Graph-based temporal credit assignment and attention-augmented value decomposition for multi-agent cooperative reinforcement learning}, volume = {655}, year = {2025}, }
• (2025). Batch-agnostic dynamic GNN for mitigating temporal discontinuity. NEUCOM, 655, 131447. (www)

  @article{ , abstract = {In Memory-based Dynamic Graph Neural Networks (MDGNNs), temporal discontinuity caused by large-batch training disrupts the sequential order of events, leading to the loss of temporal information. Our analysis shows that this information loss affects the model’s ability to effectively optimize its parameters, making convergence more difficult. Furthermore, we derive an upper Lipschitz bound revealing that larger batch sizes expand the model’s weight search space, which increases the difficulty of finding the optimal parameters. Guided by this insight, we propose Batch-Agnostic Dynamic GNN (BADGNN), which involves two methods: (1) Temporal Lipschitz Regularization (TLR), which restricts the expansion of the weight search space and stabilizes optimization under large batch sizes; and (2) Adaptive Attention Adjustment (A3), which complements TLR by refining attention distributions to mitigate information loss caused by both regularization and large batches. Together, these methods enable BADGNN to achieve robust performance during large-scale training. Extensive experiments on three real-world datasets show that BADGNN supports up to 11× larger batch sizes and achieves 2 × training speedup over TGN, without sacrificing performance. Our code is available at Code: https://github.com/YangZhou1999999999/TGN_Lip}, archive = {J_NEUCOM}, author = {Yang Zhou and Xiaoning Ren and Jing He}, doi = {10.1016/j.neucom.2025.131447}, journal = {Neurocomputing}, month = {11}, pages = {131447}, shortjournal = {Neurocomputing}, title = {Batch-agnostic dynamic GNN for mitigating temporal discontinuity}, volume = {655}, year = {2025}, }
• (2025). ML-PINN: A memory-efficient physics-informed mamba-LSTM network for fast and accurate PDE solving. NEUCOM, 655, 131446. (www)

  @article{ , abstract = {Physics-informed neural networks (PINNs) have emerged as a novel PDE solver and demonstrated significant potential. However, most existing models struggle to achieve satisfactory performance in both solution accuracy and computational efficiency, particularly in time-dependent modeling scenarios where extended input sequences are typically required to maintain precision. This inevitably leads to excessive GPU memory allocation and prolonged training durations. In this work, we develop a novel architecture integrating Mamba with Long Short-Term Memory networks (LSTM), hereby referred to as the Physics-informed Mamba-LSTM Neural Network (ML-PINN). By using shorter input sequences, ML-PINN is able to maintain high accuracy while achieving up to a 36 % reduction in GPU memory consumption and a 48 % decrease in training time compared to methods such as PINNMamba which employs state-space models and PINNsFormer (a Transformer-based PINN framework).}, archive = {J_NEUCOM}, author = {YiMing Gao and Bing Wang and Jingyi Lu and Zhou Tian}, doi = {10.1016/j.neucom.2025.131446}, journal = {Neurocomputing}, month = {11}, pages = {131446}, shortjournal = {Neurocomputing}, title = {ML-PINN: A memory-efficient physics-informed mamba-LSTM network for fast and accurate PDE solving}, volume = {655}, year = {2025}, }
• (2025). Comprehensive image restoration for robot-assisted PC-side UI automated testing using neural network. NEUCOM, 655, 131444. (www)

  @article{ , abstract = {In this paper, we propose a novel deep learning-based image restoration framework, FESDNet-SwinT, tailored for robot-assisted automated testing of PC-side user interfaces (UIs). The method addresses the compounded challenges of screen image degradation, including moiré patterns, low resolution and blurring. By integrating high-frequency enhancement modules with Swin Transformer architectures, FESDNet-SwinT effectively restores fine details and structural fidelity. We introduce PC-Restore, a large-scale dataset comprising 94,675 degraded/clean image pairs, specifically designed to empower models with robust generalization capabilities for real-world UI scenarios. Furthermore, we establish an Optical Character Recognition (OCR)-based evaluation method to objectively assess restoration performance, particularly in text-rich UI images. Experimental results demonstrate that FESDNet-SwinT achieves state-of-the-art performance with PSNR of 28.34, SSIM of 0.933, LPIPS of 0.044 and OCR accuracy of 90.4 %. This comprehensive solution ensures high-quality image restoration while maintaining real-time performance, making it highly suitable for diverse UI automated testing applications. The code is available at https://github.com/zyx-study/FESDNet-SwinT .}, archive = {J_NEUCOM}, author = {Yunxiang Zhu and Hailei Ding and Yangkun Zhu and Jingtao Hu and Ming-Che Lee and Chaklam Silpasuwanchai and Min Huang and Yibo Zou}, doi = {10.1016/j.neucom.2025.131444}, journal = {Neurocomputing}, month = {11}, pages = {131444}, shortjournal = {Neurocomputing}, title = {Comprehensive image restoration for robot-assisted PC-side UI automated testing using neural network}, volume = {655}, year = {2025}, }
• (2025). LPUDC: A laplacian pyramid neural network for restoring images from under-display cameras. NEUCOM, 655, 131442. (www)

  @article{ , abstract = {Images captured by under-display cameras (UDCs) suffer from severe degradation due to the complex optical path passing through multiple material layers, leading to light attenuation, color distortion, and diffraction artifacts. The display stack typically includes polarizing films and OLED pixel layers, which introduce nonlinear and spatially non-uniform effects on light propagation. To model this degradation, we consider two major components. First, the transparent layers mainly cause color shifts and overall intensity loss. Second, the OLED layer, containing light-emitting units and opaque wiring, introduces complex diffraction effects that result in irregular high-frequency artifacts. To address these challenges, we propose a Laplacian pyramid neural network for restoring images from UDCs (LPUDC). This method decomposes the image into high- and low-frequency components, enabling targeted correction of different types of degradation. In addition, we use a deep guided filtering (DGF) module to transfer low-resolution features for high-resolution reconstruction, improving efficiency and reducing cost. Extensive quantitative and qualitative evaluations on two widely used UDC benchmarks demonstrate that LPUDC outperforms state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Zhenyan Ding and Zhixiang Fang and Zhengning Wang and Lehan Ding and Zhenni Zeng and Xinglong Luo and Shaoqin Yuan and Haoran Ma and Binquan Leng}, doi = {10.1016/j.neucom.2025.131442}, journal = {Neurocomputing}, month = {11}, pages = {131442}, shortjournal = {Neurocomputing}, title = {LPUDC: A laplacian pyramid neural network for restoring images from under-display cameras}, volume = {655}, year = {2025}, }
• (2025). A memory-based conditional neural process for video instance segmentation. NEUCOM, 655, 131439. (www)

  @article{ , abstract = {Video instance segmentation (VIS) is an evolving research topic in computer vision that aims to simultaneously detect, segment, and track semantic objects across multiple video frames. However, existing VIS methods are typically unaware of the reliability of the training samples from insufficient and imbalanced datasets, leading to suboptimal performance. To address this challenge, we propose a memory-based conditional neural process (MemCNP) module to exploit the strengths of both memory networks and the CNP model which handles heterogeneous latent space distributions for reliable modelling with insufficient data. Our MemCNP utilises predicted uncertainty to regularise VIS predictions as well as to identify reliable samples for effective training. Notably, our MemCNP is model-agnostic and can thus be seamlessly integrated into various VIS models to improve their performance. Extensive experiments on the YouTube-VIS and OVIS datasets demonstrate the effectiveness of MemCNP regardless of the underlying model architecture.}, archive = {J_NEUCOM}, author = {Kunhao Yuan and Gerald Schaefer and Yu-Kun Lai and Xiyao Liu and Lin Guan and Hui Fang}, doi = {10.1016/j.neucom.2025.131439}, journal = {Neurocomputing}, month = {11}, pages = {131439}, shortjournal = {Neurocomputing}, title = {A memory-based conditional neural process for video instance segmentation}, volume = {655}, year = {2025}, }
• (2025). DentSeg-EDA: A 3D tooth segmentation framework for CBCT images using enhanced dual attention. NEUCOM, 655, 131437. (www)

  @article{ , abstract = {The accurate three-dimensional (3D) segmentation of individual teeth from cone-beam computed tomography (CBCT) images is essential for enhancing computer-assisted planning in various dental procedures. Despite recent advances, existing methods continue to face difficulties when dealing with complex dental anatomies, particularly in scenarios with limited annotated data. In this study, we propose a novel enhanced dual-attention mechanism integrated within the skip connections of a two-stage V-Net framework to overcome these challenges and address the clinical demand for more robust and efficient solutions. In the initial stage of our pipeline, the network employs 3D convolutions to extract regions of interest, tooth skeletons, and centroids simultaneously, thereby effectively capturing spatial relationships and preserving essential anatomical details. These features are subsequently fed into the encoder of the second-stage V-Net, which performs the final segmentation of individual teeth. The incorporation of the enhanced dual attention block within the skip connections enables the model to focus on both global contextual cues and local anatomical features, leading to superior feature representations and improved boundary delineation accuracy. The proposed method was evaluated on a dataset of CBCT images. Comprehensive experiments demonstrated that our model achieved a Dice score of 92.01 % and an intersection over union of 85.35 %, outperforming the best previous method by 2.00 % and 3.13 % points, respectively, while reducing the Hausdorff distance and average surface distance, which are critical improvements for accurate tooth root and boundary segmentation in clinical practice, by 0.67 mm and 0.11 mm, respectively. This work contributes to the advancement of digital dentistry by offering a more reliable and efficient solution for 3D tooth segmentation, with practical applications in clinical settings.}, archive = {J_NEUCOM}, author = {Muhammad Asif Jamal and Guoqing Chao and Bumshik Lee}, doi = {10.1016/j.neucom.2025.131437}, journal = {Neurocomputing}, month = {11}, pages = {131437}, shortjournal = {Neurocomputing}, title = {DentSeg-EDA: A 3D tooth segmentation framework for CBCT images using enhanced dual attention}, volume = {655}, year = {2025}, }
• (2025). Single-image signal-dependent noise parameter estimation method based on weak texture image block selection with ARDPC. NEUCOM, 655, 131435. (www)

  @article{ , abstract = {Existing relative density peak clustering algorithms require an empirical definition of the relevant parameters, which leads to unstable clustering results. To alleviate this drawback, we propose adaptive relative density peak clustering (ARDPC), which removes the effect of difference in density distribution and selects the clustering centers from datasets with lower density distribution. Specifically, ARDPC designs a method for the adaptive selection of the cutoff distance. It first uses information entropy to determine an adaptive cutoff distance for dataset, thereby avoiding the influence of empirical judgment on the clustering results. Next, an adaptive relative density is defined to replace the original local density. This adaptive relative density is obtained by comparing the density of each point with that of its surrounding points, ensuring that clustering centers with sparse densities attain higher density values. Finally, we compare ARDPC with classical clustering algorithms and other RDPC methods on synthetic and real datasets. The experimental results show that the mean values of the F-score, adjusted rand index, and normalized mutual information for ARDPC are 14 %, 30 %, and 28 % higher than those of the other algorithms, respectively. In addition, we use ARDPC to select weakly textured image blocks for the estimation of signal-dependent noise parameters, thereby overcoming the problems of incomplete image feature characterization and block threshold selection influenced by empirical judgment. Experimental results demonstrate that the proposed method achieves a smaller mean squared error (MSE) in noise parameter estimation across different noise models, including Poisson-Gaussian and Gaussian-Gaussian models.}, archive = {J_NEUCOM}, author = {Yu-qian Wu and Jing-long Wang and Jin-tao Tao and Yu Zhang}, doi = {10.1016/j.neucom.2025.131435}, journal = {Neurocomputing}, month = {11}, pages = {131435}, shortjournal = {Neurocomputing}, title = {Single-image signal-dependent noise parameter estimation method based on weak texture image block selection with ARDPC}, volume = {655}, year = {2025}, }
• (2025). Label-guided multi-beam forward-looking sonar image enhancement. NEUCOM, 655, 131434. (www)

  @article{ , abstract = {Object detection in multibeam forward-looking sonar (MFLS) images faces significant challenges due to noise and poor target features. This paper proposes LBD-Enhance, a label-guided brightness differentiation algorithm that improves target discriminability by adaptively highlighting target regions while suppressing background noise. Our solution comprises a label-guided differentiation curve and a lightweight network (HCVNet) for real-time inference. Experiments on EOD-Test dataset demonstrate significant improvements of 16.69 % in AP50 and 14.10 % in mAP on challenging scenes, while maintaining real-time performance (25.322 FPS on NVIDIA 3060Ti, 9.625 FPS on NVIDIA Orin Nano). LBD-Enhance is particularly suitable for real-time underwater detection applications. The code is publicly available at: https://github.com/PiratePai/LBD_Enhance .}, archive = {J_NEUCOM}, author = {Pai Peng and Zhongliang Jing and Han Pan and Xuyang Zhang and Kaiyao Ling}, doi = {10.1016/j.neucom.2025.131434}, journal = {Neurocomputing}, month = {11}, pages = {131434}, shortjournal = {Neurocomputing}, title = {Label-guided multi-beam forward-looking sonar image enhancement}, volume = {655}, year = {2025}, }
• (2025). Semi-SNN: Biological-inspired semi-supervised image classification with spiking neural networks. NEUCOM, 655, 131433. (www)

  @article{ , abstract = {Information is transmitted in the form of spikes between neurons in the human brain. Spiking neural networks (SNNs) emulate the brain’s information processing using spikes, demonstrating exceptional efficiency in power consumption. The human brain processes information without repeatedly training a large amount of labeled data like artificial neural networks (ANNs), and only uses a small amount of labeled data to achieve few-shot learning. Therefore, the combination of SNNs and semi-supervised learning (SSL) using only a small amount of labeled data is closer to the process of human brain processing and cognitive information. Nevertheless, effectively utilizing SNNs for processing extensive data with minimal annotations remains a formidable challenge. In this paper, we introduce a novel approach to combine SSL with SNNs for image classification tasks by leveraging surrogate gradient descent. Our SSL paradigm integrates teacher–student distillation to leverage brain-like SNNs to extract consistent information from labeled and unlabeled data through robust augmentation. To address the low learning efficiency of SNNs, we adopt a joint training scheme with a burn-in stage to warm up the model. This approach significantly enhances the training speed, mitigating the high training cost associated with network inefficiency. Experimental validation on CIFAR10, CIFAR100 and SVHN demonstrates the superiority of our method in terms of accuracy and training cost compared to other methods. Our method achieves outstanding performance on the CIFAR10 sub-dataset and competitive performance on other datasets, approaching the optimal algorithmic level. Additionally, our method exhibits shorter training times, with a single iteration consuming only 37.5 % of the Flex Match time, and reduced memory usage to 25 % of Flex Match and 71.41 % of the fastest training Pseudo Labeling.}, archive = {J_NEUCOM}, author = {Zhen Cao and Xiaoyu Chen and Biao Hou and Chuanfeng Ma and Leida Li and Hao Zhu and Licheng Jiao}, doi = {10.1016/j.neucom.2025.131433}, journal = {Neurocomputing}, month = {11}, pages = {131433}, shortjournal = {Neurocomputing}, title = {Semi-SNN: Biological-inspired semi-supervised image classification with spiking neural networks}, volume = {655}, year = {2025}, }
• (2025). Fastener anomaly detection for industrial assembly scenarios based on few-shot fine-grained component segmentation. NEUCOM, 655, 131432. (www)

  @article{ , abstract = {Accurate fastener anomaly detection is critical in complex industrial assembly (e.g., helicopter main blade hubs), yet conventional methods fail to address logical anomalies arising from assembly relations and irrelevant component interference. To overcome these limitations, we propose a novel two-stage anomaly detection framework leveraging unlabeled data with minimal segmentation annotations. First, a few-shot fine-grained segmentation model—integrating a few-shot component segmentation prediction branch (FCPB) and global semantic foreground refinement branch (GFRB)—achieves 81.78 % mIoU under 5-shot learning. Crucially, FCPB encodes assembly-related priors via assembly relation entropy and component assembly ratio losses. Second, a feature-assembly relation fusion memory bank jointly analyzes the component features and logical constraints of the assembly relations through assembly relation histogram (ARH) and component feature embedding (CFE). Extensive evaluation on our newly established fastener assembly anomaly detection (FA-AD) dataset demonstrates 99.31 % AUROC, surpassing SOTA by 13.44 % and validates efficacy in industrial assembly scenarios.}, archive = {J_NEUCOM}, author = {Chen Liu and Meng Chi and Yanding Wei and Qiang Fang and Mingjun Tang and Libin Wang}, doi = {10.1016/j.neucom.2025.131432}, journal = {Neurocomputing}, month = {11}, pages = {131432}, shortjournal = {Neurocomputing}, title = {Fastener anomaly detection for industrial assembly scenarios based on few-shot fine-grained component segmentation}, volume = {655}, year = {2025}, }
• (2025). All-in-one weather image restoration: Multi-modal attention and bi-directional mamba fusion. NEUCOM, 655, 131431. (www)

  @article{ , abstract = {To address the coexistence of scattering noise and occlusion-induced structural degradation in images captured under adverse weather conditions such as rain and snow, we propose a novel Mamba-Attention Fusion Block (MAF-Block). This module is the first to integrate frequency-domain and edge-guided attention, driven by multi-modal priors, with bidirectional Mamba-based efficient long-range dependency modeling within a single Transformer block. This design enables a unified framework for dynamic denoising and structural reconstruction. In practice, the MAF-Block is incorporated into a multi-scale encoder-decoder architecture to enhance the model’s feature representation and restoration capabilities. Experimental results demonstrate that our method achieves notable performance improvements on the Outdoor Rain and Snow100K-S (L) datasets, with PSNR gains of 0.84 dB and 0.66 dB (0.14 dB) respectively, surpassing the current state-of-the-art methods for rain and snow image restoration.}, archive = {J_NEUCOM}, author = {Mingzhang Guo and Yongping Xie}, doi = {10.1016/j.neucom.2025.131431}, journal = {Neurocomputing}, month = {11}, pages = {131431}, shortjournal = {Neurocomputing}, title = {All-in-one weather image restoration: Multi-modal attention and bi-directional mamba fusion}, volume = {655}, year = {2025}, }
• (2025). YOLO-AEF: Traffic sign detection on challenging traffic scenes via adaptive enhancement and fusion. NEUCOM, 655, 131430. (www)

  @article{ , abstract = {With the rapid development of intelligent driving technologies such as autonomous driving and driver assistance systems, traffic sign detection has become increasingly vital for road and transportation safety. To address performance degradation under complex illumination conditions, this paper proposes YOLO with Adaptive Enhancement Fusion (YOLO_AEF), a novel framework that integrates adaptive enhancement and dual-branch feature fusion. The Multiple Exposure Enhancement Module (MEEM) adaptively enhances low-quality images under challenging illumination conditions, while the Adaptive Feature Fusion Module (AFFM) fuses semantic features from both the original and enhanced images, improving robustness and contextual representation. The Fusion Detection Module (FDM) realizes robust detection of traffic signs following the framework of YOLO. Experimental results show that YOLO_AEF outperforms DEIM on the TT100K dataset by 4.8 percentage points and surpasses YOLOv12 on the CCTSDB dataset by 7.0 percentage points demonstrating strong generalization capability. Compared with various state-of-the-art methods, YOLO_AEF delivers superior accuracy and speed under diverse illumination challenges, making it highly suitable for intelligent traffic perception systems.}, archive = {J_NEUCOM}, author = {Peizhi Wang and Raihani Mohamed and Norwati Mustapha and Noridayu Manshor}, doi = {10.1016/j.neucom.2025.131430}, journal = {Neurocomputing}, month = {11}, pages = {131430}, shortjournal = {Neurocomputing}, title = {YOLO-AEF: Traffic sign detection on challenging traffic scenes via adaptive enhancement and fusion}, volume = {655}, year = {2025}, }
• (2025). Multi-dimensional opinion dynamics for social networks with asynchronous updates. NEUCOM, 655, 131428. (www)

  @article{ , abstract = {Multi-dimensional opinion dynamic models can often be used to describe the mutual influence of agents on different topics expressed in social networks. This paper mainly focuses on the multi-dimensional opinion dynamics on social networks with asynchronous updates, where networks include two types of agents: stubborn agents and non-stubborn agents. In the asynchronous update mechanism, the update time instants of each agent are independent and different from those of other agents, which leads to the complexity and time variance of opinion interaction among agents. This paper comprehensively analyzes the convergence of multi-dimensional opinion dynamic models under asynchronous updates, relying on the techniques of binary relation composition and infinite products of sub-stochastic matrices. The results show that the opinions of stubborn agents gradually influence the opinions of non-stubborn agents, and ultimately completely determine the stable opinions of non-stubborn agents. The theoretical results are validated through numerical simulations.}, archive = {J_NEUCOM}, author = {Heng Wang and Panpan Zhu and Lei Shi}, doi = {10.1016/j.neucom.2025.131428}, journal = {Neurocomputing}, month = {11}, pages = {131428}, shortjournal = {Neurocomputing}, title = {Multi-dimensional opinion dynamics for social networks with asynchronous updates}, volume = {655}, year = {2025}, }
• (2025). MP-DRA: Multi-scale memory and adaptive pseudo-anomaly enhanced open-set anomaly detection. NEUCOM, 655, 131427. (www)

  @article{ , abstract = {The efficacy of open-set anomaly detection is severely constrained by the scarcity of real anomalies. To tackle this, we argue for the critical importance of effectively learning from available anomalies and generating valuable pseudo-anomalies. We introduce MP-DRA, a novel framework featuring a multi-scale feature memory bank that is dynamically updated with representative samples during training. For anomaly scoring, our model adaptively queries the memory bank to retrieve robust and discriminative features. Critically, MP-DRA dynamically calibrates the difficulty and strategy of pseudo-anomaly synthesis after each epoch based on performance metrics, leading to a more efficient training paradigm. Extensive experiments on several real-world datasets validate that MP-DRA delivers state-of-the-art and stable performance on diverse anomaly detection tasks. Our code can be found at https://github.com/Goolubo/MLR/tree/v4.0 .}, archive = {J_NEUCOM}, author = {Yunxue Shao and Fangdi Xu and Lingfeng Wang}, doi = {10.1016/j.neucom.2025.131427}, journal = {Neurocomputing}, month = {11}, pages = {131427}, shortjournal = {Neurocomputing}, title = {MP-DRA: Multi-scale memory and adaptive pseudo-anomaly enhanced open-set anomaly detection}, volume = {655}, year = {2025}, }
• (2025). LLM-CG: Large language model-enhanced constraint graph for distantly supervised relation extraction. NEUCOM, 655, 131426. (www)

  @article{ , abstract = {Distant supervision has effectively addressed the issue of limited training data in relation extraction. However, significant challenges such as noisy labels and long-tailed distributions persist. Most existing methods depend on noisy instances for model training, leading to error propagation and accumulation. Inspired by the robust generative and reasoning capabilities of large language models (LLMs), we leverage LLMs as an auxiliary knowledge source to enhance model performance. Building on this concept, we introduce a novel distantly supervised relation extraction framework that incorporates an LLM-enhanced constraint graph. Specifically, entity types act as intermediaries to construct a constraint graph connecting various relations, facilitating the transfer of valuable information from data-rich to data-scarce relations. Furthermore, we utilize LLMs to obtain initial node representations within the constraint graph and subsequently employ a graph neural network encoder to capture both semantic and structural information. Additionally, the encoded entity types are concatenated with the corresponding sentences and relations to provide an auxiliary constraint that mitigates the impact of noisy instances. Extensive experiments demonstrate that the proposed model significantly outperforms existing competitive baselines in both denoising and long-tailed relation extraction.}, archive = {J_NEUCOM}, author = {Bao Liu and Guilin Qi}, doi = {10.1016/j.neucom.2025.131426}, journal = {Neurocomputing}, month = {11}, pages = {131426}, shortjournal = {Neurocomputing}, title = {LLM-CG: Large language model-enhanced constraint graph for distantly supervised relation extraction}, volume = {655}, year = {2025}, }
• (2025). Radial search-based graph clustering method. NEUCOM, 655, 131421. (www)

  @article{ , abstract = {Graph-based clustering methods represent data samples as nodes and their relationships as edges, which effectively capture the complex structures within similarity graphs. However, many graph-based clustering methods rely on traditional spectral clustering to partition the graph, which may overlook crucial local structural information and affect clustering accuracy. To address this, we propose a Radial Search-Based Graph Clustering (RSGC) method, which can detect clusters with arbitrary shapes and densities, even in complex scenarios such as high-dimensional or multi-peak distributions. We propose a Radial Search Allocation (RSA) method for initial partitioning of the similarity graph, which constructs well-structured single-peak sub-graphs by fully considering the local structure. Additionally, we propose a method for calculating sub-graph similarity based on the importance of cross-cluster edges in the similarity graph, and obtain the final clustering result by merging highly similar sub-graphs. Experimental validations on synthetic and real datasets demonstrate the effectiveness of the RSGC method.}, archive = {J_NEUCOM}, author = {Yangyang Zhao and Feng Zhu and Junyi Guan and Xiongxiong He and Sheng Li}, doi = {10.1016/j.neucom.2025.131421}, journal = {Neurocomputing}, month = {11}, pages = {131421}, shortjournal = {Neurocomputing}, title = {Radial search-based graph clustering method}, volume = {655}, year = {2025}, }
• (2025). A novel fault diagnosis method for electric submersible progressing cavity pumps: Addressing imbalanced mixed-domain datasets with complex subdomains. NEUCOM, 655, 131420. (www)

  @article{ , abstract = {The Electric Submersible Progressing Cavity Pump (ESPCP) is a critical device for oil and gas extraction, operating under harsh subsurface conditions. Due to its complexity and limited observability, fault diagnosis of ESPCPs remains challenging. Recent advancements in deep learning (DL) have shown promise in fault diagnosis, particularly in cross-domain learning. However, most existing methods focus on improving domain adaptability of pre-trained models, with little attention to constructing initial models capable of handling heterogeneous industrial datasets from multiple sources. In this study, we collected 27 ESPCP fault cases from SCADA systems, which present challenges such as dynamic operating conditions, unidentified discrepancies, and incomplete labels, forming a mixed domain of multiple unknown subdomains. To address this, we propose the Mixed-Domain Intra-Distribution Alignment Module (MD-IDA), which implicitly aligns domains by minimizing intra-class feature similarity and maximizing inter-class feature dissimilarity, enabling effective training on complex datasets. Additionally, we apply sample generation techniques in the aligned feature space to optimize sample distributions and mitigate inter-domain disparities. Experimental results demonstrate that our method achieves superior fault diagnosis accuracy on multi-source industrial datasets.}, archive = {J_NEUCOM}, author = {Yajie Ou and Wenzheng Chen and Xiaoyong Gao and Bin Liu and Xin Lei and Chaodong Tan}, doi = {10.1016/j.neucom.2025.131420}, journal = {Neurocomputing}, month = {11}, pages = {131420}, shortjournal = {Neurocomputing}, title = {A novel fault diagnosis method for electric submersible progressing cavity pumps: Addressing imbalanced mixed-domain datasets with complex subdomains}, volume = {655}, year = {2025}, }
• (2025). USTNet: Ultrafast style transfer between infrared and visible images. NEUCOM, 655, 131419. (www)

  @article{ , abstract = {Visible light and infrared images belong to different, modalities, which is not conductive to the image registration. To achieve better image registration and fusion, it is necessary to convert the modality of visible light images into infrared modality before image registration. The paper proposes a lightweight real-time framework named USTNet for generating infrared-like visible-light images, which can bridge the modality gap in visible and infrared domains. To accelerate the feature representation at early stage, a Channel Acceleration block is adopted; Besides, the Faster Transfer Module (FTM) and the Detail Connection Module (DCM) ensure the speed of semantic embedding and the high-quality reconstructed image, respectively. Moreover, by introducing Matting Laplacian constraint and dWCT transform, the regionally semantically consistent features as well as coherent regional styles across source and target domain are enforced in this work. The proposed USTNet is evaluated on both near-infrared and far-infrared datasets and compared against six state-of-the-art style transfer methods. Extensive experiments show that the proposed USTNet achieves superior qualitative and quantitative results with a 2x speedup over the state-of-the-art methods, making it more compatible for real-time cross-modal vision applications.}, archive = {J_NEUCOM}, author = {Haoxiang Shi and Puchun Wang and Yanqi Wu and Bo Yang and Zhaozi Zu and Zhongjun Qu}, doi = {10.1016/j.neucom.2025.131419}, journal = {Neurocomputing}, month = {11}, pages = {131419}, shortjournal = {Neurocomputing}, title = {USTNet: Ultrafast style transfer between infrared and visible images}, volume = {655}, year = {2025}, }
• (2025). A neumann-neumann acceleration with coarse space for domain decomposition of extreme learning machines. NEUCOM, 655, 131417. (www)

  @article{ , abstract = {Extreme learning machines (ELMs), which preset hidden layer parameters and solve for last layer coefficients via a least squares method, can typically solve partial differential equations faster and more accurately than Physics Informed Neural Networks. However, they remain computationally expensive when high accuracy requires large least squares problems to be solved. Domain decomposition methods (DDMs) for ELMs have allowed parallel computation to reduce training times of large systems. This paper constructs a coarse space for ELMs, which enables further acceleration of their training. By partitioning interface variables into coarse and non-coarse variables, selective elimination introduces a Schur complement system on the non-coarse variables with the coarse problem embedded. Key to the performance of the proposed method is a Neumann-Neumann acceleration that utilizes the coarse space. Numerical experiments demonstrate better condition numbers leading to a ten times reduction in iterations and four times faster training time compared to a previous DDM method for ELMs.}, archive = {J_NEUCOM}, author = {Byungeun Ryoo and Chang-Ock Lee}, doi = {10.1016/j.neucom.2025.131417}, journal = {Neurocomputing}, month = {11}, pages = {131417}, shortjournal = {Neurocomputing}, title = {A neumann-neumann acceleration with coarse space for domain decomposition of extreme learning machines}, volume = {655}, year = {2025}, }
• (2025). Multi-behavior multivariate contrastive learning for MOOC video recommendation. NEUCOM, 655, 131416. (www)

  @article{ , abstract = {In Massive Open Online Courses (MOOCs), the interaction between learners and videos is complex and diverse, but video completion rates are generally low. Since video completion rates usually depend on whether the content aligns with learners’ preferences, different completion rates can reflect various learning behavior patterns. This paper aims to explore the correlation between multiple behavior patterns and video modality information to recommend videos and improve completion rates. This brings two key challenges: (i) It is difficult to extract learning preferences from video modalities in a personalized manner and link them with behavior patterns; (ii) When modeling multiple behaviors, it is challenging to eliminate noise from behaviors with low completion rates. To address these challenges, this paper proposes a Multi-behavior Multi-behavior Multivariate Contrastive Learning framework (MMCL). First, we design a high-order search aggregation network (HSAN) in the behavior-aware embedding generator, which captures the diversity and dependencies of behaviors by identifying effective high-order node information under different behaviors. Then, to capture learning preferences from modalities, we design a modality-aware embedding generator, incorporating semantic information from various video modalities in behaviors with high completion rates. Finally, we use an adaptive fusion mechanism to represent the impact of personalized learning preferences on multiple behavior patterns. Additionally, we design a multivariate contrastive learning component to filter out behavioral noise by extracting self-supervised signals of modality-aware embeddings for learners through constructing contrastive losses. In two real-world educational datasets, our method improves the HR@5 by 19.84 % and 7.23 %, respectively, compared to the state-of-the-art baseline.}, archive = {J_NEUCOM}, author = {Qing Yang and Zhiying Li and Zezheng Wu and Yuzhao Huang and Jingwei Zhang}, doi = {10.1016/j.neucom.2025.131416}, journal = {Neurocomputing}, month = {11}, pages = {131416}, shortjournal = {Neurocomputing}, title = {Multi-behavior multivariate contrastive learning for MOOC video recommendation}, volume = {655}, year = {2025}, }
• (2025). Vertical partitioned high-dimensional data publishing with differential privacy principal component analysis. NEUCOM, 655, 131415. (www)

  @article{ , abstract = {In the era of information explosion, the massive and high-dimensional nature of data pose severe challenges to data analysis and privacy protection, particularly when multiple institutions hold different attributes of the same object. To address this challenge, this study proposes the integration of principal component analysis (PCA) and differential privacy (DP) to achieve data dimensionality reduction and enhanced privacy protection. Building on this, we have designed two differential privacy-based multi-party vertically partitioned data release methods: principal component analysis differential privacy algorithm (PCA-DP) and covariance matrix principal component analysis differential privacy algorithm (CMPCA-DP). To address the characteristics of high-dimensional data, the PCA-DP algorithm ensures privacy by introducing noise into the low-dimensional approximation of the dataset. In contrast, the CMPCA-DP algorithm employs a novel approach that involves the addition of noise to the diagonal entries of the covariance matrix. This strategy significantly enhances the efficiency of data processing and the utility of the dataset, while concurrently maintaining data privacy. Through theoretical and experimental analyses, we find that the CMPCA-DP algorithm ensures data privacy protection while also offering higher efficiency and greater data utility.}, archive = {J_NEUCOM}, author = {Junming Zhang and Jingru Wang and Shigong Long and Guangyuan Liu and Lun Wang and Di Yang}, doi = {10.1016/j.neucom.2025.131415}, journal = {Neurocomputing}, month = {11}, pages = {131415}, shortjournal = {Neurocomputing}, title = {Vertical partitioned high-dimensional data publishing with differential privacy principal component analysis}, volume = {655}, year = {2025}, }
• (2025). GCB-diff: Global context-aware and boundary-aware feature fusion based on diffusion model for medical image segmentation. NEUCOM, 655, 131414. (www)

  @article{ , abstract = {Medical image segmentation identifies anatomical structures or lesions in scans, which is vital for diagnosis, surgical planning, and treatment monitoring. Although denoising diffusion probabilistic models (DDPMs) have recently succeeded in image generation and segmentation, directly applying them to medical images remains challenging due to unclear boundaries, irregular shapes, and heterogeneous appearances. To address these challenges, we propose GCB-Diff, a novel diffusion-based framework leveraging global context modeling and boundary-aware feature extraction for more accurate and reliable medical image segmentation. The Global Context-Aware Diffusion Module (GCDM) captures long-range anatomical dependencies to establish understanding of the global context. Building on these features, the Boundary-Aware Feature Extraction Module (BFEM) refines spatial details by delineating precise object edges. The Feature Fusion Module (FFM) then integrates the global and boundary features to generate high-fidelity segmentation maps. We evaluate our model on five medical image segmentation datasets. ISLES’15, ISLES’17, ISLES’22, BraTS2020, and BTCV. On the ISLES datasets, our method achieved DSC/HD95 scores of 61.85 % / 43.23 mm, 82.74 % / 4.11 mm, 58.71 % / 22.87 mm, and 81.66 % / 3.33 mm. On BraTS2020 and BTCV, the scores were 84.31 % / 6.23 mm and 78.44 % / 9.48 mm, respectively. Comparative and ablation experiments show that GCB-Diff achieves excellent segmentation performance across various medical imaging tasks by combining global context and boundary-aware features within a diffusion framework. Its improved accuracy and boundary precision are expected to enhance radiologists’ diagnostic and treatment planning capabilities. GCB-Diff codes are available at https://github.com/ShannanChen/GCB-Diff .}, archive = {J_NEUCOM}, author = {Shannan Chen and Qingda Chen and Tongkang Zhang and Kaizhou Gao and Ronghui Ju and Peizhuo Zang and Shouliang Qi}, doi = {10.1016/j.neucom.2025.131414}, journal = {Neurocomputing}, month = {11}, pages = {131414}, shortjournal = {Neurocomputing}, title = {GCB-diff: Global context-aware and boundary-aware feature fusion based on diffusion model for medical image segmentation}, volume = {655}, year = {2025}, }
• (2025). MDSTransUNet: Multi-scale deep supervised transformer U-net for COVID-19 lung tissue and infection segmentation. NEUCOM, 655, 131413. (www)

  @article{ , abstract = {The global COVID-19 pandemic outbreak in early 2020 precipitated an unprecedented public health crisis, highlighting the urgent demand for advanced diagnostic solutions. Hybrid architectures that combine U-Net and Transformer models have attracted widespread attention in medical image segmentation. However, global attention mechanisms often compromise local edge information and tend to overlook subtle regions of lung infection. In this paper, we propose a novel multi-scale deep supervised segmentation network, named MDSTransUNet, designed for segmenting lung and infection regions from chest CT and X-ray images. Following the conventional encoder–decoder paradigm, MDSTransUNet integrates ShuffleNet and Transformer modules in the encoder, while leveraging multi-resolution inputs to enhance deep feature supervision. Two dedicated decoder branches are employed to segment lung structures and infection regions simultaneously. The interaction between the two branches ensures that the segmentation of infection regions is constrained within the lung structures. To further boost segmentation performance, we introduce three key components: the multi-scale parallel pooling (MSPP) module, the enhanced atrous spatial pyramid pooling (EASPP) module, and the global context channel attention (GCCA) module. These modules serve to reinforce feature learning, extract multi-scale features, and acquire global attention, thereby enhancing the segmentation accuracy. Extensive experiments on multi-modal datasets demonstrate that our MDSTransUNet achieves high segmentation accuracy for lung structures of varying sizes and diverse infection patterns, consistently outperforming existing state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Yidan Yan and Beibei Hou and Junding Sun}, doi = {10.1016/j.neucom.2025.131413}, journal = {Neurocomputing}, month = {11}, pages = {131413}, shortjournal = {Neurocomputing}, title = {MDSTransUNet: Multi-scale deep supervised transformer U-net for COVID-19 lung tissue and infection segmentation}, volume = {655}, year = {2025}, }
• (2025). IATA: Instance-driven advancing targeted attacks with transferable pattern embedding. NEUCOM, 655, 131412. (www)

  @article{ , abstract = {Generative attack methods have attracted increasing attention in recent years due to their potential to efficiently deceive black-box models. Among the various attacks, the targeted attacks focus on misleading victim models to produce adversary-desired predictions, which are more challenging and threatening than the untargeted attacks. However, the current methods still exhibit compromised generalization in low-data regimes due to insufficient disentanglement of the instance-level discriminative features from dataset-wide adversarial pattern distributions. To this end, we propose an instance-driven advancing targeted attacks (IATA) framework with transferable pattern embedding. IATA achieves high transferability by integrating dual-branch pattern injection and local patch guided adversarial attacks. Specifically, the instance-driven adversarial generator and the prototype discriminator are designed to fuse instance-level embeddings from a specific target and the prototype-level features of the target class, respectively. The local patch tuning strategy is designed to enhance the dual-substitute model discrepancy attack by stimulating more texture perturbations. Extensive experiments demonstrate that IATA achieves state-of-the-art performance in the black-box setting and significantly outperforms existing targeted attack methods by a margin of 14.98 %, as measured by the targeted transfer attack success rate when Inception-v3 is used as the substitute model.}, archive = {J_NEUCOM}, author = {Litan Sun and Xiaofeng Qu and Guangrun Chen and Haokun Geng and Jian Zhu and Sijie Niu}, doi = {10.1016/j.neucom.2025.131412}, journal = {Neurocomputing}, month = {11}, pages = {131412}, shortjournal = {Neurocomputing}, title = {IATA: Instance-driven advancing targeted attacks with transferable pattern embedding}, volume = {655}, year = {2025}, }
• (2025). Deep binary hyperbolic embedding for large-scale image retrieval. NEUCOM, 655, 131411. (www)

  @article{ , abstract = {With the rapid development of the web and mobile devices, the scale of multimedia data has grown exponentially. Efficiently retrieving target information from massive datasets has become an important research topic in the multimedia area. Benefiting from advantages such as fast retrieval speed and low storage cost, hashing methods have gradually become mainstream techniques in massive multimedia retrieval tasks. Most existing deep hashing methods commonly employ Euclidean space as the embedding space to measure the semantic similarity among raw samples. However, since the volume of Euclidean space grows polynomially, it struggles to effectively model latent semantic structures, especially when dealing with hierarchical semantics that are widely present in the real world. To address the above problem, by mapping the learned embeddings to hyperbolic space, a novel Deep Binary Hyperbolic Embedding (DBHE) framework is proposed to generate high-quality discrete descriptors. Specifically, based on the Poincaré ball, the feature embeddings obtained from Vision Transformer are projected into hyperbolic space, which better represents the hierarchical semantic structures among raw samples. A pairwise cross-entropy loss based on hyperbolic space is designed to encourage embeddings of semantically similar samples to stay close, while pushing those of dissimilar samples apart, thereby improving the discriminability of hash codes. Extensive experiments on three public datasets show that, compared with the Euclidean setting, our proposed DBHE achieves significant performance improvements across multiple retrieval evaluation metrics, fully demonstrating its advantages in image retrieval applications. The source code is available at https://github.com/QinLab-WFU/DBHE .}, archive = {J_NEUCOM}, author = {Mengru Zhang and Enhao Wang and Qibing Qin and Jinkui Hou and Wenfeng Zhang and Lei Huang}, doi = {10.1016/j.neucom.2025.131411}, journal = {Neurocomputing}, month = {11}, pages = {131411}, shortjournal = {Neurocomputing}, title = {Deep binary hyperbolic embedding for large-scale image retrieval}, volume = {655}, year = {2025}, }
• (2025). Quantum splitting convolutional neural network-based distributed quantum disease detection model. NEUCOM, 655, 131410. (www)

  @article{ , abstract = {Quantum computing leverages the properties of superposition and entanglement to enable high-dimensional feature extraction from medical images, demonstrating its potential to improve the accuracy of disease detection. However, practical applications are still constrained by the current performance bottlenecks of quantum hardware and affected by the excessive entangling gate operations. In this work, we propose a distributed quantum disease detection (DQDD) model using quantum splitting convolutional neural network. The DQDD employs a hybrid quantum convolutional neural network (QCNN) to capture low-dimensional and high-dimensional features in medical images. We introduce the quantum circuit splitting technique to our model, reconstructing an 8-qubit QCNN with 5 qubits, while simultaneously reducing the entangling operations in each sub-circuit. The model comprises two core components, the Information Distillation mechanism designed to extract pivotal features and the Non-saliency Filtering mechanism dedicated to eliminating irrelevant patterns. Both components operate cohesively, facilitating parallel computation across single or multiple quantum computing devices. Evaluations on ISIC2017 and HAM10000 skin cancer datasets demonstrate that our model outperforms existing quantum models in skin cancer detection, doing so with fewer parameters. Superior generalization capability is further validated through cross-domain testing on the MedMNIST dataset.}, archive = {J_NEUCOM}, author = {Yangyang Li and Zhengya Qi and Yuelin Li and Haorui Yang and Ronghua Shang and Licheng Jiao}, doi = {10.1016/j.neucom.2025.131410}, journal = {Neurocomputing}, month = {11}, pages = {131410}, shortjournal = {Neurocomputing}, title = {Quantum splitting convolutional neural network-based distributed quantum disease detection model}, volume = {655}, year = {2025}, }
• (2025). A survey on self-supervised learning: Recent advances and open problems. NEUCOM, 655, 131409. (www)

  @article{ , abstract = {Despite recent breakthroughs, deep learning still relies heavily on costly manual annotations of large-scale datasets to achieve optimal performance in supervised learning. Self-Supervised Learning (SSL) has emerged as a promising approach to leverage unlabeled data without the need for manual labeling. This survey provides a systematic and comprehensive review of state-of-the-art SSL methods in various fields, including image processing, video, audio, speech, and Natural Language Processing (NLP). We begin with an overview that outlines how this survey differs from existing ones and highlights its main contributions. We then analyze and categorize SSL pretext tasks, offering a detailed classification that divides SSL methods into six categories. For each category, we explore the frameworks, approaches, and performance comparisons, emphasizing their advantages and disadvantages. Following this, we discuss the application domains of SSL. Finally, we examine emerging trends and open challenges in SSL, providing insights into the difficulties of advancing SSL models and suggesting directions for future research and development.}, archive = {J_NEUCOM}, author = {Jianping Zhang and Lei Yang and Seyed Mahmoud Sajjadi Mohammadabadi and Feng Yan}, doi = {10.1016/j.neucom.2025.131409}, journal = {Neurocomputing}, month = {11}, pages = {131409}, shortjournal = {Neurocomputing}, title = {A survey on self-supervised learning: Recent advances and open problems}, volume = {655}, year = {2025}, }
• (2025). Dual-level information fusion-boosted fisher embedding discriminative dictionary learning for few-shot recognition. NEUCOM, 655, 131408. (www)

  @article{ , abstract = {Dictionary Learning (DL) methods have made remarkable contributions in the realm of health monitoring, attributed to their exceptional representation capabilities. In practical scenarios, abnormal samples are frequently exceedingly scarce compared to normal samples, posing a significant obstacle to health status recognition. This study introduces a dual-level information fusion-boosted Fisher embedding discriminative dictionary learning (BFEDDL) method tailored for few-shot recognition tasks. Initially, multiple-phase information is obtained and fused at the data level based on the shift-invariance property of the signals to construct an initial dictionary with augmented feature representation ability. To elevate the efficacy of DL methods, the BFEDDL objective function incorporates the discriminative fidelity and Fisher discriminative constraints, where the latter are simultaneously applied to dictionary atoms, encoding coefficients and their transposes, ensuring that the learned dictionary not only reconstructs the data accurately but also possesses strong discriminative power. Moreover, feature-level information fusion is implemented through dictionary expansion and compression, which further refines the discriminative abilities of the dictionary. Two case studies involving electrocardiographic data and a high-speed train wheelset bearing monitoring dataset are presented and the results demonstrate that BFEDDL outperforms several state-of-the-art methods, highlighting its potential for few-shot recognition in practical applications.}, archive = {J_NEUCOM}, author = {Xin Zhang and Zhongqiang Zhang and Lei Wu and Laihao Yang and Jiaxu Wang}, doi = {10.1016/j.neucom.2025.131408}, journal = {Neurocomputing}, month = {11}, pages = {131408}, shortjournal = {Neurocomputing}, title = {Dual-level information fusion-boosted fisher embedding discriminative dictionary learning for few-shot recognition}, volume = {655}, year = {2025}, }
• (2025). STNet: Seasonal-trend network for multivariate time series forecasting. NEUCOM, 655, 131407. (www)

  @article{ , abstract = {Time series forecasting is widely used in fields such as economics, energy, traffic management, and weather prediction. However, real-world time series often exhibit complex non-stationarity, with trends, seasonality, and noise deeply intertwined, posing significant challenges for accurate forecasting. To address these challenges, this paper proposes STNet (Seasonal-Trend Network) , a time series forecasting model based on a dual-branch architecture. STNet employs an Exponentially Weighted Moving Average (EWMA) decomposition framework to separate time series into trend and seasonal components, which are then processed by a Multi-Scale Convolutional Gated Unit (MCGU) and a combination of Channel-Independent Linear (CIL), Sequence Feature Extractor (SFE), and Channel Feature Extractor (CFE) layers. This design allows STNet to effectively disentangle and model the distinct characteristics of trend and seasonal components. Experimental results demonstrate that STNet achieves superior performance on multiple real-world datasets, excelling in both long-term and short-term forecasting tasks. STNet not only offers high computational efficiency but also maintains robust forecasting accuracy when dealing with complex time series.}, archive = {J_NEUCOM}, author = {Shuai Li and Cheng Zhao and Jie Zhou}, doi = {10.1016/j.neucom.2025.131407}, journal = {Neurocomputing}, month = {11}, pages = {131407}, shortjournal = {Neurocomputing}, title = {STNet: Seasonal-trend network for multivariate time series forecasting}, volume = {655}, year = {2025}, }
• (2025). Online bandit optimization with stochastic inequality constraints. NEUCOM, 655, 131406. (www)

  @article{ , abstract = {This paper investigates the online bandit optimization problem under long-term stochastic inequality constraints. The constraint set consists of the intersection of a fixed set and a stochastic function constraint set, where the stochastic function constraint set is generated independently and identically distributed at each round. Instead of requiring the constraints to be satisfied in every round, we require them to be satisfied over the long run, which makes the constraint more general. The one- and the two-point bandit drift-plus-penalty algorithms are proposed to solve this problem. The performance of the proposed algorithms is evaluated using regret and cumulative constraint violation, and it is shown that both are sublinear. Finally, the algorithms are verified to be compatible with numerical simulation.}, archive = {J_NEUCOM}, author = {Yuanyuan Li and Deming Yuan and Baoyong Zhang and Ju H. Park}, doi = {10.1016/j.neucom.2025.131406}, journal = {Neurocomputing}, month = {11}, pages = {131406}, shortjournal = {Neurocomputing}, title = {Online bandit optimization with stochastic inequality constraints}, volume = {655}, year = {2025}, }
• (2025). Phys-vim: State space model for remote physiological measurement. NEUCOM, 655, 131405. (www)

  @article{ , abstract = {Remote photoplethysmography (rPPG) is a contact-free and cost-effective method for video physiological signal estimation. Traditional networks in rPPG like CNNs are limited by local inductive bias while transformers require large datasets for training and have quadratic computational complexity. Moreover, existing methods often struggle to accurately extract periodic signals in long- range contexts, especially in the presence of noise or unseen scenarios. To address these challenges, we introduce Phys-Vim, a novel method based on the Vision Mamba (Vim), an advanced state space model for vision tasks. Phys-Vim can capture long-range dependencies with linear computational complexity. Additionally, we propose a Multi-Path Selective Scan (MPSS) mechanism to improve the consistency and global awareness for signal extraction from video. Extensive evaluations demonstrate that Phys-Vim achieves state-of-the-art performance. In the cross-dataset evaluation from UBFC-rPPG to PURE, Phys-Vim reduces MAE by 73 % and increases Pearson correlation by 8 % compared to previous best results. Moreover, it requires far fewer parameters and reduces computational cost by 11.5 % compared to the previous transformer-based Spiking-PhysFormer.}, archive = {J_NEUCOM}, author = {Zhiyuan Chang and Haoxiang Wang and Yuhao Han and Jianming Lv and Bilian Li}, doi = {10.1016/j.neucom.2025.131405}, journal = {Neurocomputing}, month = {11}, pages = {131405}, shortjournal = {Neurocomputing}, title = {Phys-vim: State space model for remote physiological measurement}, volume = {655}, year = {2025}, }
• (2025). A new learning algorithm based on neuronal activity degree for spiking neural network. NEUCOM, 655, 131402. (www)

  @article{ , abstract = {Nowadays, spiking neural networks (SNNs) with great computing capacity and low-power consumption have been widely applied in many fields. Due to their inherent discontinuity and complex nonlinear mechanisms, improving the performance of SNNs at the learning algorithm level has become a popular research topic. Unfortunately, some details inside neurons are often overlooked, and only the firing times are transmitted to postsynaptic neurons. Since the time cost for firing by neurons varies in both real biological networks and SNNs, this paper presents a new concept, called neuronal activity degree, to extract latent interval information underlying neurons. The neuronal activity degree is then embedded within the SpikeProp algorithm. This new algorithm ensures that a more active neuron, as determined by its neuronal activity degree, will have a greater impact on the change in the weight of the synapses connected to it during the learning process. The experimental results on five benchmark static datasets and two neuromorphic datasets show that the proposed method, ADSNN, improves the accuracy, generalization and stability of SNNs.}, archive = {J_NEUCOM}, author = {Lu Zhang and Fang Liu and Jie Yang and Wei Wu}, doi = {10.1016/j.neucom.2025.131402}, journal = {Neurocomputing}, month = {11}, pages = {131402}, shortjournal = {Neurocomputing}, title = {A new learning algorithm based on neuronal activity degree for spiking neural network}, volume = {655}, year = {2025}, }
• (2025). Computational modeling for gratings stimulated gamma oscillations in a large-scale cortical neuronal network. NEUCOM, 655, 131401. (www)

  @article{ , abstract = {Biological experiments have shown that the oscillatory properties of gamma oscillation (30–80 Hz) in the primary visual cortex, which may be related to visual information transmission and integration, can be modulated by visual stimuli. However, the visual stimulus-modulation mechanisms of gamma oscillation are not fully unveiled. In this paper, a comprehensive computational model incorporating global drifting sinusoidal grating and a large-scale network of primary visual cortex is constructed to investigate this issue. The model reproduces the stimulus-dependent phenomena of gamma oscillation in a series of biological experiments, including the dependence phenomena of the spectral power and peak frequency of gamma oscillation on grating’s orientation, contrast, and temporal frequency, respectively. In addition, the analysis of spike clusters and input current fluctuations found that the neuronal properties and particular function-specific connectivity patterns play important roles in modulating gamma oscillation in the cortex network. Specifically, the orientation selectivity of neurons determines the spectral power, whereas the push–pull scheme of synapses in layer 4 prevents the peak frequency from changing; the difference between excitatory and inhibitory neurons makes the peak frequency proportional to the grating’s temporal frequency and causes the power turning point to vary with grating contrast. The study provides a comprehensive computational model and approach for deeply understanding the gamma oscillatory phenomena in the primary visual cortex. It can help to understand the dynamic mechanisms of visual stimulus-evoked gamma oscillations and help predict and guide relevant biological experiments; moreover, it may lay the foundation for further research on the cognitive function of related oscillations.}, archive = {J_NEUCOM}, author = {Hao Yang and Fang Han and Wenlian Lu and Qingyun Wang}, doi = {10.1016/j.neucom.2025.131401}, journal = {Neurocomputing}, month = {11}, pages = {131401}, shortjournal = {Neurocomputing}, title = {Computational modeling for gratings stimulated gamma oscillations in a large-scale cortical neuronal network}, volume = {655}, year = {2025}, }
• (2025). Joint multi-grained similarity contrastive learning for video-text retrieval. NEUCOM, 655, 131392. (www)

  @article{ , abstract = {With the popularity of short videos and the dramatic increase in the number of users, video–text retrieval has received increasing attention from researchers. However, there are still some bottlenecks in the existing works. Firstly, many previous works have not well considered the interaction between the multi-grained features of different modalities, which results in the modal information not being fully exploited. Secondly, these works focus solely on inter-modal interactions and contrasts, neglecting the contrasts of multi-grained features within modalities, which makes the similarity measurement less accurate. Finally, similarities at different granularities are not effectively fused so that truly meaningful instance-level similarity scores are not formed. To address these bottlenecks, we devised an efficient video–text retrieval model, called Joint Multi-grained Similarity Contrastive Learning for Video–Text Retrieval (JM-VTR). Firstly, the method employs a multidimensional contrastive strategy of two modal features, including inter-modal and intra-modal multi-grained feature contrasts. Secondly, to fuse the various contrastive similarities well, we also design a joint modal attention module to fuse the various similarities into a final joint modal multi-grained similarity score. Extensive experiments show that our approach outperforms existing state-of-the-art methods by 9.2 % , 5.6 % , 13.5 % , and 11.7 % of R@10 for text-to-video retrieval task on four benchmarks, respectively.}, archive = {J_NEUCOM}, author = {Mingyong Li and Mingyuan Ge and Wenfeng Zhang}, doi = {10.1016/j.neucom.2025.131392}, journal = {Neurocomputing}, month = {11}, pages = {131392}, shortjournal = {Neurocomputing}, title = {Joint multi-grained similarity contrastive learning for video-text retrieval}, volume = {655}, year = {2025}, }
• (2025). Instance-aware global re-detection for precise and efficient long-term visual tracking. NEUCOM, 655, 131391. (www)

  @article{ , abstract = {In long-term visual tracking, the target often disappears from the field of view due to full occlusion or being out-of-view, and when the target reappears in the field of view, the tracking algorithm needs to re-detect it in order to achieve continuous tracking. Therefore, global re-detection capability is crucial for long-term visual tracking. Existing global re-detection methods suffer from high computational overhead, slow speed, and poor performance. To address these problems, this paper proposes an instance-aware global re-detection method (IGR). Firstly, we introduce the one-stage and anchor-free framework in object detection as the overall structure for our re-detection, which only requires a single feature modulation and significantly improves tracking speed. Then, we propose an instance-guided feature modulation module, which enhances the algorithm’s instance-awareness through improved feature mixing and enhancement. Finally, we introduce an instance-level dynamic sample selection strategy that adaptively balances positive and negative samples and improves the model’s robustness against interference from similar objects. The proposed re-detection method is combined with four local trackers (EVPTrack, OSTrack, TransT, SMAT) to constitute a long-term visual tracking method, and the performance is evaluated on six datasets (LaSOT, LaSOT e x t , UAV123, VideoCube, UAV20L, VOT2020LT). Experimental results show that IGR not only achieves substantial overall performance gains (with a 6.5 % improvement in success rate on LaSOT compared to the baseline), but also brings significant improvements to local trackers. For example, IGR-TransT improves success rate by 10.5 % on UAV20L and increases normalized precision by 7.9 % on VideoCube. Among the proposed variants, the best-performing method—IGR-EVPTrack—achieves a speed of up to 58 FPS, which enables precise and efficient long-term visual tracking. Code and trained models will be available at https://github.com/rvpagehhj/IGR .}, archive = {J_NEUCOM}, author = {Zhiqiang Hou and Hao Yue and Xianxin Jia and Sugang Ma and Xiaobao Yang and Lei Pu}, doi = {10.1016/j.neucom.2025.131391}, journal = {Neurocomputing}, month = {11}, pages = {131391}, shortjournal = {Neurocomputing}, title = {Instance-aware global re-detection for precise and efficient long-term visual tracking}, volume = {655}, year = {2025}, }
• (2025). Parameter-efficient hierarchical spatio-temporal graph network with specialized label smoothing for qualitative EEG emotion recognition. NEUCOM, 655, 131390. (www)

  @article{ , abstract = {This study introduces a parameter-efficient Hierarchical Spatial Temporal Network (HiSTN) specifically designed for the task of emotion classification using multi-channel electroencephalogram (EEG) data. The network incorporates a graph hierarchy constructed from bottom-up to various levels of abstraction, offering the dual-advantage of enhanced task-relevant deep feature extraction alongside a lightweight design. The model’s effectiveness is further amplified when used in conjunction with a proposed special label smoothing method. Comprehensive benchmark experiments reveal that this combined approach yields high, balanced performance in terms of both quantitative and qualitative predictions. HiSTN, which has approximately 1000 parameters, achieves mean F1 scores of 96.82 % (valence) and 95.62 % (arousal) in subject-dependent tests on the rarely-utilized 5-classification task problem from the DREAMER dataset. In subject-independent settings, the same model yields mean F1 scores of 78.34 % for valence, and 81.59 % for arousal. The adoption of the proposed Sequential Top-2 Hit Rate (Seq2HR) metric helps reveal improvements surpassing 50 % in subject-dependent tasks and 30 % in subject-independent tasks achieved through our approach when compared to the common case of training with regular one-hot labels. These results, when jointly examined, further highlight the significant enhancements our proposed methods can bring to relevant prediction tasks in terms of the balance between models’ quantitative and qualitative performance. The study also includes relevant ablation studies and case explorations to further elucidate the workings of the proposed model and enhance its interpretability. Codes relevant to this paper are available on Github at: https://github.com/dykuang/EEG-based-affective-computing .}, archive = {J_NEUCOM}, author = {Dongyang Kuang and Xinyue Song and Craig Michoski}, doi = {10.1016/j.neucom.2025.131390}, journal = {Neurocomputing}, month = {11}, pages = {131390}, shortjournal = {Neurocomputing}, title = {Parameter-efficient hierarchical spatio-temporal graph network with specialized label smoothing for qualitative EEG emotion recognition}, volume = {655}, year = {2025}, }
• (2025). EnfoMax: Domain entropy and mutual information maximization for domain generalized face anti-spoofing. NEUCOM, 655, 131388. (www)

  @article{ , abstract = {The face anti-spoofing (FAS) methods perform well under intra-domain setups. However, their cross-domain performance is unsatisfactory. As a result, the domain generalization (DG) method has gained more attention in FAS. Existing methods treat FAS as a simple binary classification task and propose a heuristic training objective to learn domain-invariant features. However, there is no theoretical explanation of what a domain-invariant feature is. Additionally, the lack of theoretical support makes domain generalization techniques such as adversarial training lack training stability. To address these issues, this paper proposes the EnfoMax framework, which uses information theory to analyze cross-domain FAS tasks. This framework provides theoretical guarantees and optimization objectives for domain generalized FAS tasks. EnfoMax maximizes the domain entropy and mutual information of live samples in source domains without using adversarial learning. Experimental results demonstrate that our approach performs well on extensive public datasets and outperforms state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Tianyi Zheng and Bo Li and Shuang Wu and Ben Wan and Guodong Mu and Shice Liu and Shouhong Ding and Jia Wang}, doi = {10.1016/j.neucom.2025.131388}, journal = {Neurocomputing}, month = {11}, pages = {131388}, shortjournal = {Neurocomputing}, title = {EnfoMax: Domain entropy and mutual information maximization for domain generalized face anti-spoofing}, volume = {655}, year = {2025}, }
• (2025). Parallel dual-branch network with multi-scale features for unsupervised domain adaption person re-identification. NEUCOM, 655, 131387. (www)

  @article{ , abstract = {Extracting discriminative features is a significant challenge for unsupervised person re-identification (Re-ID) when training from the source domain to the target domain. Most unsupervised domain adaptive methods for person Re-ID rely on single-scale appearance information during feature extraction. These methods may overlook the potentially useful explicit information of other different scales, and the pseudo-labels obtained through clustering often contain noise. Furthermore, relying on a single branch network model structure can lead to learning stagnation, limiting model generalizability. To resolve these problems, this paper proposes a parallel dual-branch network (PDBN) which comprises two branches: a backbone branch and a pyramid branch, PDBN leverages multi-scale pyramid branch features and global backbone branch features to jointly learn discriminative features. The two branches are not independent to each other but synergistically correlated. Specifically, we formulate a novel mutual teaching PDBN training strategy for unsupervised domain adaptation in person Re-ID by hard classification loss, hard triplet loss, soft classification loss and soft triplet loss. The backbone branch aims to maximise global feature discriminative capability, while the pyramid branch extracts multi-scale details from the backbone using a feature pyramid network to concurrently optimise the underlying complementary advantages across scales. To reduce pseudo-label noise, Segmenting Dynamic Clustering (SDC) is introduced, which divides training into two phases. In phase 1, SDC uses backbone branch features for clustering to ensure accurate feature learning. In phase 2, SDC fuses features from both branches by max operation as the final discriminative features for clustering, enabling interactive reinforcement between them. Extensive experiments conducted on Market1501, DukeMTMC-reID, and MSMT17 demonstrate the superiority of our method compared with state-of-the-art methods for unsupervised cross-domain person Re-ID.}, archive = {J_NEUCOM}, author = {Hao Li and Shiyi Lei and Yuyang Feng and Xin Zhao and Tao Zhang}, doi = {10.1016/j.neucom.2025.131387}, journal = {Neurocomputing}, month = {11}, pages = {131387}, shortjournal = {Neurocomputing}, title = {Parallel dual-branch network with multi-scale features for unsupervised domain adaption person re-identification}, volume = {655}, year = {2025}, }
• (2025). MambaAD: Multivariate time series anomaly detection in IoT via multi-view mamba. NEUCOM, 655, 131385. (www)

  @article{ , abstract = {Multivariate time series anomaly detection (MTSAD) in Internet of Things (IoT) systems is a crucial area of research that aims to increase cybersecurity, prevent disruptions and improve quality of service. To cope with the growing complexity of sensory data, recent approaches emphasize learning the temporal dynamics of each signal with Transformers and capturing correlations between signals with graph learning. However, these methods still struggle with two major challenges. First, the deep models using Transformer architecture and graph structure are computationally inefficient due to their quadratic complexity, which limits their scalability and applicability. Second, it is difficult to learn the generalized latent patterns from the limited but highly sensitive training data, which makes the models prone to overfitting. To address these challenges, we propose MambaAD, an efficient anomaly detection framework based on the linear-time state space model. Specifically, we segment and tokenize the time series to retain local semantic information in the embeddings. Then, we develop a bidirectional Mamba module to extract temporal dependencies and inter-signal correlations from different perspectives. Finally, the captured features are fused and projected for reconstruction and scoring. We also provide a multi-view token masking strategy and a contrastive learning mechanism to elevate both representation quality and generalization performance. Thorough experimentation with six real-world datasets across different fields indicates that the proposed method surpasses current state-of-the-art benchmarks in terms of both efficiency and accuracy.}, archive = {J_NEUCOM}, author = {Shuxin Qin and Jing Zhu and Aipeng Guo and Yansong Yang and Lu Wang and Gaofeng Tao}, doi = {10.1016/j.neucom.2025.131385}, journal = {Neurocomputing}, month = {11}, pages = {131385}, shortjournal = {Neurocomputing}, title = {MambaAD: Multivariate time series anomaly detection in IoT via multi-view mamba}, volume = {655}, year = {2025}, }
• (2025). S2Trans: Structured spectrum transformer for robust unsupervised video object segmentation. NEUCOM, 655, 131384. (www)

  @article{ , abstract = {Unsupervised Video Object Segmentation (UVOS) aims to identify and segment spatiotemporal regions of the most prominent object in videos. However, achieving accurate and robust segmentation in complex scenarios is challenging due to imbalanced multimodal features and insufficient fine-grained structural representation. To address these issues, we propose the Structured Spectrum Transformer (S2Trans) for UVOS. S2Trans decouples the Transformer into temporal and spatial components, facilitating the integration of locally correlated optical flow to guide spatial attention. Specifically, we introduce a Guided Bimodal Synthesis (GBS) module that dynamically balances the contributions of video frame and flow features by leveraging their interactions, while effectively suppressing misleading information. To capture multi-scale structural and textural details of objects, which are crucial for accurate UVOS, we propose the Dual-scale Spectrum Gated Feed-Forward (DSGFF). Additionally, to relieving global context loss, we design a Hybrid Context spatial Multi-Head Self-Attention (HC-MHSA) mechanism that integrates global tokens into window-based attention. Extensive experiments on three mainstream UVOS benchmarks demonstrate that the proposed framework achieves state-of-the-art performance, validating its effectiveness in balancing multimodal contributions and enriching feature representation.}, archive = {J_NEUCOM}, author = {Chao Zhang and Jiaqing Fan and Mengjuan Jiang and Fanzhang Li}, doi = {10.1016/j.neucom.2025.131384}, journal = {Neurocomputing}, month = {11}, pages = {131384}, shortjournal = {Neurocomputing}, title = {S2Trans: Structured spectrum transformer for robust unsupervised video object segmentation}, volume = {655}, year = {2025}, }
• (2025). G-NodeMixup: Enhancing graph neural networks reachability under extremely limited labels. NEUCOM, 655, 131378. (www)

  @article{ , abstract = {Graph Neural Networks ( Gnns ) have shown remarkable performance in semi-supervised node classification, but their effectiveness diminishes in settings with extremely limited labeled data. The scarcity of labeled nodes leads to an under-reaching issue, where unlabeled nodes receive insufficient supervision, resulting in poor generalization. In this paper, we propose G-NodeMixup , a generalized extension of our previously proposed NodeMixup framework which was designed to address under-reaching by improving communication between labeled and unlabeled nodes. G-NodeMixup introduces three novel components: (1) Multi-set Pairing, which facilitates mixup between Labeled-Labeled, Labeled-Unlabeled, and Unlabeled-Unlabeled nodes to enhance node interactions and promote smoother decision boundaries; (2) Subgraph-based Mixup, which focuses mixup within k -hop subgraphs to preserve graph locality and avoid disruptive global edge modifications; and (3) Consistency Regularization-based Mixup Loss, which reduces reliance on noisy pseudo-labels by enforcing consistency between mixed node predictions. Our framework remains architecture-agnostic and can be applied to various Gnn models without requiring significant architectural changes or excessive computational overhead. Experimental results across several benchmark datasets demonstrate that G-NodeMixup consistently improves Gnn performance in extremely limited labeled settings, achieving state-of-the-art results and establishing its practical effectiveness.}, archive = {J_NEUCOM}, author = {Ziyu Guan and Beilei Ling and Weigang Lu and Meng Yan and Yaming Yang and Wei Zhao and Yibing Zhan and Dapeng Tao}, doi = {10.1016/j.neucom.2025.131378}, journal = {Neurocomputing}, month = {11}, pages = {131378}, shortjournal = {Neurocomputing}, title = {G-NodeMixup: Enhancing graph neural networks reachability under extremely limited labels}, volume = {655}, year = {2025}, }
• (2025). Interactive triplet attention for few-shot fine-grained image classification. NEUCOM, 655, 131377. (www)

  @article{ , abstract = {Few-shot fine-grained classification aims to identify novel fine-grained classes from extremely few examples with ultra-high semantic similarity between classes, hence a notoriously hard task. To extract discriminative features from few samples for recognizing subtle differences between fine-grained classes , it is pivotal to exploit comprehensive interactions across all dimensions in space and channel, which, however, is unexplored yet by state-of-the-art methods in this challenging area. To address this issue, in this paper we show that a simple adjustment to the existing triplet attention module (TAM) can be highly effective for few-shot fine-grained image classification. More specifically, building on TAM which comprises three parallel branches for pairwise interactions between height, width, and channel dimensions, we introduce an additional interaction between the outputs of these three branches, capable of modeling the dependency across all three dimensions; the revised method is dubbed interactive triplet attention module (ITAM). ITAM is a plug-and-play module, which can be inserted into any metric-based few-shot fine-grained image classifiers for performance enhancement. Extensive experiments, on CUB-200-2011, Flowers, Stanford-Cars, and Stanford-Dogs, showcase the superiority of ITAM against state-of-the-art few-shot fine-grained image classifiers.}, archive = {J_NEUCOM}, author = {Xiaoxu Li and Shaoying Xue and Jiyang Xie and Xiaochen Yang and Zhanyu Ma and Jing-Hao Xue}, doi = {10.1016/j.neucom.2025.131377}, journal = {Neurocomputing}, month = {11}, pages = {131377}, shortjournal = {Neurocomputing}, title = {Interactive triplet attention for few-shot fine-grained image classification}, volume = {655}, year = {2025}, }
• (2025). IMSDO: Deep metric learning with incremental margin and standard deviation optimization. NEUCOM, 655, 131376. (www)

  @article{ , abstract = {This study investigates two research questions in deep metric learning, each motivated by practical intuition and theoretical grounding. First, we hypothesize that systematically increasing the margin during training can guide the model to learn from easier to harder tasks, leading to better convergence and generalization. This approach is similar to warm-up learning rate and curriculum learning strategies. Second, we propose that controlling the standard deviation of feature maps through a new loss function can suppress noise and produce more consistent and discriminative representations. To validate these hypotheses, we introduce IMSDO, a method that: (1) gradually increases the margin using an increment function during training to implement a structured learning process; and (2) minimizes the standard deviation of feature maps to enhance their quality. Despite its simplicity, IMSDO significantly improves the performance of the triplet loss and outperforms several strong baselines across standard benchmarks. These results highlight IMSDO’s practical value and open new directions for research in metric learning.}, archive = {J_NEUCOM}, author = {Jeremy Winston and Dae-Ki Kang}, doi = {10.1016/j.neucom.2025.131376}, journal = {Neurocomputing}, month = {11}, pages = {131376}, shortjournal = {Neurocomputing}, title = {IMSDO: Deep metric learning with incremental margin and standard deviation optimization}, volume = {655}, year = {2025}, }
• (2025). UCFFormer: Recognizing human actions from multimodal sensors using unified contrastive fusion transformer. NEUCOM, 655, 131374. (www)

  @article{ , abstract = {Human Action Recognition (HAR) is fundamental to intelligent systems, enabling machines to accurately interpret and respond to human activities. Multimodal sensor fusion has emerged as a key approach to enhancing HAR performance by leveraging complementary information from diverse sensing modalities. In this paper, we propose the Unified Contrastive Fusion Transformer (UCFFormer), a novel multimodal fusion architecture designed to integrate heterogeneous sensor data for robust HAR in AI-driven applications. UCFFormer employs a unified transformer framework to model interdependencies across both temporal and modality domains, effectively capturing cross-modal interactions. To improve computational efficiency, we introduce the Factorized Time-Modality Transformer (FTMT), which reduces the complexity of self-attention while preserving rich contextual representations. Additionally, we propose the Multimodal Contrastive Alignment Network (MCANet), which utilizes contrastive loss to align feature distributions across modalities, ensuring semantically consistent feature fusion. Extensive experiments on three benchmark datasets demonstrate the state-of-the-art performance of UCFFormer: 99.99 % Top-1 accuracy on UTD-MHAD, setting a new record; 97.5 % cross-subject and 99.7 % cross-view accuracies on NTU RGB+D. These concrete results highlight the effectiveness of our unified contrastive fusion transformer for robust human action recognition.}, archive = {J_NEUCOM}, author = {Kyoung Ok Yang and Junho Koh and Jun Won Choi}, doi = {10.1016/j.neucom.2025.131374}, journal = {Neurocomputing}, month = {11}, pages = {131374}, shortjournal = {Neurocomputing}, title = {UCFFormer: Recognizing human actions from multimodal sensors using unified contrastive fusion transformer}, volume = {655}, year = {2025}, }
• (2025). Dynamic-unit discrete hopfield neural network with supervised preprocessing phase: Optimizing the logic mining using S-type random 2 satisfiability for classification. NEUCOM, 655, 131373. (www)

  @article{ , abstract = {Logic mining, a specialized area of data mining, extracts logical rules from datasets by examining their underlying behavior. However, a critical limitation of many existing models is their insufficient focus on ensuring the quality of processed data for knowledge extraction. This often leads to challenges in the transparency and reliability of derived outcomes. Consequently, maintaining robust data quality, reliability, and availability is essential for producing meaningful and interpretable knowledge. This paper introduces a novel logic mining model designed to enhance both the quantity and quality of induced logic by capturing all optimal solutions. The proposed model incorporates a supervised preprocessing phase that identifies optimal attributes using Elastic Net regularization for robust feature selection. To enhance logical diversity, a non-systematic S-Type Random 2 Satisfiability structure is introduced, guided by probability distributions that reflect dataset characteristics. This probabilistic logic operates within a novel Dynamic-Unit Discrete Hopfield Neural Network, which ensures comprehensive solution space coverage through dynamic Content Addressable Memory units, leading to the production of diverse and optimal induced logical rules. Furthermore, a newly developed shifting operator in the retrieval phase facilitates the optimal arrangement of attributes in the final induced logic. Experimental evaluations demonstrate that the proposed model significantly outperforms existing approaches, achieving Accuracy 82 %, Specificity of 99.8 % and a Negative Predictive Value of 85.6 %. These findings highlight the effectiveness of the supervised preprocessing phase, the implementation of a non-systematic probabilistic structure, and the efficiency of the dynamic unit network in advancing logic mining capabilities.}, archive = {J_NEUCOM}, author = {Suad Abdeen and Mohd Shareduwan Mohd Kasihmuddin and Muhammad Fadhil Marsani and Nur Ezlin Zamri and Gaeithry Manoharam and Mohd. Asyraf Mansor and Jia Li}, doi = {10.1016/j.neucom.2025.131373}, journal = {Neurocomputing}, month = {11}, pages = {131373}, shortjournal = {Neurocomputing}, title = {Dynamic-unit discrete hopfield neural network with supervised preprocessing phase: Optimizing the logic mining using S-type random 2 satisfiability for classification}, volume = {655}, year = {2025}, }
• (2025). Adaptive motion enhancement for passive non-line-of-sight action recognition. NEUCOM, 655, 131372. (www)

  @article{ , abstract = {Most current recognition methods for human action are designed for line-of-sight (LOS) scenarios, where the targets are assumed to be directly visible. However, in numerous action recognition applications, we often encounter non-line-of-sight (NLOS) situations, e.g., rescue operations, security, and autonomous vehicles. In this case, it becomes imperative that we develop methods that can effectively identify human actions outside the line of sight, such as those behind obstacles or around corners. It is also important to note that most existing NLOS approaches rely on expensive active imaging equipment, which hinders practical deployment. To address these challenges, we propose AME-Net (Adaptive Motion Enhancement Network), a novel passive NLOS action recognition framework that identifies human actions by analyzing reflections on visible relay walls using only standard RGB cameras. AME-Net adaptively amplifies subtle motion cues and mitigates environmental variability, enabling accurate and robust recognition in NLOS conditions. Furthermore, we introduce NLOS-Action, the first dataset specifically designed for passive NLOS action recognition, containing both synthetic and real-world sequences. Pursuant to our extensive experiments based on the dataset, this paper convincingly demonstrates effectiveness and practicality of the proposed AME-Net for NLOS action recognitions.}, archive = {J_NEUCOM}, author = {Zhongqi Sun and Yuan Zhou and Shuwei Huo and Sun-Yuan Kung}, doi = {10.1016/j.neucom.2025.131372}, journal = {Neurocomputing}, month = {11}, pages = {131372}, shortjournal = {Neurocomputing}, title = {Adaptive motion enhancement for passive non-line-of-sight action recognition}, volume = {655}, year = {2025}, }
• (2025). Generating realistic synthetic tabular data with integrated LLM and diffusion models. NEUCOM, 655, 131371. (www)

  @article{ , abstract = {Generating realistic synthetic tabular data is a crucial task for privacy-preserving data sharing, data augmentation, and learning from limited samples. However, existing methods often struggle with small sample sizes, heterogeneous feature types, and generalization in real-world scenarios. We propose TabularMDLM, a novel diffusion-based generative framework that integrates masked language modeling to synthesize high-quality tabular data. Unlike prior work, TabularMDLM applies noise only to feature values—preserving the semantic relationship between column names and values—and leverages pre-trained language models to iteratively reconstruct masked tokens during reverse diffusion. This design enhances generation quality while ensuring structural consistency. To evaluate the framework, we conduct extensive experiments on six tabular datasets of varying sizes, domains, and feature types. We compare TabularMDLM against recent baselines, including TabDDPM, CTGAN, and CTGAN+, under a privacy-conscious setting where only synthetic data is used for training and real data for testing. We also assess performance under class imbalance to validate generalization. Results show that TabularMDLM achieves consistently strong classification performance across accuracy, precision, recall, and F1 score, outperforming baselines in both balanced and imbalanced settings. In contrast to existing methods, TabularMDLM scales to diverse data types and low-resource regimes, offering practical advantages in privacy-sensitive applications.}, archive = {J_NEUCOM}, author = {Tokimasa Isomura and Ryotaro Shimizu and Masayuki Goto}, doi = {10.1016/j.neucom.2025.131371}, journal = {Neurocomputing}, month = {11}, pages = {131371}, shortjournal = {Neurocomputing}, title = {Generating realistic synthetic tabular data with integrated LLM and diffusion models}, volume = {655}, year = {2025}, }
• (2025). Adaptive SpikeDeep-classifier: Self-organizing and self-supervised machine learning algorithm for online spike sorting. NEUCOM, 655, 131370. (www)

  @article{ , abstract = {Objective . Invasive brain-computer interface (BCI) research is progressing towards the realization of the motor skills rehabilitation for severely disabled patients in the real world. The size of invasive micro-electrode arrays and the selection of an efficient online spike sorting algorithm (performing spike sorting at runtime) are two key factors that play pivotal roles in the successful decoding of the user’s intentions. The process of spike sorting includes the selection of channels that record the spike activity (SA) and determines the SA of different sources (neurons), on selected channels individually. The neural data recorded with dense micro-electrode arrays is time-varying and often contaminated with non-stationary noise. Unfortunately, current state-of-the-art spike sorting algorithms are incapable of handling the massively increasing amount of time-varying data resulting from the dense micro-electrode arrays, which makes the spike sorting one of the fragile components of the online BCI decoding framework. Approach . This study proposed an adaptive and self-organized algorithm for online spike sorting, named Adaptive SpikeDeep-Classifier (Ada-SpikeDeepClassifier). Our algorithm uses SpikeDeeptector for the channel selection, an adaptive background activity rejector (Ada-BAR) for discarding the background events, and an adaptive spike deep-classifier (Ada-SpikeDeepClassifier) for classifying the SA of different neural units. The process of spike sorting is accomplished by concatenating SpikeDeeptector, Ada-BAR and Ada-SpikeDeepclassifier. Results . The proposed algorithm is evaluated on two different categories of data: a human dataset recorded in our lab, and a publicly available simulated dataset to avoid subjective biases and labeling errors. The proposed Ada-SpikeDeepClassifier outperformed our previously published SpikeDeep-Classifier and eight conventional spike sorting algorithms and produced comparable results to state of the art deep learning based algorithms. Significance . To the best of our knowledge, the proposed algorithm is the first spike sorting algorithm that autonomously adapts to shifts in the distribution of noise and SA data and performs spike sorting without human intervention in various kinds of experimental settings. In addition, the proposed algorithm builds upon artificial neural networks, which makes it an ideal candidate for being embedded on neuromorphic chips that are also suitable for wearable invasive BCI.}, archive = {J_NEUCOM}, author = {Muhammad Saif-ur-Rehman and Omair Ali and Christian Klaes and Ioannis Iossifidis}, doi = {10.1016/j.neucom.2025.131370}, journal = {Neurocomputing}, month = {11}, pages = {131370}, shortjournal = {Neurocomputing}, title = {Adaptive SpikeDeep-classifier: Self-organizing and self-supervised machine learning algorithm for online spike sorting}, volume = {655}, year = {2025}, }
• (2025). FANCL: Feature-guided attention network with curriculum learning for brain metastases segmentation. NEUCOM, 655, 131369. (www)

  @article{ , abstract = {Accurate segmentation of brain metastases (BMs) in MR images is crucial for clinical diagnosis and treatment. However, the task remains highly challenging due to the inherent complexity of BMs and the technical limitations of existing methods. To address common issues including the loss of information from small-sized BMs due to convolutional and pooling operations, the difficulty of segmentation caused by complex metastasis morphology, and the lack of effective learning strategies, the study proposes a novel feature-guided attention network with curriculum learning (FANCL) to achieve precise segmentation of metastases. Specifically, the feature-guided attention mechanism constructs a feature-guided attention matrix by calculating the cross-correlation between the input image and intermediate feature maps, thereby establishing intrinsic connections between BMs of different sizes. This effectively utilizes the rich information from larger tumors to compensate for the loss of high-level features in smaller tumors. To enhance the model’s ability to handle the complex structures of BMs, a voxel-level curriculum learning strategy is proposed, which organizes the voxels in the training images according to their segmentation difficulty, guiding the model to gradually learn the detailed features of BMs from easy to difficult. Simultaneously, by using baseline models of varying depths as the curriculum-mining networks, FANCL dynamically controls the curriculum learning process, thereby further enhancing the model’s adaptability to the heterogeneity of BMs. Comprehensive evaluations conducted on the publicly available BraTS-METS 2023 and BraTS-METS 2024 datasets indicate that FANCL outperforms current state-of-the-art methods across multiple metrics, significantly aiding clinical practice.}, archive = {J_NEUCOM}, author = {Zijiang Liu and Xiaoyu Liu and Linhao Qu and Yonghong Shi}, doi = {10.1016/j.neucom.2025.131369}, journal = {Neurocomputing}, month = {11}, pages = {131369}, shortjournal = {Neurocomputing}, title = {FANCL: Feature-guided attention network with curriculum learning for brain metastases segmentation}, volume = {655}, year = {2025}, }
• (2025). ADGAT: Anomaly detection-based graph adversarial defense framework. NEUCOM, 655, 131368. (www)

  @article{ , abstract = {Graph neural networks (GNNs) are powerful tools for graph representation learning and have achieved excellent performance on tasks such as node classification and link prediction. However, GNNs are vulnerable to adversarial attacks, which inject imperceptible perturbations into the graph and fool GNNs into making wrong predictions. Therefore, we propose the ADGAT framework, which integrates graph anomaly detection and graph purification through the dynamic attention mechanism. The anomaly detector in ADGAT leverages deep neural networks to decode latent edge values in the pre-processing stage. The attention is applied in GNN for message-passing with an extra purification layer. The attention coefficients are dynamically configured for different edge clusters based on decoded edge values. ADGAT defends against attacks through purification while preserving clean graph information as much as possible. Extensive experiments on real graph datasets demonstrate that ADGAT has certain adaptability and can achieve better performance. For instance, on the Cora dataset under Metattack, its defensive performance surpasses traditional baselines by approximately 8 % and outperforms advanced robust baselines by about 1.5 %.}, archive = {J_NEUCOM}, author = {Youhuizi Li and Yi Wang and Yuyu Yin and Tingting Liang}, doi = {10.1016/j.neucom.2025.131368}, journal = {Neurocomputing}, month = {11}, pages = {131368}, shortjournal = {Neurocomputing}, title = {ADGAT: Anomaly detection-based graph adversarial defense framework}, volume = {655}, year = {2025}, }
• (2025). Position encoding of global attention for weakly supervised entity alignment. NEUCOM, 655, 131365. (www)

  @article{ , abstract = {Entity alignment aims to identify and integrate equivalent entities across different Knowledge Graphs (KGs). Graph neural network methods perform messages passing through the graph structure to obtain the similarities of entities and achieve considerable performance. However, it is difficult for distant entity pairs to learn supervised information from the alignment seed, resulting in a sharp decline in performance in few-shot tasks. To address this issue, we propose a novel weakly supervised entity alignment framework— Position Encoding of Global Attention for Weakly Supervised Entity Alignment (PEGAE) in this manuscript. Specifically, the positional features of entities from different knowledge graphs are extracted into a unified embedding space. Which makes distant entities absorb supervision information from alignment seeds through position information rather than path information. In addition, a global attention mechanism is employed to assess the significance of supervision information conveyed by different alignment seeds, emphasizing the learning of supervision information from alignment seeds that are more relevant to the current entity. Extensive experiments on public datasets demonstrate the proposed method outperforms the state-of-the-art models.}, archive = {J_NEUCOM}, author = {Tingting Jiang and Shunxin Hu and Shuai Yang and Wentao Ma and Jin He and Qingyong Wang and Chao Wang and Lichuan Gu}, doi = {10.1016/j.neucom.2025.131365}, journal = {Neurocomputing}, month = {11}, pages = {131365}, shortjournal = {Neurocomputing}, title = {Position encoding of global attention for weakly supervised entity alignment}, volume = {655}, year = {2025}, }
• (2025). Scalable and reliable deep transfer learning for intelligent fault detection via multi-scale neural processes embedded with prior knowledge. NEUCOM, 655, 131364. (www)

  @article{ , abstract = {Deep Transfer Learning (DTL) is used to mitigate the degradation of method performance that arises from the discrepancies in data distribution between different domains. Considering the fact that fault data collection in the field of Intelligent Fault Detection (IFD) is challenging and certain faults are scarce, DTL-based methods face the limitation of available observable data. Furthermore, DTL-based methods lack comprehensive uncertainty analysis which is essential for building reliable IFD systems. To address the aforementioned problems, this paper proposes a scalable and reliable DTL-based method known as Neural Processes-based Deep Transfer Learning with Graph Convolution Network (GTNP). The graph convolution network embedded with knowledge, the joint modeling based on global and local latent variables and sparse sampling strategy are used to reduce the demand for observable data in the target domain. The multi-scale uncertainty analysis is obtained by using the distribution characteristics of global and local latent variables, which enhances the reliability of the model’s detection results. The validation of the proposed method is conducted across 3 IFD tasks, consistently showing the superior detection performance of GTNP compared to the other advanced DTL-based methods.}, archive = {J_NEUCOM}, author = {Zhongzhi Li and Jingqi Tu and Jiacheng Zhu and Rong Fan and Jianliang Ai and Yiqun Dong}, doi = {10.1016/j.neucom.2025.131364}, journal = {Neurocomputing}, month = {11}, pages = {131364}, shortjournal = {Neurocomputing}, title = {Scalable and reliable deep transfer learning for intelligent fault detection via multi-scale neural processes embedded with prior knowledge}, volume = {655}, year = {2025}, }
• (2025). Bi-directional generative retrieval-augmented diffusion models for document-level informative argument extraction. NEUCOM, 655, 131360. (www)

  @article{ , abstract = {Document-level Informative Argument Extraction (IAE) presents a significant challenge in the field of information extraction. This challenge stems from the necessity for implicit coreference reasoning and the linking of long-range dependencies between events within a document. Despite recent efforts to leverage generation-based document-level extraction to enhance cross-sentence inference capabilities and capture more interactions between different events, these methods often fall short in their generation quality due to difficulties in understanding the global context. Motivated by these observations and the high-quality generation results of recent diffusion models, we propose an effective model known as BGRD ( B i-directional G enerative R etrieval-augmented D iffusion models) for document-level IAE. In BGRD, a text diffusion model is designed to generate high-quality target event sequences that mutually benefit the retrieval stage, leveraging previously generated events as a retrieval source. Firstly, a bi-directional retrieval mechanism is investigated to refine the denoising process, effectively exploring the knowledge from retrieved samples. This enhances the text diffusion model’s ability to capture the global context interconnecting the events. Secondly, retrieval-augmented cross-attention is employed between the retrieved samples and the target event sequences (random Gaussian noise during the inference phase) within the text diffusion model. Through this interaction, the quality of the retrieval source is improved by generating highly informative event sequences, which benefits the bi-directional retrieval stage. Extensive experiments on the publicly available argument extraction datasets demonstrate the superiority of our proposed BGRD model over existing approaches.}, archive = {J_NEUCOM}, author = {Lei Luo and Xuanzhi Chen and Liming Mao and Xinjie Yang and Yajing Xu and Jun Guo}, doi = {10.1016/j.neucom.2025.131360}, journal = {Neurocomputing}, month = {11}, pages = {131360}, shortjournal = {Neurocomputing}, title = {Bi-directional generative retrieval-augmented diffusion models for document-level informative argument extraction}, volume = {655}, year = {2025}, }
• (2025). DeepEDD: Nonparametric deep graph clustering through effective distance and density. NEUCOM, 655, 131359. (www)

  @article{ , abstract = {Deep graph clustering methods have become a prominent area of research. Although nonparametric clustering techniques have clear advantages, current research indicates that nonparametric methods in the field of graph clustering are scarce. Most deep graph clustering methods are parametric and require the number of clusters k to be specified a priori. An unknown k significantly diminishes the clustering performance of the model, and alternative criteria for judgment come with a computational cost. This issue is particularly pronounced in the context of deep learning, due to the increase in the number of training sessions. To address this issue, we propose an end-to-end framework, namely, nonparametric Deep graph clustering via Effective Distance and Density (DeepEDD), which utilizes the data from the clustering process to predict the k without additional computational cost. Specifically, DeepEDD introduces a new clustering algorithm based on effective distance and density as a second decoder on top of the graph masking-based fusion autoencoder framework. The algorithm does not need to define k in advance because it can automatically determine k by polynomial curve fitting. We also design a novel fusion loss to optimize DeepEDD. Experimental results on six publicly available datasets demonstrate that our proposed method surpasses existing nonparametric methods, both classical and deep. Moreover, the mask-based approach enhances the generalization performance of our framework.}, archive = {J_NEUCOM}, author = {Yuanchi Ma and Hui He and Zhongxiang Lei and Kaize Shi and Xueping Peng and Zhendong Niu}, doi = {10.1016/j.neucom.2025.131359}, journal = {Neurocomputing}, month = {11}, pages = {131359}, shortjournal = {Neurocomputing}, title = {DeepEDD: Nonparametric deep graph clustering through effective distance and density}, volume = {655}, year = {2025}, }
• (2025). Degradation-adaptive attack-robust self-supervised facial representation learning. NEUCOM, 655, 131356. (www)

  @article{ , abstract = {Self-supervised face representation learning (SFRL) shows strong potential for scalable face-related applications, yet remains vulnerable to adversarial attacks that cause dual facial semantic degradations, namely (1) structured distortions in key facial regions ( e.g. , subtle inter-ocular distance shifts) that disrupt identity-related features, and (2) unstructured additive noise ( e.g. , illumination artifacts) that entangles with face-related features in latent space. Existing defense methods struggle to deal with both facial semantic degradations in SFRL, resulting in limited robustness. To address this, inspired by existing reverse Diffusion approaches that effectively tackle the image denoising, we propose DAR-SFRL , a novel Degradation-adaptive Attack-Robust Self-supervised Face Representation Learning framework. DAR-SFRL models adversarial attacks as a degradation-based function composed of geometric distortions and additive noise, applying a multi-stage reverse Diffusion iterative process to recover facial semantics. At each stage of the process, DAR-SFRL employs: (1) an adaptive degraded-face restoration method that progressively reverses the degradation function and recovers fine-grained details from structured distortions, and (2) a noise-orthogonal contrastive learning mechanism to mitigate the impact of unstructured additive noise by maximizing the dissimilarity between noisy and clean image features in the latent space. Extensive experiments across tasks—including face recognition, facial expression recognition, and facial action unit detection—demonstrate that DAR-SFRL significantly outperforms state-of-the-art defenses under various adversarial attacks, highlighting its robustness and generalization in real-world face-aware applications. Our evaluation code is available at https://github.com/23wk/DAR-SFRL}, archive = {J_NEUCOM}, author = {Ke Wang and Yuanyuan Liu and Chang Tang and Kun Sun and Yibing Zhan and Zhe Chen}, doi = {10.1016/j.neucom.2025.131356}, journal = {Neurocomputing}, month = {11}, pages = {131356}, shortjournal = {Neurocomputing}, title = {Degradation-adaptive attack-robust self-supervised facial representation learning}, volume = {655}, year = {2025}, }
• (2025). Multi-source adversarial domain adaptation with modulated adaptive weights. NEUCOM, 655, 131352. (www)

  @article{ , abstract = {Domain drift, characterized by the distributional discrepancy between labeled source domain training data and unlabeled target domain testing data, poses a significant challenge for deep neural networks in cross-domain tasks. This challenge is exacerbated when handling data from multiple source domains. Multi-source unsupervised domain adaptation (MUDA) aims to address this issue by leveraging information from multiple sources to enhance model performance on the target domain. Despite notable advancements, existing MUDA methods still exhibit several limitations. Currently, feature discriminability receives an excessive penalty when using adversarial learning to learn transferable features with cross-domain knowledge in multi-source domain adaptation. To address this, we introduce batch spectral penalization (BSP), which enhances the relative strength of other feature vectors, thereby improving feature discriminability. Moreover, most of the existing methods assume uniform contributions from all source domains, neglecting the inherent differences among them. Accordingly, we consider quantifying the discrepancies between each source domain and the target domain. Ultimately, we introduce multi-source adversarial domain adaptation with modulated adaptive weights (MADA-AW). This framework employs an adversarial learning strategy with an adaptive weight mechanism to dynamically assess and optimize the contribution of each source domain. Experimental results on multiple image classification benchmarks validate the effectiveness of MADA-AW.}, archive = {J_NEUCOM}, author = {Wenjun Liu and Na Chen and Jiangtao Peng and Shuoshuo Hui}, doi = {10.1016/j.neucom.2025.131352}, journal = {Neurocomputing}, month = {11}, pages = {131352}, shortjournal = {Neurocomputing}, title = {Multi-source adversarial domain adaptation with modulated adaptive weights}, volume = {655}, year = {2025}, }
• (2025). SAGA-feat: A semantic- and geometry-aware network for sparse local feature learning. NEUCOM, 655, 131349. (www)

  @article{ , abstract = {Sparse local feature learning plays a pivotal role in computer vision, underpinning applications such as simultaneous localization and mapping (SLAM), image matching, and 3D reconstruction. However, existing approaches often suffer from limited multi-scale contextual modeling, inadequate integration of spatial and semantic information, and significant detail loss during feature reconstruction. To address these challenges, we propose SAGA-Feat (Semantic- and Geometry-Aware Feature Extraction Network), a weakly supervised framework following a decoupled describe-then-detect paradigm. SAGA-Feat introduces a Semantic-and-Geometry-Aware Attention module (SAGA-Attn) to enhance directional sensitivity and local semantic aggregation, a Structure-Aware Fusion decoder (SAGA-Fuse) to enforce spatial–semantic consistency through dual-domain normalization, and a Semantic-Guided Adaptive Sampling module (SAGA-Sample) to refine feature reconstruction based on learned saliency cues. Extensive experiments across multiple benchmarks demonstrate that SAGA-Feat delivers state-of-the-art or competitive performance among learning-based methods, highlighting its overall robustness and effectiveness in complex visual scenarios.}, archive = {J_NEUCOM}, author = {Yanhan Mo and Mengxiao Yin and Guiqing Li and Junjie Liao and Zhijie Liang}, doi = {10.1016/j.neucom.2025.131349}, journal = {Neurocomputing}, month = {11}, pages = {131349}, shortjournal = {Neurocomputing}, title = {SAGA-feat: A semantic- and geometry-aware network for sparse local feature learning}, volume = {655}, year = {2025}, }
• (2025). Event-triggered optimized control for multiagent systems with input saturation from reinforcement learning viewpoint. NEUCOM, 655, 131348. (www)

  @article{ , abstract = {This paper proposes the adaptive optimized consensus control of nonstrict-feedback discrete-time switched multiagent systems through incorporating the reinforcement learning strategy with event-triggered mechanisms. A simplified smooth approximation function is developed to address asymmetric input saturation constraints, effectively alleviating saturation problems while preserving closed-loop stability. Under the actor-critic framework, an adaptive controller is developed through the integration of the reinforcement learning backstepping algorithm and the approximation capability of neural networks. An external function term resolves the nonstrict-feedback structural challenge, complemented by a minimal learning parameter technique that reduces computational complexity. Furthermore, to enhance control performance and resource utilization efficiency, both periodic and input-quantized event-triggered mechanisms are considered. Lyapunov stability analysis demonstrates that all closed-loop signals are guaranteed to be semi-globally uniformly ultimately bounded under appropriate control parameters. Finally, numerical simulations validate the theoretical feasibility and confirm the control scheme’s efficacy in coordination tasks of multiagent systems.}, archive = {J_NEUCOM}, author = {Lihua Tan and Le You and Junjian Huang and Xin Wang}, doi = {10.1016/j.neucom.2025.131348}, journal = {Neurocomputing}, month = {11}, pages = {131348}, shortjournal = {Neurocomputing}, title = {Event-triggered optimized control for multiagent systems with input saturation from reinforcement learning viewpoint}, volume = {655}, year = {2025}, }
• (2025). SMANet: Sequence-enhanced multi-head attention network for robust neural semantic learning in noisy computational environments. NEUCOM, 655, 131347. (www)

  @article{ , abstract = {Traditional communication systems often fail to efficiently transmit meaningful information in noisy and dynamic environments, prompting the adoption of neural network architectures in semantic communication to prioritize semantic content over raw data. Existing neural models face persistent challenges in mitigating high noise interference, capturing long-range dependencies in sequences, and preserving semantic fidelity under varying conditions. This paper proposes SMANet, sequence-enhanced multi-head attention network for robust neural semantic learning in noisy computational environments. SMANet integrates multi-head attention mechanisms with a Dilated Normalization Block (DNB)—a specialized neural module for extracting local temporal features and global semantic representations—to enhance sequence processing capabilities, alleviate gradient vanishing/explosion issues during training, and improve network stability. At the transmitter, a neural semantic encoder employs dilated convolutions and normalization for robust feature extraction, paired with a channel encoder to achieve noise resilience; at the receiver, neural decoders precisely reconstruct semantics, facilitating applications in machine learning-driven cognitive systems. Experimental evaluations on AWGN and Rayleigh fading channels demonstrate SMANet’s superior performance, outperforming DeepSC by 23 % in BLEU scores, achieving a sentence similarity of 0.91 at SNR=18 dB, and maintaining 85 % semantic fidelity at SNR < 6 dB, highlighting its potential for neurocomputing in resource-constrained networks.}, archive = {J_NEUCOM}, author = {Jia Guo and Xinyu Jia and Jinqi Zhu and Xiang Li and Yang Liu and Weijia Feng and Wanli Xue}, doi = {10.1016/j.neucom.2025.131347}, journal = {Neurocomputing}, month = {11}, pages = {131347}, shortjournal = {Neurocomputing}, title = {SMANet: Sequence-enhanced multi-head attention network for robust neural semantic learning in noisy computational environments}, volume = {655}, year = {2025}, }
• (2025). Neural cyberattacks applied to the vision under realistic visual stimuli. NEUCOM, 655, 131344. (www)

  @article{ , abstract = {Brain-Computer Interfaces (BCIs) are systems traditionally used in medicine and designed to interact with the brain to record or stimulate neurons. Despite their benefits, the literature has demonstrated that invasive BCIs focused on neurostimulation present vulnerabilities allowing attackers to gain control. In this context, neural cyberattacks emerged as threats able to disrupt spontaneous neural activity by performing neural overstimulation or inhibition. Previous work validated these attacks in small-scale simulations with a reduced number of neurons, lacking real-world complexity. Thus, this work tackles this limitation by analyzing the impact of two existing neural attacks, Neuronal Flooding (FLO) and Neuronal Jamming (JAM), on a complex neuronal topology of the primary visual cortex of mice consisting of approximately 230,000 neurons, tested on three realistic visual stimuli: flash effect, movie, and drifting gratings. Each attack was evaluated over three relevant events per stimulus, also testing the impact of attacking 25 % and 50 % of the neurons. The results, based on the number of spikes and shift percentage metrics, showed that the attacks caused the greatest impact on the movie, while dark and static events exhibited highest resilience. Although both attacks can significantly affect neural activity, JAM was generally more damaging, producing longer temporal delays, and had a larger prevalence. Finally, JAM did not require altering many neurons to significantly affect neural activity, while the impact of FLO increased with the number of neurons attacked.}, archive = {J_NEUCOM}, author = {Victoria Magdalena López Madejska and Sergio López Bernal and Gregorio Martínez Pérez and Alberto Huertas Celdrán}, doi = {10.1016/j.neucom.2025.131344}, journal = {Neurocomputing}, month = {11}, pages = {131344}, shortjournal = {Neurocomputing}, title = {Neural cyberattacks applied to the vision under realistic visual stimuli}, volume = {655}, year = {2025}, }
• (2025). Adaptive teaching with shared classifier for knowledge distillation. NEUCOM, 655, 131343. (www)

  @article{ , abstract = {Knowledge distillation (KD) is a technique used to transfer knowledge from an overparameterized teacher network to a less-parameterized student network, thereby minimizing the incurred performance loss. KD methods can be categorized into offline and online approaches. Offline KD leverages a powerful pretrained teacher network, while online KD allows the teacher network to be adjusted dynamically to enhance the learning effectiveness of the student network. Recently, it has been discovered that sharing the classifier of the teacher network can significantly boost the performance of the student network with only a minimal increase in the number of network parameters. Building on these insights, we propose adaptive teaching with a shared classifier (ATSC). In ATSC, the pretrained teacher network self-adjusts to better align with the learning needs of the student network based on its capabilities, and the student network benefits from the shared classifier, enhancing its performance. Additionally, we extend ATSC to environments with multiple teachers. We conduct extensive experiments, demonstrating the effectiveness of the proposed KD method. Our approach achieves state-of-the-art results on the CIFAR-100 and ImageNet datasets in both single-teacher and multi-teacher scenarios. The source code is publicly available at https://github.com/random2314235/ATSC .}, archive = {J_NEUCOM}, author = {Jaeyeon Jang and Young-Ik Kim and Hyeonseong Lee}, doi = {10.1016/j.neucom.2025.131343}, journal = {Neurocomputing}, month = {11}, pages = {131343}, shortjournal = {Neurocomputing}, title = {Adaptive teaching with shared classifier for knowledge distillation}, volume = {655}, year = {2025}, }
• (2025). Factorizing value function with hierarchical residual Q-network in multi-agent reinforcement learning. NEUCOM, 655, 131340. (www)

  @article{ , abstract = {Value function decomposition has achieved notable success in Multi-Agent Reinforcement Learning (MARL) under the centralized training with decentralized execution paradigm. Traditional value function decomposition methods typically employ monotonic mixing networks to decompose the optimal joint action-value function in order to ensure consistency between joint and local action selections. However, these networks often face limitations in representational capacity and sample efficiency, making it difficult to accurately fit the reward function and achieve stable convergence, thus leading to suboptimal results. To address these challenges, we propose a novel MARL framework called Hierarchical Residual Q-network (HRQ). The HRQ framework adheres to the Individual-Global-Max principle while applying more relaxed constraints. It features an Outer Residual Network (ORN) that adjusts the joint action-value function to enhance the representational capacity of the mixing network. Additionally, HRQ incorporates an Inner Residual Entropy Auxiliary Network (IREAN) to refine individual action-value functions, addressing credit assignment and value overestimation problems arising from task diversity and agent independence in MARL. Our approach enhances exploration efficiency, sample efficiency, and convergence stability. Extensive experiments on multi-agent cooperative benchmarks, including predator-prey and StarCraft, demonstrate that HRQ outperforms existing methods in convergence speed, stability, and adaptability. Compared with the best comparison method, HRQ achieves an overall performance improvement of 10 %–20 %.}, archive = {J_NEUCOM}, author = {Fang Gao and Yunxiang Cai and Haotian Yao and Shaodong Li and Qing Gao and Linfei Yin}, doi = {10.1016/j.neucom.2025.131340}, journal = {Neurocomputing}, month = {11}, pages = {131340}, shortjournal = {Neurocomputing}, title = {Factorizing value function with hierarchical residual Q-network in multi-agent reinforcement learning}, volume = {655}, year = {2025}, }
• (2025). Learning multi-granularity temporal characteristics for attention based knowledge tracing. NEUCOM, 655, 131338. (www)

  @article{ , abstract = {Knowledge tracing (KT) plays a crucial role in personalized learning by predicting students’ future performance based on their historical interaction data. Despite advancements in KT models, they often struggle to effectively model the intricate temporal dynamics inherent in student learning processes, particularly in capturing multi-granular temporal information, accommodating irregular time intervals, and handling varying sequence lengths within student interaction data. To address these challenges, we propose MTKT to model multi-scale time patterns and long-term behaviors in student interaction data. Specifically, MTKT introduces three key components. First, a multi-aspect embedding layer captures both temporal information and interaction behaviors from student learning data. Second, a dual attention module with linear decaying bias applies negative bias to attention scores, effectively modeling long-term forgetting in students’ knowledge states across varying sequence lengths. Finally, a causal interaction convolution module captures fine-grained representations in both short-term and long-term interactions. We evaluate MTKT on three public educational datasets and compare it with state-of-the-art KT models. Our experiments show that MTKT consistently outperforms existing models in student performance prediction and length-adaptive prediction. To facilitate reproducibility and future research, we make our code and processed datasets publicly available at https://pykt.org .}, archive = {J_NEUCOM}, author = {Youheng Bai and Xueyi Li and Zitao Liu and Yaying Huang and Teng Guo and Mingliang Hou and Weiqi Luo}, doi = {10.1016/j.neucom.2025.131338}, journal = {Neurocomputing}, month = {11}, pages = {131338}, shortjournal = {Neurocomputing}, title = {Learning multi-granularity temporal characteristics for attention based knowledge tracing}, volume = {655}, year = {2025}, }
• (2025). A comprehensive benchmark of machine and deep learning models on structured data for regression and classification. NEUCOM, 655, 131337. (www)

  @article{ , abstract = {The analysis of tabular datasets is highly prevalent both in scientific research and real-world applications of Machine Learning (ML). Unlike many other ML tasks, Deep Learning (DL) models often do not outperform traditional methods in this area. Previous comparative benchmarks have shown that DL performance is frequently equivalent to or even inferior to models such as Gradient Boosting Machines (GBMs). In this study, we introduce a comprehensive benchmark aimed at better characterizing the types of datasets where DL models excel. Although several important benchmarks for tabular datasets already exist, our contribution lies in the variety and depth of our comparison: we evaluate 111 datasets with 20 different models, including both regression and classification tasks. These datasets vary in scale and include both those with and without categorical variables. Importantly, our benchmark contains a sufficient number of datasets where DL models perform best, allowing for a thorough analysis of the conditions under which DL models excel. Building on the results of this benchmark, we train a model that predicts scenarios where DL models significantly outperform alternative methods, considering only datasets where the performance difference between the two groups is statistically significant. This filtering yields 36 datasets out of the original 111. On this subset, our model achieves 92 % accuracy. We present insights derived from this characterization and compare these findings to previous benchmarks.}, archive = {J_NEUCOM}, author = {Assaf Shmuel and Oren Glickman and Teddy Lazebnik}, doi = {10.1016/j.neucom.2025.131337}, journal = {Neurocomputing}, month = {11}, pages = {131337}, shortjournal = {Neurocomputing}, title = {A comprehensive benchmark of machine and deep learning models on structured data for regression and classification}, volume = {655}, year = {2025}, }
• (2025). The multiplex deep latent position model for the clustering of nodes in multiview networks. NEUCOM, 655, 131336. (www)

  @article{ , abstract = {Multiplex networks capture multiple types of interactions among the same set of nodes, creating a complex, multi-relational framework. A typical example is a social network where nodes (actors) are connected by various types of ties, such as professional, familial, or social relationships. Clustering nodes in these networks is a key challenge in unsupervised learning, given the increasing prevalence of multiview data across domains. While previous research has focused on extending statistical models to handle such networks, these adaptations often struggle to fully capture complex network structures and rely on computationally intensive Markov chain Monte Carlo (MCMC) for inference, rendering them less feasible for effective network analysis. To overcome these limitations, we propose the multiplex deep latent position model (MDLPM), which generalizes and extends latent position models to multiplex networks. MDLPM combines deep learning with variational inference to effectively tackle both the modeling and computational challenges raised by multiplex networks. Unlike most existing deep learning models for graphs that require external clustering algorithms (e.g., k-means) to group nodes based on their latent embeddings, MDLPM integrates clustering directly into the learning process, enabling a fully unsupervised, end-to-end approach. This integration improves the ability to uncover and interpret clusters in multiplex networks without relying on external procedures. Numerical experiments across various synthetic data sets and two real-world networks demonstrate the performance of MDLPM compared to state-of-the-art methods, highlighting its applicability and effectiveness for multiplex network analysis.}, archive = {J_NEUCOM}, author = {Dingge Liang and Marco Corneli and Charles Bouveyron and Pierre Latouche and Junping Yin}, doi = {10.1016/j.neucom.2025.131336}, journal = {Neurocomputing}, month = {11}, pages = {131336}, shortjournal = {Neurocomputing}, title = {The multiplex deep latent position model for the clustering of nodes in multiview networks}, volume = {655}, year = {2025}, }
• (2025). Bidirectional distance encoding for graph neural networks. NEUCOM, 655, 131335. (www)

  @article{ , abstract = {Graph neural networks (GNNs) have become the mainstream tools for processing graph data, despite being limited by their weak expressive power. In recent years, attempts have been made to incorporate spatial distance encoding into the network learning process which has been theoretically proven to effectively improve the expressive power of GNNs. However, most of the previous work is based on undirected graphs, thus how to effectively utilize the distance of directed graphs remains a problem. In this paper, we first introduce a novel directed bidirectional distance embedding (Dir-BD) for directed graphs, which significantly boosts the expressive power of GNNs. Second, we develop a universal framework that integrates Dir-BD into any message passing neural network (MPNN) to better handle directed graphs. In addition, we combine several mainstream GNNs as the backbone of the DirBD-MPNN framework and conduct node classification and link prediction experiments on real-world datasets, achieving significant improvements with the addition of DirBD-MPNN and obtaining state-of-the-art results.}, archive = {J_NEUCOM}, author = {Xi Wang and Changxiang He and Luchao Zhang and Xiaofei Qin and Xuedian Zhang and Ming Li and Min Ju}, doi = {10.1016/j.neucom.2025.131335}, journal = {Neurocomputing}, month = {11}, pages = {131335}, shortjournal = {Neurocomputing}, title = {Bidirectional distance encoding for graph neural networks}, volume = {655}, year = {2025}, }
• (2025). SEPT: Squeeze-and-excitation point transformer with wise composite information encoding for segmentation and classification. NEUCOM, 655, 131334. (www)

  @article{ , abstract = {Utilizing the rich information of point clouds effectively is critical for segmentation and classification. However as the network goes deeper, the processes of downsampling, convolution, and information aggregation in the network cause the loss of detailed information. This is particularly damaging to segmentation and classification. Mainstream methods typically compensate for geometric loss only through position encoding. Different from them, we propose Wise Composite Information Encoding (WCIE) to fully mine a variety of detailed information to supplement information loss. And to fully utilize the rich information after being supplemented, one must concentrate on the key information. Therefore, we propose Squeeze-and-Excitation Attention (SEA) specialized for point clouds, which uses key information to emphasize useful information. Specifically, we apply it to WCIE and feature aggregation to enhance key information. During feature aggregation, it enhances key semantic and spatial information by operating on channels and points respectively. And it is worth noting that SEA can pay more attention to key information with lower computing resources compared to other attention methods. In addition, to better capture global features to learn contextual features, this paper constructs Global-Local Feature Relationships (GLFR) and adaptively fuses them into local features. Finally, this paper proposes the Squeeze-and-Excitation Point Transformer (SEPT) network for segmentation and classification. Compared with current large model methods with huge computing resource consumption, SEPT is lightweight and easy to train, and it performs well in tasks such as scene segmentation (S3DIS), shape classification (ScanObjectNN, ModelNet40), and part segmentation (ShapeNetPart). Particularly, we achieve the competitive performance of 73.3 % (Area 5) and 79.2 % (6-fold cross-validation) mIoU on the large-scale S3DIS benchmark.}, archive = {J_NEUCOM}, author = {Chao Ji and Guofeng Tong and Hao Peng and Guocheng Wang}, doi = {10.1016/j.neucom.2025.131334}, journal = {Neurocomputing}, month = {11}, pages = {131334}, shortjournal = {Neurocomputing}, title = {SEPT: Squeeze-and-excitation point transformer with wise composite information encoding for segmentation and classification}, volume = {655}, year = {2025}, }
• (2025). Certainty-aware support vector machines for fair classification. NEUCOM, 655, 131330. (www)

  @article{ , abstract = {Support Vector Machine (SVM) is a widely adopted classification algorithm. Many fair machine learning frameworks achieve fairer classification based on SVMs, mitigating systematic bias due to historical discrimination in the training data. However, data imbalance is often inherent in sensitive subgroups, which may cause the learned decision boundary to become skewed toward the majority group, further exacerbating the disadvantages faced by minority groups. To address these fairness concerns, we propose a novel C ertainty F actor-based S upport V ector M achine (CFSVM) for fair decision-making. CFSVM is an adaptive debiasing framework that incorporates instance-level penalization guided by Certainty Factors (CFs). It dynamically adjusts the influence of individual training samples by estimating local confidence using K-Nearest Neighbors (KNN) and CFs. These scores quantify each sample’s representativeness with respect to both class labels and subgroup distributions, and are used to modulate their contribution to the decision boundary during training. By tailoring the influence of each sample, our proposed CFSVM method mitigates the effects of over-clustering among dominant subgroups and addresses local distributional disparities, especially in scenarios involving imbalanced subgroup representations. We evaluate CFSVM on multiple real-world datasets across several distinct fairness frameworks. Experimental results demonstrate that CFSVM improves fairness metrics while maintaining competitive classification performance, providing a practical and effective solution for fair decision-making in machine learning.}, archive = {J_NEUCOM}, author = {Boyan Chen and Yuxuan Hu and Guixian Zhang and Debo Cheng and Shichao Zhang}, doi = {10.1016/j.neucom.2025.131330}, journal = {Neurocomputing}, month = {11}, pages = {131330}, shortjournal = {Neurocomputing}, title = {Certainty-aware support vector machines for fair classification}, volume = {655}, year = {2025}, }
• (2025). Hermite polynomials facilitating on-line learning analysis of layered neural networks with arbitrary activation function. NEUCOM, 655, 131328. (www)

  @article{ , abstract = {Following the standard statistical mechanics methods we analyze the training by online stochastic gradient descent of two-layer neural networks in a student–teacher scenario. We focus on understanding the role that different activations play, in particular mismatches between the student and the teacher, in these learning scenarios. By expanding the activation functions in the Hermite polynomial basis, we are able to effectively approximate the relevant integrals with much less computational effort than naive numerical integration. Moreover, we also extend the framework to study scenarios of concept drift and weight decay also with arbitrary activation functions. All these extensions comprise relevant advances in the field, allowing us to obtain analytical results for more realistic scenarios.}, archive = {J_NEUCOM}, author = {Otavio Citton and Frederieke Richert and Michael Biehl and Michiel Straat}, doi = {10.1016/j.neucom.2025.131328}, journal = {Neurocomputing}, month = {11}, pages = {131328}, shortjournal = {Neurocomputing}, title = {Hermite polynomials facilitating on-line learning analysis of layered neural networks with arbitrary activation function}, volume = {655}, year = {2025}, }
• (2025). Explainable topology-guided matrix factorization for link prediction in open-source software networks. NEUCOM, 655, 131326. (www)

  @article{ , abstract = {The rapid evolution of open-source software (OSS) ecosystems necessitates accurate and interpretable link prediction to facilitate project recommendation and software reuse. Existing approaches, including matrix factorization and graph-based models, often struggle to balance predictive accuracy and semantic transparency. To address this challenge, we propose ELMF, a topology-aware and explainable latent matrix factorization framework for link prediction in OSS networks. Unlike traditional vector embeddings, ELMF represents each project node as a structured latent matrix, capturing multi-aspect semantic roles and supporting fine-grained interpretability. We introduce a topology-guided explainability regularization term that encourages latent proximity among semantically related nodes, which we prove to be equivalent to Laplacian smoothing over an explainability graph. The overall loss function integrates reconstruction fidelity, L2 regularization, semantic alignment, and structure-aware divergence. Extensive experiments conducted on four real-world datasets demonstrate that ELMF consistently outperforms state-of-the-art baselines in AUC and Precision. Ablation studies and visualizations validate the critical role of the interpretability module in improving performance and transparency. Our work provides a unified and efficient solution that bridges structural representation learning with interpretable link prediction in open-source environments.}, archive = {J_NEUCOM}, author = {Liang Chen and Yun Yang}, doi = {10.1016/j.neucom.2025.131326}, journal = {Neurocomputing}, month = {11}, pages = {131326}, shortjournal = {Neurocomputing}, title = {Explainable topology-guided matrix factorization for link prediction in open-source software networks}, volume = {655}, year = {2025}, }
• (2025). VAPCaps: A novel variance-based attention network with imbalance aware loss for better pathology detection in video capsule endoscopy. NEUCOM, 655, 131325. (www)

  @article{ , abstract = {Objective: Video Capsule Endoscopy (VCE) is used as a non-invasive method to see into the gastrointestinal (GI) tract, irrespective of its varying anatomical structure. Despite recent progress in abnormality detection and fine-grained image classification in VCE, the issue of generalizability is a point that needs to be addressed. Method: We propose a V ariance A ttention network for better P athology in video Cap sule endoscopy (VAPCaps) to learn contrastive details among inter-class images through a latent space attention mechanism. The network is designed to classify multiple categories of abnormalities within a single framework. To mitigate the challenges posed by imbalanced datasets, including disparities in data distribution and detection difficulty, a novel loss function called Imbalance-Aware Loss (IAL) has been designed to optimize the training of the proposed VAPCaps model. Results: The effectiveness of the model is evaluated using metrics such as precision, recall, F1 score, and accuracy on the Kvasir Capsule dataset and the CAD-CAP database. The VAPCaps model shows significant improvement over other state-of-the-art techniques with an accuracy of 99.26 % for Kvasir Capsule dataset and 92.99 % for CAD-CAP database. Ablation analyses are performed to show the significance of the proposed IAL loss function in improving the diagnosis accuracy. The efficacy of the proposed scheme is demonstrated using LIME, t_SNE and CAM analyses.}, archive = {J_NEUCOM}, author = {Jithin Joseph and Sudhish N. George and Kiran Raja}, doi = {10.1016/j.neucom.2025.131325}, journal = {Neurocomputing}, month = {11}, pages = {131325}, shortjournal = {Neurocomputing}, title = {VAPCaps: A novel variance-based attention network with imbalance aware loss for better pathology detection in video capsule endoscopy}, volume = {655}, year = {2025}, }
• (2025). Real-time human pose estimation and tracking on monocular videos: A systematic literature review. NEUCOM, 655, 131309. (www)

  @article{ , abstract = {Real-time human pose estimation and tracking on monocular videos is a fundamental task in computer vision with a wide range of applications. Recently, benefiting from deep learning-based methods, it has received impressive progress in performance. Although some works have reviewed and summarised the advancements in this field, few have specifically focused on real-time performance and monocular video-based solutions. The goal of this review is to bridge this gap by providing a comprehensive understanding of real-time monocular video-based human pose estimation and tracking, encompassing both 2D and 3D domains, as well as single-person and multi-person scenarios. To achieve this objective, this paper systematically reviews 68 papers published between 2014 and 2024 to answer six research questions. This review brings new insights into computational efficiency measures and hardware configurations of existing methods. Additionally, this review provides a deep discussion on trade-off strategies for accuracy and efficiency in real-time systems. Finally, this review highlights promising directions for future research and provides practical solutions for real-world applications.}, archive = {J_NEUCOM}, author = {Yingying Chen and Zhenan Feng and Daniel Paes and Daniel Nilsson and Ruggiero Lovreglio}, doi = {10.1016/j.neucom.2025.131309}, journal = {Neurocomputing}, month = {11}, pages = {131309}, shortjournal = {Neurocomputing}, title = {Real-time human pose estimation and tracking on monocular videos: A systematic literature review}, volume = {655}, year = {2025}, }
• (2025). A systematic review of long document summarization methods: Evaluation metrics and approaches. NEUCOM, 655, 131287. (www)

  @article{ , abstract = {The rapid growth of complex textual data in domains such as medicine, law, and science has heightened the relevance of Long Document Summarization (LDS). Effective summarization not only requires advanced techniques but also robust evaluation metrics capable of capturing summary quality, coherence, and factual accuracy. We analyze 113 peer-reviewed studies from last two years, selected through comprehensive searches in SCOPUS, Web of Science, and PubMed, following PRISMA 2020 guidelines. We focus on LDS methods and the metrics used to evaluate them. Results indicate a rising adoption of hybrid models combining extractive and abstractive strategies, frequently powered by deep learning and optimization. Concurrently, evaluation practices have shifted from traditional overlap-based metrics (e.g., ROUGE) toward semantic measures such as BERTScore and MoverScore. However, these metrics still face challenges related to interpretability, domain adaptation, and computational cost. We advocate for the development of holistic, explainable, and reference-free evaluation frameworks aligned with human judgment to enhance the reliability and applicability of LDS systems across domains.}, archive = {J_NEUCOM}, author = {Bady Gana and Héctor Allende-Cid and Stefan Rüping and Marcelo Becerra-Rozas and Juan Zamora}, doi = {10.1016/j.neucom.2025.131287}, journal = {Neurocomputing}, month = {11}, pages = {131287}, shortjournal = {Neurocomputing}, title = {A systematic review of long document summarization methods: Evaluation metrics and approaches}, volume = {655}, year = {2025}, }
• (2025). XMF-GNN: A cross-modality dynamic fusion heterogeneous graph neural network for network intrusion detection. NEUCOM, 655, 131285. (www)

  @article{ , abstract = {With the rapid development of the Internet, network attacks have become increasingly complex and covert. However, most existing intrusion detection methods focus on extracting static features from a single modality, making it difficult to fully exploit the complementary information across different modalities. Furthermore, these methods often lack structure-awareness and adaptive feature fusion capabilities when modeling complex attack behaviors, resulting in limited effectiveness in detecting diverse intrusion patterns. To address these challenges, we propose XMF-GNN, a heterogeneous graph neural network (HGNN) model enhanced with a cross-modality attention fusion mechanism. This model constructs a heterogeneous graph representation based on flow and packet modalities of network traffic, and employs an attention-driven modality fusion mechanism to dynamically learn the importance weights between modalities, enabling adaptive cross-modal feature fusion. The model architecture consists of a two-layer heterogeneous graph convolutional encoder and a modality attention module, which together effectively capture the intricate relationships between traffic flows and payload data in complex network environments. Experiments are conducted on two publicly available network intrusion detection datasets, CIC-IDS2017 and CIC-IoT2023. The results show that the proposed model achieves an F1 score of 0.977 on CIC-IDS2017 and an even higher F1 score of 0.985 on CIC-IoT2023 in multi-class classification tasks. It also significantly outperforms state-of-the-art methods in terms of accuracy and other mainstream evaluation metrics, demonstrating the effectiveness and robustness of the proposed cross-modality fusion mechanism in handling complex intrusion detection scenarios.}, archive = {J_NEUCOM}, author = {Zhengxiang Ma and Yan Liu and Yu Chen and Zhibin Liu and Yinghao Li}, doi = {10.1016/j.neucom.2025.131285}, journal = {Neurocomputing}, month = {11}, pages = {131285}, shortjournal = {Neurocomputing}, title = {XMF-GNN: A cross-modality dynamic fusion heterogeneous graph neural network for network intrusion detection}, volume = {655}, year = {2025}, }
• (2025). RM-BGNN: A weakly informative bayesian graph neural network based on residual mechanism. NEUCOM, 655, 131284. (www)

  @article{ , abstract = {Graph neural networks have shown excellent performance on ideal graph data in recent years, but often face weak information challenges such as feature loss, structural loss, and label loss in reality. To address these challenges, this article proposes the RM-BGNN model. It adopts long-distance message propagation and graph structure enhancement techniques to solve the problems of feature and structure loss, and uses residual connections to maintain the original information and improve stability. Introducing Bayesian linear layers to handle parameter uncertainty and generating pseudo labels and confidence for unlabeled data through prototype alignment loss effectively alleviates the problem of label shortage. Experiments on multiple real datasets have shown that RM-BGNN outperforms existing baseline models significantly in various weak information scenarios, validating its effectiveness and practicality.}, archive = {J_NEUCOM}, author = {Jihao Dong and Zhaowei Liu and Dong Yang and Peng Song and Jinglei Liu and Anzuo Jiang}, doi = {10.1016/j.neucom.2025.131284}, journal = {Neurocomputing}, month = {11}, pages = {131284}, shortjournal = {Neurocomputing}, title = {RM-BGNN: A weakly informative bayesian graph neural network based on residual mechanism}, volume = {655}, year = {2025}, }
• (2025). IRAGKR:Iterative retrieval augmented generation with fine-grained knowledge refinement. NEUCOM, 655, 131282. (www)

  @article{ , abstract = {Iterative Retrieval-Augmented Generation has emerged as a promising strategy for maintaining the timeliness of information. Nevertheless, existing iterative retrieval-augmented generation methods are often plagued by irrelevant text noise during multi-round retrieval cycles, which can lead to degraded performance. To tackle this challenge, we propose a novel approach called Iterative Retrieval Augmented Generation with Fine-grained Knowledge Refinement (IRAGKR). IRAGKR comprises three key components: a dynamic retrieval gating module, a query semantic expansion module, and a document information compression module. Firstly, the dynamic retrieval gating module leverages the LLM’s self-attention mechanism, entropy calculation, and part-of-speech metrics to assess the necessity of retrieving generated answer fragments. Secondly, when the retrieval process is triggered, the query semantic expansion module formulates retrieval queries tailored to the model’s information needs based on attention mechanisms and prompt instructions to enhance retrieval precision. Finally, the document information compression module performs fine-grained knowledge filtering on the retrieved documents. IRAGKR is evaluated along with five different types of baseline methods on four open-domain question– answering datasets. Experimental results demonstrate the effectiveness of our method.}, archive = {J_NEUCOM}, author = {Kangning Du and Wenyu Wang and Benkui Zhang and Pengcheng Wang and Fan Zhang and Lin Cao and Yanan Guo}, doi = {10.1016/j.neucom.2025.131282}, journal = {Neurocomputing}, month = {11}, pages = {131282}, shortjournal = {Neurocomputing}, title = {IRAGKR:Iterative retrieval augmented generation with fine-grained knowledge refinement}, volume = {655}, year = {2025}, }
• (2025). Physics-aware parameter diffusion for spatio-temporal continuum field modeling for governance. NEUCOM, 655, 131262. (www)

  @article{ , abstract = {This paper proposes a two-stage framework named ST-PAD for spatio-temporal continuum field modeling in the field of earth sciences, aiming to achieve high-precision simulation and prediction of fluid dynamics through spatio-temporal physics awareness and parameter diffusion guidance. In the upstream stage, we design a vector quantization reconstruction module with temporal evolution characteristics, ensuring balanced and resilient parameter distribution by introducing general physical constraints. In the downstream stage, a diffusion probability network involving parameters is utilized to generate high-quality future states of fluids, while enhancing the model’s generalization ability by perceiving parameters in various physical setups. Extensive experiments on multiple benchmark datasets have verified the effectiveness and robustness of the ST-PAD framework, which showcase that ST-PAD outperforms current mainstream models in fluid dynamics modeling and prediction, especially in effectively capturing local representations and maintaining significant advantages in OOD generations.}, archive = {J_NEUCOM}, author = {An Zhang and Sheng Chen and Pengcheng Wu and Chao Wu and Hao Wu and Fan Xu and Kun Wang}, doi = {10.1016/j.neucom.2025.131262}, journal = {Neurocomputing}, month = {11}, pages = {131262}, shortjournal = {Neurocomputing}, title = {Physics-aware parameter diffusion for spatio-temporal continuum field modeling for governance}, volume = {655}, year = {2025}, }
• (2025). Multi-objective adaptive nesterov-like acceleration algorithm for multi-task deep learning. NEUCOM, 655, 131251. (www)

  @article{ , abstract = {With the wide application of multi-task learning in many deep learning problems, how to balance different tasks has become the most important challenge in multi-task learning training. This paper proposes a fast multi-objective adaptive momentum algorithm to address the multi-objective characteristics of multi-task learning and the slow descent of adaptive algorithms. Firstly, the algorithm adopts multi-objective adaptive technology to separate the cumulative gradient of each parameter in the adaptive algorithm, thereby separating the learning rate of each target task. Secondly, a Nesterov-like momentum acceleration is added to each task-level gradient, and it avoids the additional overhead of Nesterov acceleration algorithm in calculating gradients at extrapolation points. For the task-dominated problem, we describe the task weights by using the relative loss of the task and the normalized gradient paradigm, and normalize them, which can effectively balance each task. Finally, theoretically, the descending boundary of the multi-task aggregation loss function is characterized. The experimental results show that, compared with usual adaptive algorithms, the proposed algorithm has higher performance and faster convergence in training the scene parsing multi-task learning problems.}, archive = {J_NEUCOM}, author = {Fanyun Meng and Zhenyu Zhu and Liping Pang and Xiaoliang Wang}, doi = {10.1016/j.neucom.2025.131251}, journal = {Neurocomputing}, month = {11}, pages = {131251}, shortjournal = {Neurocomputing}, title = {Multi-objective adaptive nesterov-like acceleration algorithm for multi-task deep learning}, volume = {655}, year = {2025}, }
• (2025). Blending 3D geometry and machine learning for multi-view stereopsis. NEUCOM, 655, 131250. (www)

  @article{ , abstract = {Traditional multi-view stereo (MVS) methods primarily depend on photometric and geometric consistency constraints. In contrast, modern learning-based algorithms often rely on the plane sweep algorithm to infer 3D geometry, applying explicit geometric consistency (GC) checks only as a post-processing step, with no impact on the learning process itself. In this work, we introduce GC-MVSNet + + , a novel approach that actively enforces geometric consistency of reference view depth maps across multiple source views (multi-view) and at various scales (multi-scale) during the learning phase (see Fig. 1 ). This integrated GC check significantly accelerates the learning process by directly penalizing geometrically inconsistent pixels, effectively halving the number of training iterations compared to other MVS methods. Furthermore, we introduce a densely connected cost regularization network with two distinct block designs – simple and feature-dense – optimized to harness dense feature connections for enhanced regularization. Extensive experiments demonstrate that our approach achieves a new state-of-the-art on the BlendedMVS dataset, and competitive performance on the DTU and Tanks and Temples benchmarks. To our knowledge, GC-MVSNet + + is among the few approaches that enforce supervised geometric consistency across multiple views and at multiple scales during training. Our code is available at https://github.com/vkvats/GC-MVSNet-PlusPlus .}, archive = {J_NEUCOM}, author = {Vibhas Vats and Md Alimoor Reza and David Crandall and Soon-heung Jung}, doi = {10.1016/j.neucom.2025.131250}, journal = {Neurocomputing}, month = {11}, pages = {131250}, shortjournal = {Neurocomputing}, title = {Blending 3D geometry and machine learning for multi-view stereopsis}, volume = {655}, year = {2025}, }
• (2025). Control chart pattern recognition with variable window size for imbalanced data based on convolutional neural network with a convolutional block attention module. NEUCOM, 655, 131226. (www)

  @article{ , abstract = {In modern manufacturing processes, the complexity and variability of process conditions as well as the high degree of automation, impose greater demands on the real-time monitoring of production status. Control Chart Pattern Recognition (CCPR) is a crucial method for evaluating the condition of the production process, improving the ability to detect abnormalities and ensuring product quality and process stability. However, current CCPR methods typically employ a fixed-window design, which makes adapting to dynamically changing window requirements challenging. Additionally, these methods often exhibit low accuracy when handling imbalanced control chart pattern data. To address these challenges, this paper proposes a control chart pattern recognition method, MFF-CNN-CBAM, which combines a convolutional neural network with a convolutional block attention module and multi-feature fusion to recognize control chart patterns with variable window sizes in imbalanced data. The method first unifies the window size through a resampling technique. It then applies the synthetic minority oversampling technique (SMOTE) algorithm to balance the dataset, followed by extracting six mixed features, which include five statistical features and one shape feature. These features are fused with the convolutional features extracted by CNN-CBAM, and classification is conducted using a Softmax layer to increase recognition performance. The experimental results indicate that, compared to the raw data-only approach, the machine learning classifier approach, and other CCPR methods, this method demonstrates superior recognition performance, achieving an accuracy of 99.73 % when handling variable-window and imbalanced data. In addition, this paper develops an online monitoring system based on the CNN-CBAM model, which combined with the sliding window technique, can recognize normal and abnormal patterns in dynamic data in real time, achieving high recognition accuracy.}, archive = {J_NEUCOM}, author = {Li Xue and Dongsheng Zhu and Ruonan Li and Shubin Si and Haochen Wu}, doi = {10.1016/j.neucom.2025.131226}, journal = {Neurocomputing}, month = {11}, pages = {131226}, shortjournal = {Neurocomputing}, title = {Control chart pattern recognition with variable window size for imbalanced data based on convolutional neural network with a convolutional block attention module}, volume = {655}, year = {2025}, }
• (2025). TransDiff: Transformer-based diffusion model for low-light image enhancement. NEUCOM, 655, 131174. (www)

  @article{ , abstract = {Low-Light Image Enhancement (LLIE) addresses the challenge of augmenting image luminance to achieve visually natural, well-lit images. Diffusion models, recognized for their iterative denoising refinement and capacity to generate realistic images, have garnered significant attention within the realm of image synthesis. Inspired by the fidelity of diffusion models, we investigate their application to the low-light image enhancement domain, with the intent of generating high-quality, lifelike images with normal illumination. Throughout our exploration, two distinct challenges emerged: (1) diffusion model-based enhancement results are susceptible to global color bias, and (2) due to the inherent low visibility of dimly lit images, the structural fidelity of images generated by the diffusion model in such conditions is compromised, with blurring or even loss of structural details. To address these challenges, we propose TransDiff, a denoising diffusion model founded upon the Transformer architecture. Leveraging the global modeling prowess of the Transformer, TransDiff enhances the perceptual capabilities of the denoising network to mitigate the global color bias. Moreover, in response to the issue of missing intricate structures in regions of reduced luminance, we introduce a structure prior map. This map informs the design of a dual-stream denoising network governed by a priori conditions, guiding the diffusion model to produce normal-light images replete with enhanced structural intricacies. Extensive experimentation across various publicly accessible datasets attests to the superiority of our approach compared to existing state-of-the-art methods. This assessment encompasses both visual quality assessments and quantitative metrics.}, archive = {J_NEUCOM}, author = {Yonghua Zhang and Yujie Ding and Hongxuan Xie and Shuo Zhang and Hui Xu and Lvchen Cao and Ziqing Huang}, doi = {10.1016/j.neucom.2025.131174}, journal = {Neurocomputing}, month = {11}, pages = {131174}, shortjournal = {Neurocomputing}, title = {TransDiff: Transformer-based diffusion model for low-light image enhancement}, volume = {655}, year = {2025}, }
• (2025). PINE: Local patch reweighting and mixed independent neural encoder for datacentre workload prediction. NEUCOM, 655, 131170. (www)

  @article{ , abstract = {Datacentres are key pillars of modern digital services but consume vast amounts of energy and produce significant carbon emissions, largely due to inefficient resource utilisation that results from over-provisioning of computational resources. Obtaining prior knowledge of workload execution profiles can facilitate an optimum provisioning of resources in a way that reduces resource wastage whilst ensuring smooth workload execution. Herein, accurately forecasting the execution profiles of workloads can significantly reduce unnecessary energy consumption. However, achieving a reliable forecast of the execution profiles remains challenging due to the complexity, variability and interdependencies that are inherent in contemporary datacentre workloads. To promote datacentre sustainability, this paper proposes PINE: Local Patch Reweighing and Mixed Independent Neural Encoder, a novel Multilayer Perceptron (MLP)-based Neural Encoder for datacentre workload prediction on heterogeneous datacentre environments. PINE introduces a novel Mixed Independence (MI) architecture to overcome the long-debated Channel Dependence (CD) versus Channel Independence (CI) issue in time series prediction which enhances the model’s ability to learn inter-series dependencies. PINE also proposes Graph-based Equitable Resource Matching (GERM) algorithm for equitable pairing of interdependent series which avoids redundancy or imbalance. Finally, PINE provides novel Local Patch Reweighing (LPR) Block to capture local and multi-scale patterns in heterogeneous datacentre workloads. Experiments on real-world datasets from Alibaba and Google Clouds demonstrate the superior performance of the proposed PINE model against state-of-the-art approaches in terms of Mean Squared Error and Mean Absolute Error. The results validate the dependability of the proposed PINE model in contributing to sustainable datacentre operations whilst maintaining performance quality via a proactive management of datacentre workloads, thereby offering a wide range of industrial and environmental benefits.}, archive = {J_NEUCOM}, author = {Haowen Zheng and Yao Lu and Lu Liu and Zekun Sun and John Panneerselvam and Xiang Sun}, doi = {10.1016/j.neucom.2025.131170}, journal = {Neurocomputing}, month = {11}, pages = {131170}, shortjournal = {Neurocomputing}, title = {PINE: Local patch reweighting and mixed independent neural encoder for datacentre workload prediction}, volume = {655}, year = {2025}, }
• (2025). SAS-MRI: A semantic-guided approach for arbitrary scale super-resolution in multi-contrast brain MRI. NEUCOM, 654, 131389. (www)

  @article{ , abstract = {Super-resolution (SR) techniques have become increasingly significant in medical imaging. In clinical practice, radiologists are frequently zooming in on the magnetic resonance imaging (MRI) scans at arbitrary scales to fully visualize the lesions. However, most current arbitrary-scale SR techniques predominantly focus on learning pixel-to-pixel mapping from low-resolution (LR) to high-resolution (HR) images, thereby neglecting crucial high-level semantic information such as gray/white matter and inter-slice volumetric context. To tackle this problem, we introduce an innovative Semantic-guided Arbitrary-scale Super-resolution method for brain MRI (SAS-MRI). Our approach centers around leveraging semantic information from three main components: (i) a semantic-aware embedding module that effectively incorporates semantic priors into the feature representation space; (ii) a convolutional attention module that utilizes the information from adjacent slices within a volume; and (iii) a semantic-guided perceptual loss that enhances image quality by leveraging prior knowledge from medical image segmentation. Extensive experiments on two publicly available MRI datasets, fastMRI and IXI, demonstrate that our proposed SAS-MRI achieves superior Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) performance across multiple scaling factors. It consistently outperforms leading methods, achieving an average PSNR/SSIM of 31.997 dB / 0.899 on fastMRI and 37.871 dB / 0.985 on IXI, surpassing the next best methods by up to 0.20 dB / 0.002 and 0.78 dB / 0.006, respectively.}, archive = {J_NEUCOM}, author = {Zhen Li and Bo Li}, doi = {10.1016/j.neucom.2025.131389}, journal = {Neurocomputing}, month = {11}, pages = {131389}, shortjournal = {Neurocomputing}, title = {SAS-MRI: A semantic-guided approach for arbitrary scale super-resolution in multi-contrast brain MRI}, volume = {654}, year = {2025}, }
• (2025). WEI-SNNs: Spiking neural networks based on excitation-inhibition neurons and widening learnable time constants. NEUCOM, 654, 131386. (www)

  @article{ , abstract = {Spiking Neural Networks (SNNs) have gained significant attention in neuromorphic computing for their ability to process spatiotemporal information. However, the discontinuity and non-differentiability of spiking neuron models often lead to gradient vanishing during training, limiting network performance. Inspired by the spatial organization of excitation-inhibition (E-I) mechanisms in biological systems and the modularity of spiking neurons, a novel Excitation-Inhibition Leaky Integrate-and-Fire (EI-LIF) neuron model is proposed. Through dynamic regulation of the membrane potential, the model facilitates more effective gradient propagation. Furthermore, to construct high-performance SNNs from a temporal perspective, this study develops a step-wise time constant update strategy, jointly modulated by excitatory drive and temporal gating. This approach mitigates convergence difficulties in time constant optimization under strong E-I coupling. By integrating both spatial and temporal mechanisms, the Widened Learnable Time Constants and Excitation-Inhibition Spiking Neural Networks (WEI-SNNs) are proposed to enhance training stability and efficiency while maintaining biological plausibility. Experimental validation on static (CIFAR-10, CIFAR-100) and dynamic (DVS-Gesture, DVS-CIFAR10) datasets demonstrates Top-1 accuracies of 94.47 %, 75.66 %, 98.26 %, and 80.50 %, respectively, and shows good generalization on large-scale datasets such as ImageNet. Compared to state-of-the-art (SOTA) methods, WEI-SNNs achieve comparable accuracy with a significantly smaller number of parameters. This paper offers innovative algorithmic insights into efficient SNN design, supporting the development of brain-inspired intelligence.}, archive = {J_NEUCOM}, author = {Yuwei Chen and Jiawei Chen and Zhefei Cai and Yingle Fan and Yanming Wang and Minwei Zhu}, doi = {10.1016/j.neucom.2025.131386}, journal = {Neurocomputing}, month = {11}, pages = {131386}, shortjournal = {Neurocomputing}, title = {WEI-SNNs: Spiking neural networks based on excitation-inhibition neurons and widening learnable time constants}, volume = {654}, year = {2025}, }
• (2025). Frequency-conditional diffusion model for low-light image enhancement. NEUCOM, 654, 131383. (www)

  @article{ , abstract = {Low-light image enhancement (LLIE) aims to improve the visibility and quality of low-light images. The latest developments in diffusion model-based low-light enhancement have shown improvements in generating realistic, detailed images through iterative denoising. However, diffusion models often produce structural incompleteness and blurred contours due to their inability to effectively optimize the frequency–domain characteristics of noise. Specifically, existing methods fail to resolve the inconsistencies between low-frequency image features and noise components, leading to structurally incomplete restorations. Additionally, high-frequency details are easily disturbed by noise during restoration, leading to texture degradation. To address these problems, we propose a frequency-conditional diffusion model for low-light image enhancement. In the low-frequency domain, we employ Two-Dimensional Discrete Wavelet Transform (2D-DWT) to align noise and image features, maintaining the integrity of the image’s structural information. And in order to tackle detail reconstruction challenges, we introduce the Multi-directional Sparse High-frequency Reconstruction Module (MSHR). This module employs a learnable top-k operator to retain key values and suppress low-correlation interference, while integrating horizontal, vertical and diagonal details to enhance high-frequency detail reconstruction. Our method achieves excellent results on mainstream benchmark datasets, such as LOLv1 and LOLv2. Visualization results demonstrate that our method performs better in terms of structural and detail features. The source code is available at https://github.com/JunhaoWang17/Frediff .}, archive = {J_NEUCOM}, author = {Li Zhao and Junhao Wang and Chenxiang Fan and Zhonglong Zheng and Xiaoqin Zhang}, doi = {10.1016/j.neucom.2025.131383}, journal = {Neurocomputing}, month = {11}, pages = {131383}, shortjournal = {Neurocomputing}, title = {Frequency-conditional diffusion model for low-light image enhancement}, volume = {654}, year = {2025}, }
• (2025). A multi-view and multi-granularity emotional semantic interaction framework combining graph and attention mechanism for multimodal sentiment analysis. NEUCOM, 654, 131380. (www)

  @article{ , abstract = {Multimodal sentiment analysis aims to distinguish emotional leanings within data by examining information across multiple modalities. The primary responsibility is to effectively harness both the internal and external emotional correlations among modalities while thoroughly extracting implicit emotional cues to accommodate a wide range of semantic contexts. We present MMIF, a Multi-view and Multi-granularity Sentiment Semantic Interaction Framework that tackles these challenges in four concise steps: the Quantum-inspired Temporal Feature Extraction (QTFE) enriches each modality’s temporal dynamics with a quantum-structured LSTM; the Graph Neural Networks Enhanced Intra-modal Representation (GEIR) builds modality-specific graphs and taps diverse GNN variants to strengthen intra-modal reasoning; the Inter-modal Deep Interaction Fusion (IDIF) fuses cross-modal cues using a bidirectionally enhanced attention mechanism; the Emotion Representation Distribution Matching (ERDM) refines the final emotion predictions by capturing multi-granularity distributions of sentiment intensity information. Experimental results on the public datasets CMU-MOSI, CMU-MOSEI, and CH-SIMS show that the proposed model outperforms the compared methods in terms of performance, achieving the best accuracy scores of 89.12 %, 86.79 %, and 80.53 %, as well as F1 scores of 89.13 %, 86.80 %, and 80.71 %, respectively, demonstrating a significant performance advantage.}, archive = {J_NEUCOM}, author = {Zuhe Li and Xiang Guo and Huaiguang Wu and Jun Yu and Haoran Chen and Yifan Gao and Xiaowei Huang and Yushan Pan}, doi = {10.1016/j.neucom.2025.131380}, journal = {Neurocomputing}, month = {11}, pages = {131380}, shortjournal = {Neurocomputing}, title = {A multi-view and multi-granularity emotional semantic interaction framework combining graph and attention mechanism for multimodal sentiment analysis}, volume = {654}, year = {2025}, }
• (2025). An open-set few-shot face recognition framework with balanced adaptive cohesive mixing. NEUCOM, 654, 131379. (www)

  @article{ , abstract = {Face recognition has made great progress in closed-set and big data-supported scenarios. However, open-set few-shot face recognition (OSFSFR) is also in great demand for real-world applications, but the research on it remains at a rather preliminary and limited stage. The limited research on OSFSFR methods focuses on the optimization of models and metrics, yet they discard valuable base-class knowledge after pre-training. Although using pseudo-labels to retain this knowledge is a recent trend, it is hampered by two significant drawbacks: the injection of feature noise from unreliable labels and aggravated class imbalance caused by the base-novel class distribution gap. Therefore, we propose an innovative framework named Balanced Adaptive Cohesive Mixing (BACM), which effectively leverages base-class samples through three key strategies to tackle data scarcity and generate samples that align with novel class distributions: (1) a quality-based sample filtering mechanism to select informative base samples, (2) Center-guided Adaptive Cohesive Mixing (CGACM) for enhanced sample, and (3) Alleviating the Class Imbalance (ACI) which integrates linear and semantic pseudo-labels to mitigate class imbalance caused by pseudo-labeling. Extensive experimental evaluations on the CASIA and IJB-C datasets demonstrate the superior performance of our method in optimizing recognition accuracy for known identities and maintaining robust rejection capabilities for unknown subjects, surpassing the state-of-the-art approaches. The code is available at: https://github.com/chenxialiao/BACM .}, archive = {J_NEUCOM}, author = {Chenxia Liao and Huan Wan and Jianfeng Xu and Pengxiang Su and Xin Wei}, doi = {10.1016/j.neucom.2025.131379}, journal = {Neurocomputing}, month = {11}, pages = {131379}, shortjournal = {Neurocomputing}, title = {An open-set few-shot face recognition framework with balanced adaptive cohesive mixing}, volume = {654}, year = {2025}, }
• (2025). A holistic power optimization approach for microgrid control based on deep reinforcement learning. NEUCOM, 654, 131375. (www)

  @article{ , abstract = {Microgrid systems integrated with renewable energy sources (RES) and energy storage systems (ESS) have played a crucial role in providing more secure and reliable energy and deepening the penetration of renewables. However, optimizing the operational control of such an integrated energy system lacks the joint consideration of environmental, infrastructural and economic objectives. This paper presents a holistic data-driven power optimization approach based on deep reinforcement learning (DRL) for microgrid control, considering the multiple needs of decarbonization, sustainability and cost-efficiency. First, two control schemes, namely the prediction-based (PB) and prediction-free (PF) schemes, are devised to formulate the control problem. Second, a holistic reward function is designed to account for the operational costs, carbon emissions, peak load and battery degradation together. Third, a double dueling deep Q network (D3QN) is built to optimize the ESS charging and discharging strategies for real-time energy management. Finally, extensive simulations are conducted on a US microgrid to demonstrate the effectiveness of the proposed approach. Results show that the PB scheme outperforms the PF scheme when the prediction error is below 12.5 %, while the PF scheme becomes more effective as the error increases. It is also found that the D3QN with PB scheme can significantly reduce the annual operational cost and carbon emissions, while also greatly mitigating battery degradation and peak load.}, archive = {J_NEUCOM}, author = {Fulong Yao and Wanqing Zhao and Matthew Forshaw and Yang Song}, doi = {10.1016/j.neucom.2025.131375}, journal = {Neurocomputing}, month = {11}, pages = {131375}, shortjournal = {Neurocomputing}, title = {A holistic power optimization approach for microgrid control based on deep reinforcement learning}, volume = {654}, year = {2025}, }
• (2025). DABLN：An intelligent classification network based on breadth-based learning for noisy redundant signals. NEUCOM, 654, 131366. (www)

  @article{ , abstract = {In the field of artificial intelligence recognition and classification, deep-based learning and breadth-based learning have attracted extensive attention and research as two major methods. However, deep-based learning requires a large amount of data for training, and the training process is relatively slow. In contrast, breadth-based learning is fast to train and works well with small samples of data. Unfortunately, breadth-based learning may be more susceptible to noise and redundant features. Therefore, a new Distributed Adversarial Broad Learning Network (DABLN) is proposed. In DABLN, the risk of overfitting in the random mapped features is minimized through introducing distributed constraint terms in the optimization problem, resulting in a stable model that eliminates redundant information. Furthermore, a pair of non-parallel adversarial hyperplanes is obtained in the random feature space, which can capture different aspects and variations in the data, thereby reducing the sensitivity of the model to noise and other redundant information. The different experimental results show that compared to the breadth-based methods, DABLN method has significant advantages in precision, Recall, F 1 -score, Kappa and Accuracy.}, archive = {J_NEUCOM}, author = {Haoran Liu and Haiyang Pan and Jinde Zheng and Rui Yuan and Jinyu Tong}, doi = {10.1016/j.neucom.2025.131366}, journal = {Neurocomputing}, month = {11}, pages = {131366}, shortjournal = {Neurocomputing}, title = {DABLN：An intelligent classification network based on breadth-based learning for noisy redundant signals}, volume = {654}, year = {2025}, }
• (2025). FOATLBL: Federated online active transfer learning via broad learning system. NEUCOM, 654, 131363. (www)

  @article{ , abstract = {Federated learning (FL) is a new paradigm of machine learning in which clients send trained local models to server, and the server aggregates these models into a global model. FL has received a lot of attention due to its security and effectiveness. However, the existing methods depend heavily on a large number of labeled samples from clients, and FL performance deteriorates when clients have only a few labeled samples or poor quality labeled samples. To face this challenge, we propose federated online active transfer learning via broad learning (FOATLBL). In FOATLBL, clients can obtain more important labeled samples in the target domain and adapt to model training with limited labeled samples. FOATLBL also features a new personalized aggregation algorithm that allows the aggregated model to better adapt to local data on the client. Finally, extensive experiments on the transfer learning public datasets and the human activity recognition datasets show superior performance of FOATLBL over the baseline.}, archive = {J_NEUCOM}, author = {Chang-E Ren and Siyao Cheng and Zehua Xuan}, doi = {10.1016/j.neucom.2025.131363}, journal = {Neurocomputing}, month = {11}, pages = {131363}, shortjournal = {Neurocomputing}, title = {FOATLBL: Federated online active transfer learning via broad learning system}, volume = {654}, year = {2025}, }
• (2025). Eigen-informed neural ordinary differential equations: Incorporating knowledge of system properties. NEUCOM, 654, 131358. (www)

  @article{ , abstract = {Using neural ordinary differential equations (ODEs) to model complex systems is still a challenging venture and fails for various reasons, often resulting in convergence to unsatisfactory local minima or unintended termination of the training process due to solver issues. To take a step back, the root of the problem lies in either the complexity of the optimization problem or the lack of data. The negative effects of both aspects can be reduced by incorporating a priori knowledge, a common strategy in Scientific Machine Learning . Especially when modeling physical systems, there is almost always more system knowledge available than is used for training. Examples range from knowledge of generic properties such as “the system is stable” to more quantifiable attributes such as “the system oscillates in a known frequency range”. To close the loop, such knowledge can be used to achieve faster convergence and better generalization. In this paper, we focus specifically on system properties that can be expressed based on eigenvalues, such as (partial) stability, oscillation capability, frequencies, damping, and stiffness. We explain how such properties can be intuitively integrated into training neural ODEs, provide open-source software for this, and finally show in three academic examples that such eigen-informed neural ODEs are able to converge in fewer steps (median up to factor 6.41 ), generalize better (median up to factor 6.92 ) and are solvable more efficiently (median up to factor 1.12 ) compared to pure neural ODEs, and can even reconstruct a given system on undersampled data.}, archive = {J_NEUCOM}, author = {Tobias Thummerer and Lars Mikelsons}, doi = {10.1016/j.neucom.2025.131358}, journal = {Neurocomputing}, month = {11}, pages = {131358}, shortjournal = {Neurocomputing}, title = {Eigen-informed neural ordinary differential equations: Incorporating knowledge of system properties}, volume = {654}, year = {2025}, }
• (2025). Text-graph alignment based on LLMs instruction learning and wavelet graph transformers for brain disease analysis. NEUCOM, 654, 131355. (www)

  @article{ , abstract = {Brain graphs from functional magnetic resonance imaging (fMRI) combined with phenotypic texts serve as fundamental data modalities for brain disease analysis. The lack of alignment between self-supervised signals in the graph modalities of brain graphs and the text modalities of phenotypic data in current multimodal learning methods has led to suboptimal diagnostic performance and limited interpretability of brain disease patterns. To address these limitations, we introduce the Brain Text–Graph Alignment Model (BrainTGA), which integrates brain disease diagnosis with pathological analysis. Our model creates a shared semantic space that aligns text and graph data by using Large Language Models (LLMs) instruction learning to enhance phenotypic texts and implementing a wavelet graph Transformer model to encode brain graphs. Our comprehensive evaluation of benchmark datasets with heterogeneity and site bias demonstrates that BrainTGA outperforms existing state-of-the-art methods in diagnostic accuracy. It also has advantages in terms of time and space complexity and provides superior interpretability of neural patterns consistent with the diagnoses of clinical experts. The open source model implementation of our BrainTGA is available at https://github.com/Mmurphyy/BrainTGA .}, archive = {J_NEUCOM}, author = {Wenfeng Tan and Hong Yang and Shanshan Ye and Ji Zhang and Peng Zhang and Zhi Qiao}, doi = {10.1016/j.neucom.2025.131355}, journal = {Neurocomputing}, month = {11}, pages = {131355}, shortjournal = {Neurocomputing}, title = {Text-graph alignment based on LLMs instruction learning and wavelet graph transformers for brain disease analysis}, volume = {654}, year = {2025}, }
• (2025). Triple contrastive learning framework for domain-level zero-shot recommendation. NEUCOM, 654, 131354. (www)

  @article{ , abstract = {Domain-level zero-shot recommendation (DZSR) poses a significant challenge in recommender systems (RS), representing a cold-start scenario where the domain completely lacks interaction history. Previous works leveraged universal knowledge from item content to bridge domains, transferring user preferences from the source domain to the target domain. However, previous works relied solely on item content similarity for DZSR, while overlooking both co-interaction relevance and topic relevance. These limitations not only restrict the exploration of user interests, but also further hinder the transfer of user preferences across domains. To address limitations, this paper proposes a T riple C ontrastive L earning F ramework (TCLF) to perform three tailored alignments. Specifically, we first design a self-contrastive learning to align the embeddings generated from item descriptions and knowledge graph triples of the same item. Then, we measure the co-interaction relevance from both the user and item perspectives, and align co-interacted source items by a source-domain contrastive learning. Finally, we measure the topic relevance in an indirect manner, and design a cross-domain contrastive learning to align topic-relevant source and target items. Extensive experiments on four public datasets demonstrate the superiority of TCLF for DZSR tasks compared to state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Huanyuan Zhou and Xue Chen and Wenjuan Cha and Ruonan Gu and Peng Wang}, doi = {10.1016/j.neucom.2025.131354}, journal = {Neurocomputing}, month = {11}, pages = {131354}, shortjournal = {Neurocomputing}, title = {Triple contrastive learning framework for domain-level zero-shot recommendation}, volume = {654}, year = {2025}, }
• (2025). RIANet: Regional interaction attention network for the segmentation of ultrasound images. NEUCOM, 654, 131353. (www)

  @article{ , abstract = {Ultrasound, a widely used medical imaging technique, plays a crucial role in diagnosis and treatment. The precise and efficient segmentation of ultrasound images is essential for localizing organs and identifying pathological changes. Recent models for ultrasound image segmentation, drawing inspiration from Vision Transformers, predominantly utilize self-attention mechanisms. Although self-attention mechanisms excel in capturing global feature information, they often overlook essential local feature details while facing challenges of high computational complexity. To address these issues, we propose Regional Interaction Attention, a lightweight multi-scale attention mechanism. Regional Interaction Attention, introducing the concept of Regional Interaction, interacts with self-attention components to effectively address two critical challenges inherent in self-attention mechanisms: the lack of local feature information and high computational complexity. Furthermore, we propose the Regional Interaction Attention Network, a hierarchical network architecture designed for effective semantic segmentation in ultrasound images. Experimental results show that our RIANet outperforms its counterparts in segmentation tasks, achieving the overall best results with an mIoU of 80.72 % and a Dice coefficient of 76.12 % in the DDTI dataset, an mIoU of 76.09 % and a Dice coefficient of 70.66 % in the BUSI dataset, and an mIoU of 76.13 % in the AUS dataset. External validation on TN-SCUI and BUID datasets further verifies the effectiveness and generalization of RIANet. RIANet exhibits promising prospects for ultrasound image segmentation tasks.}, archive = {J_NEUCOM}, author = {Shiyao Sun and Chong Fu and Yingyou Wen and Xingwei Wang and Chiu-Wing Sham}, doi = {10.1016/j.neucom.2025.131353}, journal = {Neurocomputing}, month = {11}, pages = {131353}, shortjournal = {Neurocomputing}, title = {RIANet: Regional interaction attention network for the segmentation of ultrasound images}, volume = {654}, year = {2025}, }
• (2025). GLNet: A network with different cascading methods for global and local features. NEUCOM, 654, 131346. (www)

  @article{ , abstract = {Segmenting the liver in medical images is crucial for a variety of clinical applications, providing important information for early diagnosis and surgical planning of liver-related diseases. While existing models commonly rely on convolutional neural networks for hierarchical modeling of local features for a long time, Transformer-based architectures provide a new and enhanced perspective through the ability of long-range dependency capture of global context, and the complementarity between the two in terms of local accuracy and global consistency is becoming a research focus. However, Transformers often struggle to preserve detailed features, which may limit their segmentation performance. To address this question, we propose a novel segmentation model that achieves two-way dynamic fusion of local details and global context through a synergistic and complementary mechanism of hierarchical local feature enhancement and global semantic association modeling. At the core of our approach is a Gaussian weighting mechanism that enhances texture and edge information by assigning higher weights to pixels closer to the center of the feature while suppressing interference from surrounding regions. This mechanism enables the model to preserve key details while effectively capturing global information. In addition, we introduce the Fourier transform to extract high-frequency components from low-resolution feature maps, further enhancing the model’s ability to capture features, which are critical for accurate segmentation. Experimental results performed on four publicly available datasets show that our proposed model achieves state-of-the-art performance in both segmentation accuracy and generalization ability, highlighting its effectiveness in medical image segmentation tasks. The code will be available at https://github.com/chengxin-su/GLNet.git .}, archive = {J_NEUCOM}, author = {Jiao Du and Chengxin Su and Zhao Qun and Xiaoyu Yu}, doi = {10.1016/j.neucom.2025.131346}, journal = {Neurocomputing}, month = {11}, pages = {131346}, shortjournal = {Neurocomputing}, title = {GLNet: A network with different cascading methods for global and local features}, volume = {654}, year = {2025}, }
• (2025). PAttL-GAN: Pixel attention localization generative adversarial network for unsupervised anomaly detection in medical images. NEUCOM, 654, 131345. (www)

  @article{ , abstract = {Medical anomaly detection (MAD) plays a vital role in clinical diagnosis because the detected results provide interpretable pathological information for actual diagnosis. Unlabeled datasets, which include abundant normal and abnormal images, have not been effectively utilized. Further, most works cannot achieve fine-grained localization of pixel-level anomaly region. To solve these problems, we propose P ixel Att ention L ocalization G enerative A dversarial N etwork ( PAttL-GAN ). We first design a multi-level efficient fusion (MEF) module to adaptively assign weights to different layers, enabling the effective fusion of both local and global information. We then create a pixel-level anomaly localization (PAL) module which consists of an adaptive multi-scale attention (AMA) mechanism and a pixel-level weighting (PLW) strategy to improve the saliency of localized abnormal regions and distinguish abnormal regions from complex textures. Comprehensive experiments on four publicly available datasets validate the effectiveness and generalization capability of PAttL-GAN. It surpasses state-of-the-art baselines while maintaining a relatively lightweight structure. PAttL-GAN is a robust, efficient, and interpretable solution for MAD across different imaging modalities and body regions. We also provide sufficient visualization results to enhance the model’s interpretability, offering an innovative perspective for eXplicable Artificial Intelligence (XAI).}, archive = {J_NEUCOM}, author = {Guangli Li and Yuxing Zou and Boyang Liu and Nan Jiang and Jingqin Lv and Wenhao Ai and Donghong Ji and Qingzhong Zhou and Hongbin Zhang}, doi = {10.1016/j.neucom.2025.131345}, journal = {Neurocomputing}, month = {11}, pages = {131345}, shortjournal = {Neurocomputing}, title = {PAttL-GAN: Pixel attention localization generative adversarial network for unsupervised anomaly detection in medical images}, volume = {654}, year = {2025}, }
• (2025). Dual-decoder conditional diffusion model based on spatial-domain difference compensation pre-fusion for infrared and visible image fusion. NEUCOM, 654, 131342. (www)

  @article{ , abstract = {Infrared and visible image fusion aims to preserve both intensity and texture details from the two modalities. While generative adversarial networks (GANs) have shown promise, they suffer from mode collapse and limited interpretability. To address these challenges, we propose a dual-decoder conditional diffusion model based on spatially compensated pre-fusion, termed SDC-DDF. This method tackles the lack of prior knowledge in fusion distribution and reduces the excessive inference steps typical of diffusion models. By leveraging spatial discrepancies between the two modalities, pre-fused images are generated as noisy "fused ground truth", serving as coarse prior in the forward diffusion process. Additionally the refinement decoder refines the pre-fused images with varying levels of noise. During the reverse denoising, coarse reconstruction decoder restores a coarse fused image from random noise in just five denoising steps, which is then refined by the refinement decoder to produce a high-quality final result. Experimental results on five publicly available datasets demonstrate that SDC-DDF outperforms eleven state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Zhenjia Zhang and Kai Huo and Jun Li and Shuaifeng Zhi}, doi = {10.1016/j.neucom.2025.131342}, journal = {Neurocomputing}, month = {11}, pages = {131342}, shortjournal = {Neurocomputing}, title = {Dual-decoder conditional diffusion model based on spatial-domain difference compensation pre-fusion for infrared and visible image fusion}, volume = {654}, year = {2025}, }
• (2025). DBAE: Decision boundary based adversarial example generation via multi-step optimization. NEUCOM, 654, 131341. (www)

  @article{ , abstract = {Adversarial examples are deceptive texts generated by introducing subtle perturbations into original samples, often imperceptible to humans. Existing adversarial example generation methods primarily focus on text modality, achieving high attack success rates but often at the cost of large perturbations, which compromise the quality of generated examples and may cause semantic drift. To address these limitations, we propose Decision Boundary-based Adversarial Example Generation (DBAE), a sentence-level method that leverages the decision boundary characteristics of the target model’s vector space. By optimizing the relative positioning of feature vectors to the decision boundary, DBAE generates high-quality adversarial examples that remain semantically close to the original text. Our approach begins with a Tree-based Autoencoder (Tree-AE), which generates initial adversarial example feature vectors by injecting perturbations into the root node of the text. Guided by these initial vectors, the Particle Swarm Optimization (PSO) algorithm searches for target feature vectors near the decision boundary while preserving semantic similarity. Finally, a Conditional LLM (c-LLM) maps the optimized feature vectors to high-quality adversarial examples. Experiments on multiple models and datasets demonstrate that DBAE not only maintains a high attack success rate but also improves perplexity and semantic similarity compared to traditional methods. Despite a slightly higher average generation time of 7.82 s on an RTX 4090, our decision-boundary method attains > 85 % semantic similarity and strong human-evaluation scores, confirming its effectiveness and practical value for adversarial text generation.}, archive = {J_NEUCOM}, author = {Yu Sun and Peiqi Sun and Yanhui Du}, doi = {10.1016/j.neucom.2025.131341}, journal = {Neurocomputing}, month = {11}, pages = {131341}, shortjournal = {Neurocomputing}, title = {DBAE: Decision boundary based adversarial example generation via multi-step optimization}, volume = {654}, year = {2025}, }
• (2025). RETO: Reinforcement learning enhanced terminology optimization for cyber threat intelligence summarization. NEUCOM, 654, 131339. (www)

  @article{ , abstract = {Cyber threat intelligence summarization consolidates raw intelligence into concise summaries for decision-makers, playing a crucial role in rapid, accurate detection and response to cyber threats. However, existing research primarily focuses on improving the fluency, overlap, and accuracy of summaries, often neglecting the optimization of the correctness and coverage of specialized terminology. While recent advances in LLMs technology show promise, they frequently produce hallucinations, especially in specialized areas, undermining the reliability and accuracy of summaries. In this paper, a R einforcement Learning E nhanced T erminology O ptimization ( RETO ) approach is proposed for cyber threat intelligence summarization. RETO ’s three modules include a two-stage terminology annotator, a threat intelligence-specific terminology generator, and a multi-reward-aware terminology optimizer. The first module provides high-quality annotated data, the second module trains an automatic terminology generator, and the final module optimizes the terminology generator, guiding LLMs to generate more accurate summaries within the reinforcement learning framework. Additionally, a multi-reward-aware function is designed to improve the terminology generation quality by maximizing a combination of ROUGE-L, Terminology Correctness, and Homologous Terminology Coverage. Experimental results on the CTISum dataset show that RETO achieves state-of-the-art performance in zero-shot summarization tasks, and the introduction of domain-specific terminology effectively reduces hallucinations generated by LLMs, leading to more accurate and reliable summaries. We also make our code publicly available at https://github.com/JmeiDing/RETO .}, archive = {J_NEUCOM}, author = {Junmei Ding and Wei Peng and Yueming Lu and Zhiqiang Wang and Yan Feng}, doi = {10.1016/j.neucom.2025.131339}, journal = {Neurocomputing}, month = {11}, pages = {131339}, shortjournal = {Neurocomputing}, title = {RETO: Reinforcement learning enhanced terminology optimization for cyber threat intelligence summarization}, volume = {654}, year = {2025}, }
• (2025). Fusing beyond homophily and low-frequency information in graph neural networks. NEUCOM, 654, 131332. (www)

  @article{ , abstract = {It is significant to repeatedly aggregate neighborhood information for graph representation. However, because most current graph neural networks (GNNs) are based on the assumption of homophily, they can only perform well for graphs with strong homophily, but are not able to be directly generalized to those graphs with strong heterophily. Therefore, many existing researches have been devoted to fusing higher-order neighborhoods or high-frequency signals for these graphs with strong heterophily. In this paper, a beyond homophily and low-frequency information graph convolutional network, HFGCN, is proposed for better generalizing graph convolutional network (GCN) to strong heterophily settings. Specifically, the adaptive-fusion of the low-frequency and high-frequency information of first-order neighborhood is aimed at substituting for fusing higher-order neighborhood. And then, the separate ego-embedding, neighbor-embedding, and integration of the intermediate representations are used in the HFGCN. Finally, the above three key designs are combined into a GCN framework. Extensive experiments on nine real-world datasets validate that our model outperforms the state-of-the-art under both graphs with strong homophily and heterophily, and further relieves the over-smoothing problem.}, archive = {J_NEUCOM}, author = {Yangcai Xie and Jiecheng Li and Wenzhen Zhang and Jian Wei and Guangquan Lu and Shichao Zhang}, doi = {10.1016/j.neucom.2025.131332}, journal = {Neurocomputing}, month = {11}, pages = {131332}, shortjournal = {Neurocomputing}, title = {Fusing beyond homophily and low-frequency information in graph neural networks}, volume = {654}, year = {2025}, }
• (2025). A measure of conditional dependence for multivariate compositional data. NEUCOM, 654, 131331. (www)

  @article{ , abstract = {Recent advances in predictive modeling of multivariate compositional data, which encompass scenarios with multiple compositional predictors and either compositional or scalar responses, have highlighted critical needs for effective dependence measurement. Current methodologies face limitations in quantifying conditional dependencies between responses and individual predictors while controlling for confounding variables, particularly when addressing nonlinear associations. To bridge this gap, we present a novel Compositional Conditional Dependence Coefficient (CoCDC) featuring four key innovations. (1) We introduce a nearest neighbor analysis designed for compositional predictors, and establish a coefficient for measuring conditional dependence between scalar responses and multivariate compositional predictors. (2) Through optimal transport theory, we propose a multivariate rank for compositions with formal definitions and estimation procedures. (3) We further extend CoCDC to handle predictions involving both compositional responses and predictors through integration of the proposed multivariate rank for compositional responses and the nearest neighbor analysis for compositional predictors. (4) We propose a Feature Ordering by Compositional Conditional Dependence (FOCCD) algorithm that leverages the CoCDC to enable effective variable selection. Extensive simulations reveal that CoCDC achieves competitive performance in detecting both linear and nonlinear marginal associations compared to benchmarks, while establishing the first measurement for quantifying joint dependencies in compositional predictions with either compositional or scalar responses. The FOCCD algorithm demonstrates superior performance in complex prediction scenarios involving nonmonotonic marginal associations and joint dependencies, outperforming existing methods in variable selection accuracy. Empirical studies through real-world datasets further confirm the superiority of CoCDC, with FOCCD-selected predictors yielding parsimonious and interpretable models.}, archive = {J_NEUCOM}, author = {Keyao Wang and Huiwen Wang and Shan Lu and Lihong Wang}, doi = {10.1016/j.neucom.2025.131331}, journal = {Neurocomputing}, month = {11}, pages = {131331}, shortjournal = {Neurocomputing}, title = {A measure of conditional dependence for multivariate compositional data}, volume = {654}, year = {2025}, }
• (2025). StyleMorpheus: Learning a StyleGAN-based 3D-aware morphable face model with a disentangled style space. NEUCOM, 654, 131329. (www)

  @article{ , abstract = {Recent advances in 3D-aware neural rendering have enabled photorealistic face image synthesis from arbitrary viewpoints. However, achieving disentangled control over facial attributes typically depends on large, curated datasets collected in controlled environments. To overcome this limitation, we introduce StyleMorpheus, a 3D-aware, StyleGAN-based morphable face model that can be trained entirely on in-the-wild face images. StyleMorpheus surpasses traditional 3D-aware morphable models in rendering quality, despite relying solely on unconstrained 2D training data. Unlike conventional StyleGAN-based methods, StyleMorpheus also provides disentangled control over facial identity, expression, and appearance, allowing each attribute to be adjusted independently without unintended changes to the others. StyleMorpheus employs an auto-encoder structure, where the encoder learns a representative, disentangled style code space, and the decoder enforces disentanglement by using shape- and appearance-related codes at different levels of the network. Furthermore, we fine-tune the decoder through StyleGAN-based generative adversarial learning to achieve photorealistic rendering quality. StyleMorpheus is computationally lightweight and achieves real-time rendering speeds, making it suitable for virtual reality applications. We further demonstrate the disentanglement capabilities of StyleMorpheus through face editing tasks such as style mixing, face morphing, and color editing. Project homepage: https://peizhiyan.github.io/docs/morpheus .}, archive = {J_NEUCOM}, author = {Peizhi Yan and Rabab K. Ward and Dan Wang and Qiang Tang and Shan Du}, doi = {10.1016/j.neucom.2025.131329}, journal = {Neurocomputing}, month = {11}, pages = {131329}, shortjournal = {Neurocomputing}, title = {StyleMorpheus: Learning a StyleGAN-based 3D-aware morphable face model with a disentangled style space}, volume = {654}, year = {2025}, }
• (2025). In depth analysis for securing the truth: Addressing the fake news challenge with graph neural networks. NEUCOM, 654, 131327. (www)

  @article{ , abstract = {The fake news phenomenon has a significant impact on societies, homeland security, democracy and the functioning of the public space. The spread of false information is becoming an increasing challenge in the context of the dynamic growth of the volume of content shared by news outlets and social media. The overwhelming amount of this information makes manual verification of every news item or press release practically impossible. The current development of technology in the field of natural language processing (NLP) opens up new possibilities for the development of automatic content verification systems. The automation of this process not only improves but also significantly speeds up the detection of unreliable information, which is a key tool in the fight against fake news. In this article, we propose an innovative approach that involves a multi-factor assessment of the content of documents, as opposed to the frequently used approach of binary classification into fake and non-fake. Our classification system is based on analysis using graph neural networks, which allows for a more complex and contextual understanding of the data. The obtained results indicate a significant improvement in effectiveness compared to the baseline approaches, which suggests a potential for enhanced mitigation of misinformation dissemination.}, archive = {J_NEUCOM}, author = {Gracjan Kątek and Rafał Kozik and Aleksandra Pawlicka and Marek Pawlicki and Michał Choraś}, doi = {10.1016/j.neucom.2025.131327}, journal = {Neurocomputing}, month = {11}, pages = {131327}, shortjournal = {Neurocomputing}, title = {In depth analysis for securing the truth: Addressing the fake news challenge with graph neural networks}, volume = {654}, year = {2025}, }
• (2025). S2F: Border peeling clustering from cluster skeletons to final clusters. NEUCOM, 654, 131324. (www)

  @article{ , abstract = {The border-peeling clustering algorithm (BP) has demonstrated superior performance in clustering tasks, however, its accuracy in identifying core points and border points remains insufficient. To address this limitation, this paper proposes a novel BP framework termed S2F, which transitions from clustering skeletons to final clusters. Specifically, S2F innovatively introduces a bidirectional k -nearest neighbor ( BkNN ) structure to reconstruct density influence calculations, enabling non-iterative partitioning of core and border points based on the density differences between each data point and its bidirectional neighbors. Subsequently, an undirected graph is constructed using the core points, and connected components are extracted to form the initial clusters (i.e., clustering skeletons). Finally, border points are assigned to the skeletons based on their associations with core points, generating the final clustering results. To validate the effectiveness of S2F, comparative experiments were conducted on 10 synthetic datasets and 8 real-world datasets. The results demonstrate that S2F significantly outperforms the original BP algorithm and its advanced variants (such as ROBP, FBPStream, and TWStream) in most cases.}, archive = {J_NEUCOM}, author = {Shuntai Zhang and Yizhang Wang and Ming Yang and Witold Pedrycz}, doi = {10.1016/j.neucom.2025.131324}, journal = {Neurocomputing}, month = {11}, pages = {131324}, shortjournal = {Neurocomputing}, title = {S2F: Border peeling clustering from cluster skeletons to final clusters}, volume = {654}, year = {2025}, }
• (2025). Fast semantic scene completion via two-stage representation. NEUCOM, 654, 131323. (www)

  @article{ , abstract = {Semantic scene completion, or occupancy prediction, aims to complete and segment environments simultaneously based on incomplete sensor observations. Due to its comprehensive perception capabilities, this technology is becoming a trend in 3D scene understanding. However, semantic scene completion methods require high computational resources, which hinders their real-world application. In this work, we propose an efficient two-stage fast semantic scene completion method called Fast-SSC. In the first stage, a Geometric Completion Network built upon the Hybrid Parallel Dilated Block is designed to extract multi-scale geometric features and complete the scene. In the second stage, a Bird’s Eye View network with the Spatial Awareness Enhancement Attention mechanism is employed to achieve fast scene segmentation. Extensive experiments on the SemanticKITTI dataset show that the proposed Fast-SSC achieves state-of-the-art performance. Specifically, the Fast-SSC ranks first among BEV-based methods and operates at 41.8 FPS on NVIDIA GeForce GTX 1080 Ti. These results indicate that our Fast-SSC efficiently utilizes scene information and holds the potential for practical deployment. Code is available at https://github.com/six-wood/Fast-SSC .}, archive = {J_NEUCOM}, author = {Musen Lin and Wenguang Wang}, doi = {10.1016/j.neucom.2025.131323}, journal = {Neurocomputing}, month = {11}, pages = {131323}, shortjournal = {Neurocomputing}, title = {Fast semantic scene completion via two-stage representation}, volume = {654}, year = {2025}, }
• (2025). Federated dynamic graph neural network for cross-modal organizational real-time community detection. NEUCOM, 654, 131322. (www)

  @article{ , abstract = {As organizational networks grow increasingly dynamic and multimodal, real-time community detection has become critical for multinational restructuring and emergency response. Existing methods face challenges in modeling event cascades, resolving cross-modal semantic alignment, and balancing privacy protection with detection accuracy. This paper presents Federated Dynamic Cross-Modal Community Detection Framework(FedDyCM), the core innovation is mainly reflected in: (1) integration of Hawkes processes with temporal graph neural networks to quantify event cascades via self-exciting intensity matrices for modeling sudden structural shifts; (2) gated cross-modal attention that dynamically weights text, topological, and behavioral features through context-aware attention to enhance semantic alignment by adaptively elevating critical modality weights during crisis scenarios; (3) a federated privacy-alignment mechanism combining local differential privacy with Procrustes embedding alignment to preserve data sovereignty while ensuring global semantic consistency and limiting privacy attack rates to < 5 %; and (4) entropy-constrained fairness to prevent resource monopolization by limiting maximum community size. Experimental results show: This framework satisfies the criteria of cascade effects, cross-modal fusion, balanced distribution, and privacy preservation, achieving real-time community partitioning under privacy constraints. FedDyCM achieves a dynamic modularity score Q of 0.50 ± 0.01 , elevates the F1-score for community reorganization detection to 0.88 ± 0.01 , and improves Normalized Mutual Information (NMI) by 13.3 percentage points compared to baseline models. This framework provides efficient, secure, and equitable decision support for multinational organizational management, advancing practical application of dynamic graph neural networks in complex real-world scenarios.}, archive = {J_NEUCOM}, author = {You-hong Li and Yinwei Zhan and Le Gao and Tian-Yu Wang and Hao-Ming Mo and Adel Nasri}, doi = {10.1016/j.neucom.2025.131322}, journal = {Neurocomputing}, month = {11}, pages = {131322}, shortjournal = {Neurocomputing}, title = {Federated dynamic graph neural network for cross-modal organizational real-time community detection}, volume = {654}, year = {2025}, }
• (2025). VS-SMOTE: Leveraging high-value spaces to balance noise control and data diversity for class imbalanced learning. NEUCOM, 654, 131321. (www)

  @article{ , abstract = {In machine learning, a key approach for addressing the class imbalance problem is the Synthetic Minority Over-sampling TEchnique (SMOTE). However, the problem of introducing low-quality samples or noise remains. Many methods have been proposed to address this problem, which have failed to accurately assess spatial quality. Such methods focused overly on the neighborhoods of the original samples, rather than those of the synthetic ones. Additionally, they struggle to balance noise control with data diversity. To address these challenges, a novel method called VS-SMOTE is proposed in this paper. First, the space in which synthetic samples are located in multiple subspaces is viewed and middle samples are used to assess their spatial quality. Second, high-value spaces are selected to generate synthetic samples using two spatial selections. Using efficient spatial assessment and selection methods, VS-SMOTE avoids noise generation and maximizes synthetic sample generation in high-value spaces. Extensive experimental results for 30 real datasets, two synthetic datasets and six classical classifiers demonstrate that VS-SMOTE outperformed seven representative state-of-the-art methods. Particularly, VS-SMOTE improved the average AUC, F1 and G-mean scores by 2.49 % - 9.47 %, 8.19 % - 22.63 % and 7.83 % - 17.15 %, respectively, compared to SMOTE. Thus, VS-SMOTE effectively handles the class imbalance problem by constraining numerous new samples in multiple high-value linear spaces.}, archive = {J_NEUCOM}, author = {Fuxing Qiu and Fengyu Yang and Yu Xiong and Tong Chen and Peng Xiao and Wei Zheng and Xingui Chen and Kejin Zhou}, doi = {10.1016/j.neucom.2025.131321}, journal = {Neurocomputing}, month = {11}, pages = {131321}, shortjournal = {Neurocomputing}, title = {VS-SMOTE: Leveraging high-value spaces to balance noise control and data diversity for class imbalanced learning}, volume = {654}, year = {2025}, }
• (2025). Learning to select representative instances based on neighborhood distribution. NEUCOM, 654, 131320. (www)

  @article{ , abstract = {As an important technique in data mining and machine learning, instance selection aims to select an instance subset from an original dataset, which can maintain the characteristics of the original dataset as much as possible. Then, the representative instance subset can be used for model training or data mining, in order to achieve or approximately achieve the effect of using all of the instances. Since data distribution usually affects the performance, generalization capability and accuracy of the classification or regression models, it indeed plays a pivotal role in the construction of the models. Therefore, data distribution is also crucial for instance selection. By considering both the density and the decision information of the neighborhood of an instance, we propose a neighborhood distribution-based instance selection algorithm (NDIS). Specifically, the concept of neighborhood density is firstly given to reflect the dispersion of data points around an instance. Then, by combining the neighborhood density and the neighborhood decision information, a measure NDLC is provided for selecting representative instances. Finally, an instance selection algorithm NDIS is proposed in which NDLC is utilized to formulate an instance selection strategy, and a wrapper technique is provided to select the optimal instance subset. Numerical experiments show that NDIS possesses the satisfactory performance.}, archive = {J_NEUCOM}, author = {Xiao Zhang and Qian Yang and Ting Qian}, doi = {10.1016/j.neucom.2025.131320}, journal = {Neurocomputing}, month = {11}, pages = {131320}, shortjournal = {Neurocomputing}, title = {Learning to select representative instances based on neighborhood distribution}, volume = {654}, year = {2025}, }
• (2025). Interpretable local-global monitoring network for three-phase flow processes with concurrent analysis of independence and high-order temporal correlation. NEUCOM, 654, 131319. (www)

  @article{ , abstract = {Oil-gas-water three-phase flow is a dynamic and coupled process with numerous flow states, which is often monitored by multiple sensors. Different evolutions of flow states will impact various combinations of variables, resulting in local process correlations. Additionally, effective monitoring of three-phase flow requires a global analysis of phase relations (oil-water relation and gas-liquid relation) hidden in the coupled process data. For a comprehensive monitoring of three-phase flow process, this work proposes an interpretable local-global monitoring network (MoniNet) based on convolutional independent slow and steady feature analysis (Conv-IS 2 FA). In this MoniNet, Conv-IS 2 FA integrates the concepts of static independence and dynamic slowness, simultaneously exploring the high-order statistics and temporal correlation to extract latent Gaussian or non-Gaussian components from mixed multi-sensor signals. Firstly, the initial layer of MoniNet extracts local intra-block correlations among adjacent variables, emphasizing subtle local fluctuations. Then, the higher layer integrates low-level features to learn correlations among all variable blocks, capturing global phase relations in two subspaces representing distinct static and dynamic information. The convolutional operations in these layers are both conducted by Conv-IS 2 FA with physical interpretability in an unsupervised manner. Finally, the monitoring statistics are designed in both local and global layers, and Bayesian inference is employed to produce intuitive results and support decision fusion. Dynamic experiments of three-phase flow show the effectiveness of the proposed MoniNet.}, archive = {J_NEUCOM}, author = {Linghan Li and Shumei Zhang and Feng Dong}, doi = {10.1016/j.neucom.2025.131319}, journal = {Neurocomputing}, month = {11}, pages = {131319}, shortjournal = {Neurocomputing}, title = {Interpretable local-global monitoring network for three-phase flow processes with concurrent analysis of independence and high-order temporal correlation}, volume = {654}, year = {2025}, }
• (2025). Stagewise edge expansion physics-informed neural networks for solving PDEs. NEUCOM, 654, 131318. (www)

  @article{ , abstract = {Physics-Informed Neural Networks (PINNs), as a promising application of neural networks in scientific computing, have shown great potential for solving partial differential equations (PDEs). However, training instability remains a common issue, particularly under globally unified training strategies. Although some prior studies have proposed improvements from the perspective of training strategies, such as curriculum learning and sequential training, to mitigate this problem, the trade-off between solution accuracy and computational efficiency remains a fundamental challenge in solving complex physical problems. To address this, we introduce a dynamic training strategy termed stagewise edge expansion physics-informed neural networks (SEE-PINN). Specifically, SEE-PINN incorporates a gated fusion mechanism into the network architecture to enhance the model's representation capacity by capturing multi-scale features and balancing local and global solution components. During the training process, SEE-PINN initially focuses on the initial and boundary points of the PDE. Once the model achieves a predefined accuracy threshold on these critical regions, the training domain is gradually expanded to include neighboring transitional areas, ensuring high-precision learning in the expanded regions and mitigating the difficulty of handling complex boundary conditions. Subsequently, the remaining configuration points are introduced in batches with weight allocation, enabling the model to progressively accommodate increasing complexity while maintaining stability and accuracy. We validated the performance of SEE-PINN through four numerical experiments. The results demonstrate that SEE-PINN exhibits a stable training process, achieves relatively high solution accuracy, and maintains a competitive level of computational efficiency.}, archive = {J_NEUCOM}, author = {Shuaijie Zhang and Ping Yuan and Mingjiu Pan and Chao Zhang}, doi = {10.1016/j.neucom.2025.131318}, journal = {Neurocomputing}, month = {11}, pages = {131318}, shortjournal = {Neurocomputing}, title = {Stagewise edge expansion physics-informed neural networks for solving PDEs}, volume = {654}, year = {2025}, }
• (2025). Esophageal cancer lesion detection with information preservation for multi-modal diagnostic images. NEUCOM, 654, 131317. (www)

  @article{ , abstract = {Esophageal cancer represents a significant public health concern, necessitating accurate lesion detection via multi-modal endoscopic imaging. However, existing image detection methods often fail to effectively model cross-modal discrepancies and frequently incur substantial information loss, resulting in suboptimal feature representations and degraded performance. To overcome these limitations, we introduce EsophagealDET, a novel model tailored for lesion detection in multi-modal endoscopic images. EsophagealDET is designed to explicitly capture modality-specific characteristics through a Modal Evidence Module, which enhances the model’s capacity to leverage complementary modality information. In addition, we propose a Cross Layer Information Alignment structure to preserve critical features and alleviate information degradation. EsophagealDET was evaluated on both a self-collected multi-modal ECD-Dataset and the publicly available single-modal Kvasir dataset, demonstrating its effectiveness in the medical image detection domain. Extensive ablation studies were conducted to assess the individual contributions of the proposed modules. Experimental results show that EsophagealDET achieves notable improvements in mean average precision (mAP), with gains of 4.09 % on the ECD-Dataset and 1.56 % on the Kvasir dataset, respectively.}, archive = {J_NEUCOM}, author = {Keyu Chen and Zhiyou Yang and Ma Luo and Wei Liu and Yuhang Zhang and Bing Hu and Xianglei Yuan and Wenyu Chen}, doi = {10.1016/j.neucom.2025.131317}, journal = {Neurocomputing}, month = {11}, pages = {131317}, shortjournal = {Neurocomputing}, title = {Esophageal cancer lesion detection with information preservation for multi-modal diagnostic images}, volume = {654}, year = {2025}, }
• (2025). Shuffle-reshuffle gradient mamba for multimodal medical image fusion. NEUCOM, 654, 131316. (www)

  @article{ , abstract = {Multimodal Medical Image Fusion (MMIF) aims to integrate information from different modalities into a single image with comprehensive information and rich spatial details. In recent years, significant progress has been made in MMIF tasks due to advancements in deep neural networks. State Space Models, such as Mamba, have demonstrated a strong capability in modeling long-range dependencies with linear complexity, offering a promising solution to MMIF. However, the fixed sequence scanning mechanism in Mamba may introduce bias, compromising its global receptive field and preventing effective global modeling capabilities. Additionally, Mamba tends to overly focus on global dependency modeling, hindering its effectiveness in representing long-range dependencies while retaining rich fine-grained spatial details, which are critical for MMIF. Furthermore, directly applying Mamba for cross-modal relationship exploration may result in feature redundancy and attenuation of critical information, making it challenging to effectively integrate complementary cross-modal features. To address these issues, we propose the Shuffle-Reshuffle Gradient Mamba (SRGM) tailored to MMIF. Specifically, we design the Local and Global Gradient Mamba (LGGM) to extract modality-specific features while preserving abundant spatial details. Additionally, we introduce the Attention-Guided Cross Mamba (AGCM) to facilitate effective cross-modal feature fusion and reduce feature redundancy. To further enhance fusion performance, we incorporate a Shuffle-Reshuffle Scanning (SRC) strategy into LGGM to overcome the limitations of Mamba’s fixed scanning, achieving unbiased local and global modeling. Through extensive experiments on CT-MRI, PET-MRI, and SPECT-MRI image fusion tasks, our proposed approach surpasses state-of-the-art (SOTA) methods, showcasing superior fusion results.}, archive = {J_NEUCOM}, author = {Lewu Lin and Yufa Duan and Yingying Wang and Jialing Huang and Rongjin Zhuang and Xiaotong Tu and Xinghao Ding and Na Shen and Qing Lu}, doi = {10.1016/j.neucom.2025.131316}, journal = {Neurocomputing}, month = {11}, pages = {131316}, shortjournal = {Neurocomputing}, title = {Shuffle-reshuffle gradient mamba for multimodal medical image fusion}, volume = {654}, year = {2025}, }
• (2025). A novel practically designated-time fuzzy optimal control via reinforcement learning for stochastic constrained nonlinear systems. NEUCOM, 654, 131315. (www)

  @article{ , abstract = {This paper investigates the practically designated time fuzzy optimal control problem for stochastic constrained nonlinear systems with input delay, using reinforcement learning (RL) algorithms. For the uncertain stochastic nonlinear system, a new practically designated time control scheme is proposed and proved to be effective. Meanwhile, a fourth-order asymmetric time-varying barrier Lyapunov function (BLF) is leveraged to mitigate the severe uncertainty caused by the constraints, and Padé approximation and intermediate variables are deployed to counteract the impact of input delay. Additionally, the optimal RL-based instruction filtering control scheme is implemented by exploiting the designated time stability, which not only achieves a prescribed convergence time based on dynamic event triggering, but also ensures that the necessary state constraints are satisfied. The bounded command filtering design avoids the need to prove the continuity of the higher order derivatives of the virtual control signals, thus significantly reducing the complexity of the stability analysis associated with the designated-time convergence of the closed-loop signals. In the optimal control design, the controller put forward ensures that all signals within the closed-loop system remain bounded in probability and the tracking error meets the performance requirements. Finally, the simulation results of two examples verify the effectiveness and applicability.}, archive = {J_NEUCOM}, author = {Lifang Qiu and Junsheng Zhao and Zong-Yao Sun and Chaoxu Mu}, doi = {10.1016/j.neucom.2025.131315}, journal = {Neurocomputing}, month = {11}, pages = {131315}, shortjournal = {Neurocomputing}, title = {A novel practically designated-time fuzzy optimal control via reinforcement learning for stochastic constrained nonlinear systems}, volume = {654}, year = {2025}, }
• (2025). Image augmentation agent for weakly supervised semantic segmentation. NEUCOM, 654, 131314. (www)

  @article{ , abstract = {Weakly Supervised Semantic Segmentation (WSSS), which utilizes only image-level annotations, has gained considerable attention for its efficiency and reduced cost. However, most existing WSSS methods focus on designing new network structures and loss functions to generate more accurate dense labels, overlooking the limitations imposed by fixed datasets, which can constrain performance improvements. We argue that more diverse trainable images provide WSSS with richer information and help model understand more comprehensive semantic patterns. Therefore in this paper, we introduce a novel approach called Image Augmentation Agent (IAA) which shows that it is possible to enhance WSSS from data generation perspective. IAA mainly designs an augmentation agent that leverages large language models (LLMs) and diffusion models to automatically generate additional images for WSSS. In practice, to address the instability in prompt generation by LLMs, we develop a prompt self-refinement mechanism. It allows LLMs to re-evaluate the rationality of generated prompts to produce more coherent prompts. Additionally, we insert an online filter into diffusion generation process to dynamically ensure the quality and balance of generated images. Experimental results show that our method significantly surpasses state-of-the-art WSSS approaches on the PASCAL VOC 2012 and MS COCO 2014 datasets. Our source code will be released.}, archive = {J_NEUCOM}, author = {Wangyu Wu and Xianglin Qiu and Siqi Song and Zhenhong Chen and Xiaowei Huang and Fei Ma and Jimin Xiao}, doi = {10.1016/j.neucom.2025.131314}, journal = {Neurocomputing}, month = {11}, pages = {131314}, shortjournal = {Neurocomputing}, title = {Image augmentation agent for weakly supervised semantic segmentation}, volume = {654}, year = {2025}, }
• (2025). Frequency-sensitive diffusion model for personalized sequential recommendation. NEUCOM, 654, 131313. (www)

  @article{ , abstract = {Sequence recommendation (SR) is crucial for many applications, aiming to predict users’ next interactions based on their history. However, data sparsity and the long-tail problem remain challenges. Some models use data augmentation to address these issues, but often rely on rigid global distributions and embedding variations, neglecting dynamic fluctuations in data. This leads to low-quality content that fails to meet users’ needs. To address this, we propose a novel frequency-aware diffusion model for personalized sequence recommendation (FSDRec). FSDRec constructs an adaptive information-enhanced framework to address dynamic changes in user preferences. It detects shifts in user behavior and integrates both global and local information by selectively retaining high- and low-frequency signals during noise addition, effectively adapting to evolving user needs. A harmonic attention mechanism, based on time intervals, guides the noise restoration in the reverse process. FSDRec also employs a frequency ramp structure to comprehensively capture data features, enhancing prediction accuracy through contrastive learning. Extensive experiments conducted on four benchmark datasets demonstrate that FSDRec achieves superior performance compared to state-of-the-art methods, particularly showing significant improvements of 13.8 % in NDCG@5 on ML-1 M and 7.0 % in HR@10 on the Sports dataset.}, archive = {J_NEUCOM}, author = {Lina Wei and Ruoxin Ni and Jinhui Wei and Yuncheng Jiang}, doi = {10.1016/j.neucom.2025.131313}, journal = {Neurocomputing}, month = {11}, pages = {131313}, shortjournal = {Neurocomputing}, title = {Frequency-sensitive diffusion model for personalized sequential recommendation}, volume = {654}, year = {2025}, }
• (2025). DHS-PE: A more suitable positional encoding for directed hypergraphs. NEUCOM, 654, 131312. (www)

  @article{ , abstract = {Positional encoding is a key skill to capture node structure information in graph neural networks. The existing directed graph encoding and hypergraph encoding have good performance in tasks such as node classification and link prediction. However, these methods still face significant challenges when dealing with directed hypergraphs with complex and high-order interactions. Traditional positional encoding mechanism is difficult to effectively model direction-sensitive topological dependencies, and cannot complete tasks such as node classification well. Although the existing directed hypergraph neural networks method can solve this challenge, it is difficult to effectively model the dynamic semantics of multi-order directional interactions, resulting in limited performance in tasks in sparse topology scenarios. In this paper, we propose a positional encoding framework D irected H ypergraph S tructural-aware P ositional E ncoding ( DHS-PE ) for directed hypergraphs. We provide a random walk based on hyperedge transfer into the original direction propagation to effectively capture the directionality of the directed hypergraph. Moreover, the transition probability between hyperedges is introduced into the spectral decomposition to improve the clusterability perception of the method. When combining different method modules, the transfer matrix is used to dynamically adjust the weight of multi-scale convolution to enhance local pattern capture. Experiments demonstrate that DHS-PE outperforms baseline methods on multiple public datasets, confirming its superiority in modeling high-order relations.}, archive = {J_NEUCOM}, author = {Ruixin He and Yang Shi and Mingxuan Huang and Liangsi Lu and Zhaorong Dai and Yingtong Shi and Jingchao Wang and Zhuoming Xie}, doi = {10.1016/j.neucom.2025.131312}, journal = {Neurocomputing}, month = {11}, pages = {131312}, shortjournal = {Neurocomputing}, title = {DHS-PE: A more suitable positional encoding for directed hypergraphs}, volume = {654}, year = {2025}, }
• (2025). Global attractivity criteria for a discrete-time hopfield neural network model with unbounded delays via singular m—matrices. NEUCOM, 654, 131311. (www)

  @article{ , abstract = {In this work, we establish two global attractivity criteria for a multidimensional discrete-time non-autonomous Hopfield neural network model with infinite delays and delays in the leakage terms. The first criterion, which applies when the activation functions are bounded, is based on M -matrices that are not necessarily invertible. The second criterion, relevant for unbounded activation functions, requires that a related singular M -matrix be irreducible. We contrast our findings with existing results in the literature and present numerical simulations to illustrate the novelty of the proposed criteria.}, archive = {J_NEUCOM}, author = {José J. Oliveira and Ana Sofia Teixeira}, doi = {10.1016/j.neucom.2025.131311}, journal = {Neurocomputing}, month = {11}, pages = {131311}, shortjournal = {Neurocomputing}, title = {Global attractivity criteria for a discrete-time hopfield neural network model with unbounded delays via singular m—matrices}, volume = {654}, year = {2025}, }
• (2025). A generic class-agnostic object counting network with adaptive offset deformable convolution. NEUCOM, 654, 131310. (www)

  @article{ , abstract = {Class-agnostic object counting (CAC) aims at counting the number of objects in the unseen category in an image. Traditional methods often focus on a single setting, such as Few-shot Counting (FSC) and Zero-shot Counting (ZSC), and they struggle to handle objects with varying scales within the same image and fail to utilize self-similarity cues for accurate localization and counting. In this paper, we design a generic class-agnostic object counting network with Adaptive Offset Deformable Convolution (AODC), which initially focuses on the Reference-less class-agnostic object Counting task without any exemplar, and can be easily extended to FSC and ZSC. Our method calculates the self-similarity maps of the image features and performs a 4D convolution on these maps, obtaining the adaptive offsets for the deformable convolution, so that the model can obtain complete information about the object at that location. Through this process, AODC is able to recognize objects of different scales in a same sample. We further extend our approach to both zero-shot setting and few-shot setting, the former with semantic text and the latter with visual exemplars as references. We conduct experiments on the few-shot object counting dataset FSC-147, as well as other large-scale datasets, and show that our method significantly outperforms state-of-the-art approaches in all the three settings. Code is available at https://github.com/WJWu20/AODC .}, archive = {J_NEUCOM}, author = {Wenjun Wu and Yuanwu Xu and Haofeng Zhang}, doi = {10.1016/j.neucom.2025.131310}, journal = {Neurocomputing}, month = {11}, pages = {131310}, shortjournal = {Neurocomputing}, title = {A generic class-agnostic object counting network with adaptive offset deformable convolution}, volume = {654}, year = {2025}, }
• (2025). Short-term wind power forecasting with small-sample datasets using an attention-enhanced domain-adversarial neural network. NEUCOM, 654, 131308. (www)

  @article{ , abstract = {Accurate short-term wind power forecasting is crucial for grid stability but is often challenged by limited data availability at newly built or remote wind farms, hindering state-of-the-art models from capturing complex temporal dynamics. While domain-adversarial neural networks (DANN) offer potential for cross-domain knowledge transfer, conventional frameworks often suffer from insufficient temporal feature extraction capabilities and potential conflicts between domain alignment and prediction tasks. To address these limitations, this paper proposes an enhanced DANN framework specifically tailored for small-sample forecasting. The model incorporates a shared feature extractor based on attention-enhanced temporal convolutional networks (TCNs) to effectively capture both local dependencies and global context inherent in wind power series. Crucially, it employs dual domain-specific regressors alongside the adversarial domain classifier, optimizing predictive performance for each domain while leveraging the gradient reversal layer (GRL) to enforce domain-invariant feature learning. Extensive experimental validation demonstrates that the proposed approach significantly improves prediction accuracy and robustness in data-scarce scenarios by effectively transferring domain-invariant temporal features from data-rich source domains. Comparative analysis with various transfer strategies confirms the superior performance of this integrated adversarial transfer mechanism for small-sample wind power forecasting.}, archive = {J_NEUCOM}, author = {Shang Xiang and Kai Qu and Ying Tian and Shuangsi Xue and Hui Cao}, doi = {10.1016/j.neucom.2025.131308}, journal = {Neurocomputing}, month = {11}, pages = {131308}, shortjournal = {Neurocomputing}, title = {Short-term wind power forecasting with small-sample datasets using an attention-enhanced domain-adversarial neural network}, volume = {654}, year = {2025}, }
• (2025). VSTRD: An end-to-end video spatio-temporal relation detection transformer. NEUCOM, 654, 131307. (www)

  @article{ , abstract = {Video Visual Relation Detection (VidVRD) is crucial for understanding dynamic object interactions in videos. However, current methods struggle with error accumulation, computational redundancy, and limited capabilities in modeling long-time series within multi-stage frameworks. In this paper, we propose an end-to-end transformer-based framework, VSTRD ( V ideo S patio- T emporal R elation D etection Transformer). We innovatively design a decoupled query generation mechanism to decompose the query into object prototype embedding, temporal position coding, and shared space prior coding, explicitly modeling cross-frame consistency to improve trajectory association accuracy. We design the Spatio-Temporal Relation Aggregation Decoder to alleviate the problem of relationship breakage in long segments. Experiments on the ImageNet-VidVRD and VidOR datasets show that VSTRD achieves competitive performance compared to recent methods. Notably, on the VidOR dataset, our model outperforms state-of-the-art methods in the relation detection subtask, with improvements of 0.41 % in R@50 and 0.09 % in R@100, respectively.}, archive = {J_NEUCOM}, author = {Ruyi Chang and Haopeng Wang and Zhitian Zhang and Dejiao Huang and Shuai Guo}, doi = {10.1016/j.neucom.2025.131307}, journal = {Neurocomputing}, month = {11}, pages = {131307}, shortjournal = {Neurocomputing}, title = {VSTRD: An end-to-end video spatio-temporal relation detection transformer}, volume = {654}, year = {2025}, }
• (2025). F-ADMD: Fast anomaly detection via multi-scale guided denoising diffusion model. NEUCOM, 654, 131305. (www)

  @article{ , abstract = {Visual anomaly detection has become a significant solution in industrial production due to its remarkable effectiveness and efficiency. However, traditional generative model-based methods are constrained in overall performance due to limitations in reconstruction quality and speed. Recently, generative approaches based on diffusion models have brought new opportunities to visual tasks, owing to their high-quality and diverse generation capabilities. Therefore, this paper proposes a novel anomaly detection framework based on diffusion models, termed F-ADMD, which is primarily composed of two parts: image reconstruction and defect segmentation. In the image reconstruction portion, we introduce a multi-scale guided rapid reconstruction method that is not only significantly faster than traditional diffusion models but also effectively handles the reconstruction of various anomaly-type regions. In the defect segmentation section, we integrate a Transformer attention mechanism that can capture both local and global context information at different feature levels, aiding the model in focusing more effectively on key regions of the image and thus improving the accuracy of identifying the boundaries of anomalous areas and enhancing segmentation precision. F-ADMD achieves state-of-the-art performance in image-level detection and anomaly localization on the challenging and widely used MVTec dataset, demonstrating the framework's effectiveness and broad applicability. This new method not only enhances detection accuracy but also significantly boosts processing speed, providing a novel comprehensive solution for visual anomaly detection in industrial production.}, archive = {J_NEUCOM}, author = {Congying An and Jingjing Wu and Huanlong Zhang}, doi = {10.1016/j.neucom.2025.131305}, journal = {Neurocomputing}, month = {11}, pages = {131305}, shortjournal = {Neurocomputing}, title = {F-ADMD: Fast anomaly detection via multi-scale guided denoising diffusion model}, volume = {654}, year = {2025}, }
• (2025). SAGM-net: Spatio-temporal action graph modeling network for weakly-supervised temporal action localization. NEUCOM, 654, 131304. (www)

  @article{ , abstract = {Weakly-supervised temporal action localization (WS-TAL) is a challenging task in computer vision that aims to identify and locate specific action instances in untrimmed videos with only video-level annotations. To achieve this task, some existing methods build upon multiple instance learning (MIL) frameworks or design feature enhancement modules based on pre-trained extracted features. However, due to the insufficient exploration of snippet-level contextual relationships, these methods still suffer from imprecise temporal localization. Furthermore, some approaches leverage the advantage of graph convolution networks in relational modeling to address these issues. Nonetheless, they overlook the differences in spatio-temporal relation graphs generated by different modalities, confusing some action snippets with ambiguous information. To this end, we propose a novel weakly-supervised temporal action localization method, namely spatio-temporal action graph modeling network (SAGM-Net). On the one hand, we design a semantic-to-structural graph modeling (SSGM) module to refine feature representation by leveraging temporal dependencies and semantic discriminability of different modalities. On the other hand, we introduce a separation-aware feature enhancement (SFE) module, which enhances the semantic representation capability of features and maximizes the separation between pseudo-action and pseudo-background features. Experiments conducted on THUMOS14 and ActivityNet v1.2 demonstrate that SAGM-Net outperforms most WS-TAL methods, achieving an average mAP of 45.9 % on THUMOS14 and 26.5 % on ActivityNet v1.2, with performance comparable to the AICL method (AAAI 2023) on THUMOS14 and the LPR method (TCSVT 2024) on ActivityNet v1.2.}, archive = {J_NEUCOM}, author = {Yutong Li and Miao Ma and Chao Yao and Zhao Pei and Jie Wu}, doi = {10.1016/j.neucom.2025.131304}, journal = {Neurocomputing}, month = {11}, pages = {131304}, shortjournal = {Neurocomputing}, title = {SAGM-net: Spatio-temporal action graph modeling network for weakly-supervised temporal action localization}, volume = {654}, year = {2025}, }
• (2025). Visual speech recognition using attention-enhanced ResNet and hybrid recurrent-transformer encoder with quantity-ratio curriculum learning for low-resource languages. NEUCOM, 654, 131303. (www)

  @article{ , abstract = {People with hearing impairments use visual cues through lipreading to get a better understanding of their communication. While many studies focus on lipreading research for English and Mandarin, there are still limited studies that address lipreading research for low-resource languages. Therefore, we establish a novel framework based on deep learning to solve the problem, which consists of several key components. First, it includes an optimized preprocessing stage and additional data augmentation for the lipreading task. Second, it incorporates a Spatial-Temporal Feature Extractor (STFE) to extract better lip movement features. Our STFE uses a channel and spatial attention module incorporated with ResNet-18 as an attention-enhanced ResNet and a combination of BiGRU with a multi-stage residual transformer encoder. Lastly, the model integrates a word classification module to make the final prediction. Moreover, we train the model using the quantity-ratio curriculum learning strategy by starting with easier categories and incrementally increasing difficulty. It maintains constant exposure to easier samples while gradually introducing more difficult ones until reaching the final stage with a defined quantity. In addition, we also present an expanded lipreading dataset for the Indonesian language (ID-LR), which is one of the low-resource languages. ID-LR contains more than 105,000 videos with 230-word classes spoken by diverse persons in natural conditions. With this configuration and learning strategy, our approach reaches remarkable performance in terms of accuracy and becomes state-of-the-art in multiple low-resource languages of lipreading datasets (Turkish, Romanian, German, and Indonesian) based on our presented benchmarks.}, archive = {J_NEUCOM}, author = {Griffani Megiyanto Rahmatullah and Shanq-Jang Ruan and Lieber Po-Hung Li}, doi = {10.1016/j.neucom.2025.131303}, journal = {Neurocomputing}, month = {11}, pages = {131303}, shortjournal = {Neurocomputing}, title = {Visual speech recognition using attention-enhanced ResNet and hybrid recurrent-transformer encoder with quantity-ratio curriculum learning for low-resource languages}, volume = {654}, year = {2025}, }
• (2025). PACER: A fully push-forward-based distributional reinforcement learning algorithm. NEUCOM, 654, 131301. (www)

  @article{ , abstract = {In this paper, we propose the first fully push-forward-based distributional reinforcement learning algorithm, named PACER, which consists of a distributional critic, a stochastic actor and a sample-based encourager. Specifically, the push-forward operator is leveraged in both the critic and actor to model the return distributions and stochastic policies respectively, enabling them with equal modeling capability and thus enhancing the synergetic performance. Since it is infeasible to obtain the density function of the push-forward policies, novel sample-based regularizers are integrated into the encourager to incentivize efficient exploration and alleviate the risk of trapping into local optima. Moreover, a sample-based stochastic utility value policy gradient is established for the push-forward policy update, which circumvents the explicit demand for the policy density function in existing REINFORCE-based stochastic policy gradients. As a result, PACER fully utilizes the modeling capability of the push-forward operator and is able to explore a broader class of the policy spaces compared with limited policy classes used in existing distributional actor-critic algorithms (i.e. Gaussians). We validate the critical role of each component in our algorithm with extensive empirical studies. Experimental results demonstrate the superiority of our algorithm over state-of-the-art methods, achieving an average score improvement of 10 % on the Mujoco continuous control benchmark.}, archive = {J_NEUCOM}, author = {Wensong Bai and Chao Zhang and Yichao Fu and Peilin Zhao and Hui Qian}, doi = {10.1016/j.neucom.2025.131301}, journal = {Neurocomputing}, month = {11}, pages = {131301}, shortjournal = {Neurocomputing}, title = {PACER: A fully push-forward-based distributional reinforcement learning algorithm}, volume = {654}, year = {2025}, }
• (2025). GaitSTAGCN: Spatial-temporal attention graph convolutional networks for gait recognition. NEUCOM, 654, 131300. (www)

  @article{ , abstract = {Gait recognition is an important biometric identification technology, which is widely applied in fields such as security and medical care. Due to its rich temporal and spatial information, skeletal data has become a research hotspot in the field of gait recognition. However, under complex conditions such as occlusion, changes in viewing angles, and outdoor environments, the recognition accuracy of existing models still needs to be improved. To address this issue, this paper proposes a model named GaitSTAGCN, aiming to enhance the recognition ability of the model under complex conditions and thus improve its generalization performance. GaitSTAGCN contains two key modules: the Enhanced Graph Attention Integration Module (EGAIM) and the Multi-scale Temporal Graph Attention Integration Module (MTGAIM). These two modules share the Graph Attention Integration Module (GAIM). By strengthening the topological modeling of key points, they can simultaneously capture the long-distance correlations and local structural features, thus enhancing the ability to analyze complex spatial patterns. On this basis, the EGAIM module captures more complex feature distributions and richer fine-grained information by enhancing the efficiency of cross-channel feature fusion. The MTGAIM module, through analyzing data on multiple time scales, can identify short-term and long-term dynamic changes in gait data, ensuring a comprehensive understanding of temporal information and further improving the ability to capture and analyze the temporal characteristics of gait. Experiments show that GaitSTAGCN achieves an accuracy of 93.91 % in the cross-view scenarios of CASIA-B, representing an improvement of 1.91 % compared to GaitTR. Its precision of 86.96 % under the single-view condition with coat occlusion highlights its robustness against occlusion. It achieves an accuracy of 58.46 % on the GREW outdoor dataset, which is 1.42 % higher than that of GPGait. However, on the extremely complex Gait3D dataset, the model has only achieved a basic performance of 11.6 %. This indicates that when dealing with such highly complex data, the model still has a long way to go to significantly improve its recognition ability and generalization performance. It requires further optimization and improvement in the future. Nevertheless, the model has demonstrated certain advantages and application potential in other specific complex scenarios.}, archive = {J_NEUCOM}, author = {AoFei Wang and ZhenJie Hou and En Lin and Xing Li and JiuZhen Liang and XinWen Zhou}, doi = {10.1016/j.neucom.2025.131300}, journal = {Neurocomputing}, month = {11}, pages = {131300}, shortjournal = {Neurocomputing}, title = {GaitSTAGCN: Spatial-temporal attention graph convolutional networks for gait recognition}, volume = {654}, year = {2025}, }
• (2025). Age-at-death estimation from 3D bone models of the pubic symphysis by using a deep multi-view learning approach. NEUCOM, 654, 131299. (www)

  @article{ , abstract = {Forensic anthropology is a sub-field of physical anthropology that analyzes human skeletal remains of medical and legal interest. Within it, estimating an individual’s biological profile plays a crucial role for forensic anthropologists and pathologists in identifying victims and documenting evidence. In particular, the estimation of the age-at-death of an individual is crucial to reduce the range of possible coincidences during the human identification process. Most existing approaches rely on the forensic expert’s subjective evaluation and do not consider relevant characteristics that are not detectable by traditional visual methods. This paper proposes a deep multi-view learning approach to automatically estimate an individual’s age-at-death from three 2D panoramic views generated from a 3D bone model of their pubic symphysis. These three images correspond to the three principal coordinate axes and consist of 3-channel images generated from the spatial distribution map, normals’ deviation map, and gradient normals’ deviation map projections. Our approach not only detects high-level characteristics of the pubic symphysis undetectable by the human eye, but also reduces the age estimation error to 6.74, becoming, as far as we know, the best result in the state-of-the-art. In addition, the proposed approach includes a post-hoc explanation stage that visualizes, through a 3D saliency map of the pubic symphysis, the most relevant pixel areas on which the proposed neural architecture relies to make a given decision. This provides forensic scientists with valuable information and confidence to support their decision-making process.}, archive = {J_NEUCOM}, author = {Alejandro Manzanares and Marilyn Bello and Fernando Navarro and Javier Irurita and Inmaculada Alemán and Pablo Mesejo and Óscar Cordón}, doi = {10.1016/j.neucom.2025.131299}, journal = {Neurocomputing}, month = {11}, pages = {131299}, shortjournal = {Neurocomputing}, title = {Age-at-death estimation from 3D bone models of the pubic symphysis by using a deep multi-view learning approach}, volume = {654}, year = {2025}, }
• (2025). Convergence of batch gradient descent with learnable smooth masks for pruning feedforward neural networks. NEUCOM, 654, 131298. (www)

  @article{ , abstract = {Designing a near-optimal feed-forward neural network (FNN) architecture for a specific task remains a difficult challenge in the field of deep learning. To overcome this problem, we propose an effective pruning method that uses a specific smooth function, called a smooth mask, to mask and identify unnecessary network parameters. The original idea was to estimate the significance of each weight or neuron using binary variables. This leads to an NP-hard combinatorial multi-objective optimization problem. To address this issue, we used learnable smooth masks to approximate the binary nature of variables. Furthermore, we reformulate the problem as a composition of unsupervised classification problems. The batch gradient method is used to solve the problem by optimizing the proposed regularized error function. This error function is a weighted sum of the empirical training error and three proposed regularization terms, which are crucial for controlling and optimizing important weights and hidden neurons. After training, the mask functions determine the class of each weight and hidden node, allowing us to identify and prune the unimportant ones. Additionally, we provide theoretical results that demonstrate the convergence of our proposed method. Numerical analysis on a range of benchmark datasets and various FNN architectures supports the theoretical findings and demonstrates the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Mohamed Quasdane and Hassan Ramchoun and Tawfik Masrour}, doi = {10.1016/j.neucom.2025.131298}, journal = {Neurocomputing}, month = {11}, pages = {131298}, shortjournal = {Neurocomputing}, title = {Convergence of batch gradient descent with learnable smooth masks for pruning feedforward neural networks}, volume = {654}, year = {2025}, }
• (2025). PHLA: A pre-trained model based method incorporating HLA-peptide binding and immunogenicity. NEUCOM, 654, 131292. (www)

  @article{ , abstract = {The discovery of neoantigen peptides plays a vital role in tumor immunotherapy. Currently, the prediction of neoantigen peptides mainly considers the binding relationship between human leukocyte antigen (HLA) and peptides. However, by ignoring more comprehensive biological considerations, some results may not be sufficiently reliable in tumor immunotherapy. In this study, we applied natural language processing encoding technology and deep learning techniques for prediction, considering both the HLA-peptide binding probability (binding model) and the immunogenicity of the peptide-HLA complex (immune model). Experimental results show that the binding model of the peptide-HLA complex achieves better performance than some benchmarks on two external datasets. Meanwhile, the immunogenicity model of the peptide-HLA complex significantly improves the prediction performance. Finally, we perform data analysis on the prediction results and discover a close correlation with tumors, suggesting that the prediction results of the peptide-HLA complex may provide substantial assistance for tumor vaccine design.}, archive = {J_NEUCOM}, author = {Jixiang Fan and Jianzhang Zhang and Ying Huang and Xiu-Xiu Zhan and Huajun Hu and Chuang Liu}, doi = {10.1016/j.neucom.2025.131292}, journal = {Neurocomputing}, month = {11}, pages = {131292}, shortjournal = {Neurocomputing}, title = {PHLA: A pre-trained model based method incorporating HLA-peptide binding and immunogenicity}, volume = {654}, year = {2025}, }
• (2025). Fixed/preassigned-time synchronization of delayed fuzzy inertial memristive neural networks with parameter disturbance via quantized fault-tolerant control. NEUCOM, 654, 131289. (www)

  @article{ , abstract = {This article investigates fixed-time synchronization (FXTS) and preassigned-time synchronization (PATS) for fuzzy inertial memristive neural networks (FIMNNs) with parameter disturbance and discontinuous activation functions. Different from the existing research, by directly designing effective Lyapunov functionals and novel quantized fault-tolerant controllers, some sufficient criteria are established for ensuring FXTS and PATS through algebraic inequalities. Particularly, the controller devised in this paper enjoys greater universality, with its exponential powers being mutually independent. It can reduce the consumption of communication and cope with the problems of time-varying actuator faults more effectively. In the end, one instance is given to verify the efficacy of the synchronization conditions obtained.}, archive = {J_NEUCOM}, author = {Xin Huang and Shiming Chen and Zheng Zhang}, doi = {10.1016/j.neucom.2025.131289}, journal = {Neurocomputing}, month = {11}, pages = {131289}, shortjournal = {Neurocomputing}, title = {Fixed/preassigned-time synchronization of delayed fuzzy inertial memristive neural networks with parameter disturbance via quantized fault-tolerant control}, volume = {654}, year = {2025}, }
• (2025). Zero-shot load forecasting for integrated energy systems: A large language model-based framework with multi-task learning. NEUCOM, 654, 131288. (www)

  @article{ , abstract = {The growing penetration of renewable energy sources in power systems has increased the complexity and uncertainty of load forecasting, especially for integrated energy systems with multiple energy carriers. Traditional forecasting methods heavily rely on historical data and demonstrate limited transferability across different scenarios, creating significant challenges for emerging applications in smart grids and energy internet. This paper proposes the TSLLM-Load Forecasting Mechanism, a novel zero-shot load forecasting framework based on large language models (LLMs) to address these challenges. The framework consists of three key components: a data preprocessing module that handles multi-source energy load data, a time series prompt generation module that bridges the semantic gap between energy data and LLMs through multi-task learning and similarity alignment, and a prediction module that leverages pre-trained LLMs for accurate forecasting. We validate the framework’s effectiveness on real-world datasets comprising load profiles from Australian solar-powered households. In conventional testing, our method achieves competitive performance, ranking third among nine methods. In zero-shot prediction, our framework demonstrates 10.8 % MSE improvement and 12.5 % MAE improvement compared to Informer, ranking second overall after TimeMixer. Most significantly, few-shot learning experiments reveal exceptional capabilities under extreme data constraints, with our method achieving optimal performance when trained with only 1 % of available data, representing 40.8 % MSE improvement compared to the best conventional method and 78.9 % improvement compared to existing LLM-based approaches. Large-scale transferability analysis shows our model outperforms baselines for 82 % of households when trained with minimal data from a single household. These results demonstrate the framework’s potential for accurate and transferable load forecasting in integrated energy systems, particularly beneficial for renewable energy integration and smart grid applications where historical data availability is limited.}, archive = {J_NEUCOM}, author = {Jiaheng Li and Donghe Li and Ye Yang and Huan Xi and Wanju Yu and Yu Xiao and Qingyu Yang}, doi = {10.1016/j.neucom.2025.131288}, journal = {Neurocomputing}, month = {11}, pages = {131288}, shortjournal = {Neurocomputing}, title = {Zero-shot load forecasting for integrated energy systems: A large language model-based framework with multi-task learning}, volume = {654}, year = {2025}, }
• (2025). MPSD-net: A novel multiple progressive self-distillation network for semantic segmentation in adverse weather. NEUCOM, 654, 131286. (www)

  @article{ , abstract = {Semantic segmentation is a core technology for environmental perception in autonomous driving. Although numerous studies have focused on semantic segmentation under adverse weather conditions, existing methods suffer from two key limitations: (i) reliance on datasets trained under normal weather conditions (i.e., source domain data), posing risks of data privacy leakage; and (ii) a predominant focus on single-stage knowledge transfer, neglecting the “coarse-to-fine” progressive learning paradigm. To address these challenges, we propose a novel Multiple Progressive Self-Distillation Network (MPSD-Net), which hierarchically transfers knowledge from normal to adverse weather conditions through coarse-to-fine distillation. Firstly, to achieve coarse-grained distillation, we propose the novel Intra-Inter Self-Decoupled Distillation (IISDdist) based on an intra-inter coupling mechanism. This approach aligns features through dense matching while capturing intra-class and inter-class relationships to enforce directional consistency, thereby realizing spatially decoupled coarse-grained distillation. Secondly, to ensure strong consistency at the decision level, we propose Fine-Grained Logits Distillation (FGLdist) based on a knowledge response mechanism. This approach generates logit masks through uncertainty modeling and enforces multiple strict constraints to guarantee precise fine-grained distillation. Finally, to enhance global knowledge learning, we propose Cross-Stage Adaptive Distillation (CSAdist) based on dynamic feature mapping. This method constructs a feature mapping space to transfer low-frequency global structural knowledge from deep to shallow features. The experimental results show that MPSD-Net achieves competitive performance on two public datasets, ACDC and RobotCar.}, archive = {J_NEUCOM}, author = {Chong Chen and Hui Liu and Chunsheng Liu and Faliang Chang and Tuo Li and Penghui Hao}, doi = {10.1016/j.neucom.2025.131286}, journal = {Neurocomputing}, month = {11}, pages = {131286}, shortjournal = {Neurocomputing}, title = {MPSD-net: A novel multiple progressive self-distillation network for semantic segmentation in adverse weather}, volume = {654}, year = {2025}, }
• (2025). Enhancing echo state network with reservoir state selection for time series forecasting. NEUCOM, 654, 131283. (www)

  @article{ , abstract = {Echo state networks (ESNs) have been widely adopted for time series forecasting due to their remarkable efficacy. The standard approach for training ESN-based prediction models involves computing output weights through mappings between sequential reservoir states and target values. During this process, an initial period of reservoir states is typically discarded to eliminate the influence of zero initialization. However, this washout strategy leads to the loss of potentially valuable information contained in the discarded initial states, while simultaneously preserving redundant representations within the retained reservoir states. To address this limitation, we propose a novel particle swarm Bayesian optimization (PSBO) approach to yield an optimized ESN (PSBO-ESN) for time series forecasting, enabling the ESN model to adaptively identify and retain the more salient and indicative reservoir states. In PSBO-ESN, the solution to the reservoir state selection problem is formulated as a binary vector. The particle swarm optimization (PSO) algorithm is leveraged to facilitate global exploration of the solution space, while a particle-specific Bayesian optimization (BO) is designed to perform the local search, thereby strengthening the exploitation ability of our proposed method. Comprehensive experiments are conducted on two simulated and two real-world datasets. The results demonstrate that PSBO-ESN outperforms all baseline methods and significantly enhances the ESN models for time series forecasting.}, archive = {J_NEUCOM}, author = {Qi Sima and Yukun Bao and Xinze Zhang and Kun He and Xin Lai}, doi = {10.1016/j.neucom.2025.131283}, journal = {Neurocomputing}, month = {11}, pages = {131283}, shortjournal = {Neurocomputing}, title = {Enhancing echo state network with reservoir state selection for time series forecasting}, volume = {654}, year = {2025}, }
• (2025). Si-CA MobileNet: A lightweight and efficient convolutional neural network for distracted driver detection. NEUCOM, 654, 131281. (www)

  @article{ , abstract = {In recent years, the significant increase in traffic accidents caused by distracted driving has drawn widespread attention. As a result, identifying distracted drivers has become essential for improving road safety and advancing intelligent driver assistance systems. Deep learning has been used for real-time driver monitoring to detect risks like distraction, fatigue, and unsafe behaviors. To deploy CNNs on mobile devices, it is important to reduce number of parameters and computation while keeping detection accuracy high. In this paper we propose a lightweight convolutional neural network for distracted driver detection named Si-CA MobileNet. The model incorporates the Si-Block module based on the Simple, Parameter-free Attention Mechanism (SimAM) and the CA-Block module based on the Coordinate Attention (CA) mechanism. Experimental results show that Si-CA MobileNet achieves Top-1 accuracy of 99.80 % and 92.35 % on the StateFarm and 100-Driver datasets, respectively, surpassing most existing methods. Additionally, the model has only 3.17 M parameters and 175.95 M FLOPs. It also supports real-time processing at 82.95 FPS, making it well-suited for lightweight deployment scenarios.}, archive = {J_NEUCOM}, author = {Minghuan Lv and Yingjian Liu and Zonghai Zha and Xiangyun Zheng and Hao Wang and Yindong Wen and Zhongwen Guo}, doi = {10.1016/j.neucom.2025.131281}, journal = {Neurocomputing}, month = {11}, pages = {131281}, shortjournal = {Neurocomputing}, title = {Si-CA MobileNet: A lightweight and efficient convolutional neural network for distracted driver detection}, volume = {654}, year = {2025}, }
• (2025). Adaptive imbalanced node classification graph contrastive learning. NEUCOM, 654, 131280. (www)

  @article{ , abstract = {Graph Contrastive Learning (GCL) is a powerful self-supervised technique for learning node and graph representations. However, real-world graph data often exhibit imbalanced class distributions, which pose significant challenges to GCL’s effectiveness. Our experiments show that current state-of-the-art (SOTA) methods perform poorly under imbalanced settings. To address this, we propose a novel GCL framework called AIGCL for imbalanced node classification. This framework automatically and adaptively balances the node representations learned by GCL. Specifically, we introduce a new data augmentation method that retains more information from minority class nodes during graph augmentation. Additionally, we use an imbalance rate adaptive sampling strategy to balance the data. We also incorporate a Variational Graph Autoencoder (VGAE) with an encoder–decoder structure to pretrain the data and generate high-quality pseudo-labels. Our experiments demonstrate that under imbalanced settings, our model improves classification accuracy by 4 %-12 % compared to baseline models, significantly enhancing the performance of minority class nodes.}, archive = {J_NEUCOM}, author = {Xinyu Hu and Donghai Guan and Weiwei Yuan and Qi Zhu and Çetin Kaya Koç}, doi = {10.1016/j.neucom.2025.131280}, journal = {Neurocomputing}, month = {11}, pages = {131280}, shortjournal = {Neurocomputing}, title = {Adaptive imbalanced node classification graph contrastive learning}, volume = {654}, year = {2025}, }
• (2025). D2Fusion: Dual-AE and latent-guided diffusion network for infrared and visible image fusion. NEUCOM, 654, 131279. (www)

  @article{ , abstract = {Infrared and visible image (IVI) fusion retains the advantages of both modalities and is widely used in advanced vision tasks such as object detection. To improve the quality and visualization of fused images, a dual auto-encoder (dual-AE) and latent-guided diffusion fusion network (D2Fusion) for IVI is proposed. Firstly, D2Fusion uses a dual-branch residual AE to separately extract features from IVI, addressing the challenges of feature extraction discrepancies between the two modalities. Secondly, this paper introduces a dynamic cross-attention fusion network (DCAFNet) to deeply explore the spatial correlations between IVI features, enabling higher-quality fusion. Thirdly, the fused features are used as condition to guide the diffusion models, enhancing the fusion effect while improving cross-scene generalization. Finally, a multi-level dense-connected decoder (MDD) is proposed to generate the fused image. MDD allows deeper global information to complement shallow features better through dense connections, thereby enhancing the feature representation capability of the image. Extensive experiments on several public datasets show that our method achieves superior fusion performance and generates fused images better than existing state-of-the-art methods. Specifically, D2Fusion achieved EN (6.9131), SD (43.0563), and SSIM (0.9921) on the MSRS dataset.}, archive = {J_NEUCOM}, author = {Shizun Sun and Junwei Xu and Shuo Han and Jie Zhao and Bo Mo}, doi = {10.1016/j.neucom.2025.131279}, journal = {Neurocomputing}, month = {11}, pages = {131279}, shortjournal = {Neurocomputing}, title = {D2Fusion: Dual-AE and latent-guided diffusion network for infrared and visible image fusion}, volume = {654}, year = {2025}, }
• (2025). Coexistence of non-ideal threshold logic units and multiplicative input noise in DNN-kWTA networks. NEUCOM, 654, 131278. (www)

  @article{ , abstract = {While the classical dual neural network k -winner-take-all (DNN- k WTA) model demonstrates desirable properties under ideal conditions, we examine its limitations in non-ideal scenarios, specifically with non-ideal threshold logic units (TLUs) and multiplicative input noise. We consider two types of multiplicative noise: Gaussian and non-Gaussian distributions. To address these challenges, we propose an analogous dynamic model that captures the behavior of the DNN- k WTA network under these non-ideal conditions. Our analysis provides valuable insights into how TLUs and multiplicative noise affect the functional correctness of the model. Additionally, we develop two practical methods that allow us to estimate the performance of a non-ideal DNN- k WTA network without resorting to time-consuming simulations. These methods empower us to assess the likelihood of functional correctness. This assessment facilitates making informed decisions regarding the necessary levels of realization quality aligned with desired performance targets.}, archive = {J_NEUCOM}, author = {Jiajie Mai and Eric Wong and Chi-Sing Leung}, doi = {10.1016/j.neucom.2025.131278}, journal = {Neurocomputing}, month = {11}, pages = {131278}, shortjournal = {Neurocomputing}, title = {Coexistence of non-ideal threshold logic units and multiplicative input noise in DNN-kWTA networks}, volume = {654}, year = {2025}, }
• (2025). Efficient large margin adversarial training based on decision boundaries for adversarial robustness. NEUCOM, 654, 131277. (www)

  @article{ , abstract = {Recent advances in adversarial training have been proposed to improve adversarial robustness by increasing feature margins, where robustness strongly correlates with improved input margins. However, we identify a fundamental limitation of these approaches in this work: increasing feature margins does not guarantee better input margins due to distance distortions in DNNs’ non-linear hierarchical structures, leading to limited robustness improvement. To address this challenge, we propose an E fficient L arge M argin A dversarial T raining method based on D ecision B oundaries ( ELM-DB-AT ). It reformulates input margin optimization as feature margin optimization by leveraging adversarial examples that minimally cross decision boundaries. Specifically, ELM-DB-AT first extracts input-margin-related features with pre-crafted misclassified examples that marginally cross the decision boundary. Then, it optimizes these features to achieve better adversarial robustness. This process establishes a crucial connection between feature margin optimization and input margin optimization. Furthermore, ELM-DB-AT regulates cross-entropy loss by combining three metrics (including intra-class feature compactness, inter-class feature separability, and feature margin), saving computational and time consumption. Extensive experiments demonstrate the effectiveness of our approach. For example, compared with the baseline methods using ResNet50 on CIFAR10, ELM-DB-AT improves robust accuracy and clean accuracy by up to 5.81 % and 5.18 %, respectively. Additional experiments on the ImageNette dataset further validate the superior capability of the proposed method.}, archive = {J_NEUCOM}, author = {Mengjun Xu and Ziqiang Li and Lei Liu and Bin Li}, doi = {10.1016/j.neucom.2025.131277}, journal = {Neurocomputing}, month = {11}, pages = {131277}, shortjournal = {Neurocomputing}, title = {Efficient large margin adversarial training based on decision boundaries for adversarial robustness}, volume = {654}, year = {2025}, }
• (2025). Large-capacity and robust video watermarking via DWT coefficient separation/reconstruction and multi-scale spatiotemporal fusion. NEUCOM, 654, 131275. (www)

  @article{ , abstract = {Deep learning-based video watermarking has been demonstrated to be effective in enhancing visual quality and robustness. However, a common limitation of existing methods is their embedding capacity, typically constrained to a specific number of bits. Increasing embedding capacity often results in visual quality distortion in the watermarked video and a decrease in the robustness of the watermarking network. To address these issues, we propose a large-capacity and robust video watermarking network (LCR-VWN) based on Discrete Wavelet Transform (DWT) coefficient separation/reconstruction (DCSR) and multi-scale spatiotemporal fusion. To reduce visual quality distortion caused by large embedding capacities, Invertible Neural Network (INN) is employed. Additionally, a multi-scale spatiotemporal fusion structure is designed and integrated with INN, aiming to capture features of distant video frames, and minimize visual quality degradation during large-capacity embedding process. To bolster robustness further, we devise a DCSR structure for separation and reconstruction of DWT coefficients. It is utilized to separate and reconstruct high and low-frequency coefficients of DWT-transformed videos in specific proportions, and strategically selects optimal embedding positions for watermark information. Moreover, we introduce a Simulated Compression Composite Network (SCCN), which integrates neural networks and noise layer to effectively simulate video compression process. Experimental results show that LCR-VWN exhibits superior robustness and visual quality compared to state-of-the-art algorithms, offering an effective solution for copyright protection.}, archive = {J_NEUCOM}, author = {Hanbin Sun and Zhenzhen Zhang and Zichen Li and Jiaoyun Liu}, doi = {10.1016/j.neucom.2025.131275}, journal = {Neurocomputing}, month = {11}, pages = {131275}, shortjournal = {Neurocomputing}, title = {Large-capacity and robust video watermarking via DWT coefficient separation/reconstruction and multi-scale spatiotemporal fusion}, volume = {654}, year = {2025}, }
• (2025). Analysis and adaptive control for robust h∞ synchronization of complex coupled parabolic networks. NEUCOM, 654, 131274. (www)

  @article{ , abstract = {This paper delves into the robust H ∞ synchronization problem for a type of complex coupled parabolic networks (CPNs). Considering the time delay and spatial diffusion that might exist in the nodal dynamics, a complex CPN with delayed state coupling and spatial diffusion coupling is established by semi-linear parabolic partial differential equations (PDEs). An H ∞ synchronization criterion is derived for the considered complex CPNs utilizing the Green’s formula, the boundary conditions and the gradient operations. Furthermore, a distributed adaptive control scheme is presented to guarantee the H ∞ synchronization for CPNs with delayed state and spatial diffusion couplings. Finally, two numerical examples are utilized to validate the correctness of the obtained control criteria and to illustrate the advantages of the presented approaches.}, archive = {J_NEUCOM}, author = {Chenguang Liu and Jin-Liang Wang and Qing Gao and Jianbin Qiu}, doi = {10.1016/j.neucom.2025.131274}, journal = {Neurocomputing}, month = {11}, pages = {131274}, shortjournal = {Neurocomputing}, title = {Analysis and adaptive control for robust h∞ synchronization of complex coupled parabolic networks}, volume = {654}, year = {2025}, }
• (2025). Rank-constrained correspondence network for few-shot segmentation. NEUCOM, 654, 131273. (www)

  @article{ , abstract = {Few-shot segmentation (FSS) is a key technology in computer vision, aiming to segment images from unseen classes using only a few labeled examples. The main challenge lies in extracting pixel-wise relationships between the support and query images and effectively transferring label information from the support images to the query images. Through theoretical analysis, we have found that the correlation matrix in FSS is inherently low-rank, due to the limited shared representative structure between the support and query images. However, existing methods often fail to explicitly model this low-rank property, leading to redundant or noisy correspondences, thereby reducing segmentation performance. To address this issue, we propose a novel Rank-Constrained Correspondence Network (RCCNet), which imposes a low-rank constraint on the correlation matrix and explicitly utilizes its low-rank prior, effectively filtering out irrelevant relationships, enhancing the robustness of correspondences between the support and query images, and ultimately improving segmentation accuracy. Extensive experiments on datasets such as PASCAL-5 i , COCO-20 i , and FSS-1000 demonstrate that our method significantly outperforms the baseline, achieving competitive results compared to state-of-the-art methods, validating the effectiveness and importance of the low-rank constraint in few-shot segmentation.}, archive = {J_NEUCOM}, author = {Xiaoliu Luo and Ningsheng Liao and Huiqing Qi}, doi = {10.1016/j.neucom.2025.131273}, journal = {Neurocomputing}, month = {11}, pages = {131273}, shortjournal = {Neurocomputing}, title = {Rank-constrained correspondence network for few-shot segmentation}, volume = {654}, year = {2025}, }
• (2025). Exponential synchronization of variable-order fractional spatio-temporal networks via impulsive control: A direct error technique. NEUCOM, 654, 131272. (www)

  @article{ , abstract = {This study is devoted to exponential synchronization of variable-order fractional spatio-temporal networks by means of impulsive control and a direct analysis technique. Firstly, a rigorous mathematical derivation is developed for the translation of the controlled variable-order fractional models into impulsive partial differential systems, which is nontrivial for variable-order fractional systems and improves existing results. Secondly, as theoretical preparation, a differential inequality with impulse about a non-negative function and an inequality about Mittag-Leffler function with single-parameter are established. Besides, some synchronization criteria dependent on variable fractional order and impulse information are established by introducing a direct error technique, which overcomes the potential difficulty caused by the unknown or undetermined synchronous state. Two numerical results are presented to validate the theoretical research.}, archive = {J_NEUCOM}, author = {Lingbi Tang and Cheng Hu and Juan Yu}, doi = {10.1016/j.neucom.2025.131272}, journal = {Neurocomputing}, month = {11}, pages = {131272}, shortjournal = {Neurocomputing}, title = {Exponential synchronization of variable-order fractional spatio-temporal networks via impulsive control: A direct error technique}, volume = {654}, year = {2025}, }
• (2025). Secure synchronization of stochastic neural networks under deception attacks: An event-based intermittent impulsive approach. NEUCOM, 654, 131271. (www)

  @article{ , abstract = {This paper investigates the secure synchronization of stochastic neural networks (SNNs) under deception attacks by proposing an event-based intermittent impulsive control (EBIIC) strategy. The EBIIC approach incorporates the benefits of both impulsive control and event-based intermittent control (EBIC). Unlike existing EBIC methods, which necessitate continuous control signals over the control interval, our EBIIC strategy leverages discrete impulse signals within the control interval, thus significantly reducing control costs. Additionally, we propose a periodic EBIIC strategy to eliminate the need for continuous monitoring of the triggering condition. Sufficient conditions are provided for secure synchronization under deception attacks, revealing the interplay among deception attack probability, control width, impulsive gain, and event parameters. Finally, we illustrate the effectiveness of the EBIIC strategy via a Chua’s circuit example.}, archive = {J_NEUCOM}, author = {Xiaotao Zhou and Jieqing Tan and Lulu Li and Yangang Yao and Huihui Zhang}, doi = {10.1016/j.neucom.2025.131271}, journal = {Neurocomputing}, month = {11}, pages = {131271}, shortjournal = {Neurocomputing}, title = {Secure synchronization of stochastic neural networks under deception attacks: An event-based intermittent impulsive approach}, volume = {654}, year = {2025}, }
• (2025). TiM4Rec: An efficient sequential recommendation model based on time-aware structured state space duality model. NEUCOM, 654, 131270. (www)

  @article{ , abstract = {The modeling paradigm of sequential recommendation is shifting from Transformer to Mamba, including Mamba1, based on the State Space Model (SSM), and Mamba2, based on State Space Duality (SSD). While SSD achieves higher computational efficiency, it faces performance degradation in low-dimensional scenarios which is crucial for sequential recommendation tasks. Considering that time-aware enhancement methods are commonly employed to mitigate performance loss, our analysis reveals that the performance decline of SSD can similarly be fundamentally compensated by leveraging mechanisms in time-aware methods. To overcome challenges in adapting existing time-aware methods like TiSASRec, such as compatibility with SSD and computational inefficiencies in time-difference modeling, we propose Ti me-Aware M amba for Rec ommendation (TiM4Rec), which integrates a novel Time-aware Structured Masked Matrix into SSD to enhance performance while improving efficiency. Extensive experiments on four real-world datasets demonstrate its superiority, marking the first time-aware method tailored to Mamba for sequential recommendation. The code for our model is accessible at https://github.com/AlwaysFHao/TiM4Rec .}, archive = {J_NEUCOM}, author = {Hao Fan and Mengyi Zhu and Yanrong Hu and Hailin Feng and Zhijie He and Hongjiu Liu and Qingyang Liu}, doi = {10.1016/j.neucom.2025.131270}, journal = {Neurocomputing}, month = {11}, pages = {131270}, shortjournal = {Neurocomputing}, title = {TiM4Rec: An efficient sequential recommendation model based on time-aware structured state space duality model}, volume = {654}, year = {2025}, }
• (2025). Context-dependent neuro-symbolic AI through self-supervised learning with large language models. NEUCOM, 654, 131269. (www)

  @article{ , abstract = {This study proposes methods to enable neuro-symbolic artificial intelligence (AI) to learn context dependency through self-supervised learning with large language models (LLMs). While neuro-symbolic AI can utilize extensive and ambiguous knowledge bases to prove theorems, LLMs encounter significant challenges in this domain. However, conventional neuro-symbolic AI faces practical issues, as the search space for reasoning expands rapidly with the increasing size of the knowledge base. Therefore, this study utilizes self-supervised learning with LLMs to incorporate context dependency into neuro-symbolic AI, thereby reducing the search space for reasoning. Existing approaches cannot achieve this because they fail to learn context dependency. The proposed network is the first to combine neuro-symbolic AI and LLMs to reduce the search space for reasoning. The authors trained and evaluated the proposed network using a novel reasoning dataset with context and developed an evaluation metric to validate the effectiveness of search space reduction. Results demonstrate that the proposed network significantly reduces the search space while maintaining accuracy compared to existing baseline methods. This network provides a solution to the frame problem and has potential applications in laboratory automation and autonomous robotics, where reasoning with a large search space is required.}, archive = {J_NEUCOM}, author = {Hiroshi Honda and Masafumi Hagiwara}, doi = {10.1016/j.neucom.2025.131269}, journal = {Neurocomputing}, month = {11}, pages = {131269}, shortjournal = {Neurocomputing}, title = {Context-dependent neuro-symbolic AI through self-supervised learning with large language models}, volume = {654}, year = {2025}, }
• (2025). TopiCOT: Neural topic model aligning with pre-trained clustering and optimal transport. NEUCOM, 654, 131268. (www)

  @article{ , abstract = {Recent research on VAE-based neural topic models has focused on enhancing the encoder network by incorporating pre-trained language models (PLMs) and refining topic–word relationships within the generative process. Despite these improvements, the integration of PLMs often results in increased inference costs, and document-topic distributions can still exhibit suboptimal representation. Additionally, existing neural topic models have not addressed the topic–cluster relationships. In this study, we present TopiCOT (Neural Topi c Model Aligning with Pre-trained C lustering and O ptimal T ransport), a novel VAE-based topic model designed to overcome these limitations. TopiCOT effectively bridges the gap between the document clustering capabilities of PLMs and the core topic model, avoiding the need for direct PLMs integration. Moreover, we model the correlation between topics and pre-trained clusters through the Optimal Transport (OT) problem, which also enhances document representation and efficiently captures topic associations. Experimental results on popular benchmark datasets demonstrate that our method effectively improves document-topic distributions while preserving a high level of topic coherence comparable to other state-of-the-art baselines. Notably, our approach boosts inference speed by about 600 times compared to UTopic, a leading VAE-based method that leverages pre-trained language models.}, archive = {J_NEUCOM}, author = {Tung Nguyen and Duy-Tung Pham and Quang Duc Nguyen and Linh Ngo Van and Anh Nguyen Duc and Sang Dinh Viet}, doi = {10.1016/j.neucom.2025.131268}, journal = {Neurocomputing}, month = {11}, pages = {131268}, shortjournal = {Neurocomputing}, title = {TopiCOT: Neural topic model aligning with pre-trained clustering and optimal transport}, volume = {654}, year = {2025}, }
• (2025). Strategic evolution: Adaptive peak replacement and fused response techniques for dynamic constrained multimodal optimization. NEUCOM, 654, 131267. (www)

  @article{ , abstract = {This paper addresses the challenges posed by Dynamic Constrained Multimodal Problems (DCMMOPs), characterized by the integration of constraints, dynamics, and multimodality. We introduce an innovative adaptive neighborhood strategy known as Dynamic Neighborhood Adaptive Peak Replacement (DNAPR) to enhance the local search capability of the population. DNAPR strategically generates individuals in proximity to the current optimal solution and integrates dominant individuals to form the next generation. Additionally, we employ an adaptive strategy based on the mean difference between superior individuals of parent and offspring generations to prevent local optimum issues. Extending the multi-population concept, we enhance the traditional Memory-based Immigrants (MI) dynamic response strategy, introducing a Maximum Extension Distance (MED) to preserve population diversity amidst environmental changes. Finally, we propose a new algorithm framework, DMMCSA-DNAPR, integrating DNAPR with the Dynamic Multimodal Clonal Selection Algorithm (DMMCSA). Experimental results demonstrate the superior performance of DMMCSA-DNAPR across various accuracies and problem instances, establishing its efficacy in solving DCMMOPs.}, archive = {J_NEUCOM}, author = {Yong Wang and Xiao-Wei Wang and Hongyan Sang and Gai-Ge Wang and Swagatam Das}, doi = {10.1016/j.neucom.2025.131267}, journal = {Neurocomputing}, month = {11}, pages = {131267}, shortjournal = {Neurocomputing}, title = {Strategic evolution: Adaptive peak replacement and fused response techniques for dynamic constrained multimodal optimization}, volume = {654}, year = {2025}, }
• (2025). Latent attribute augmented network for few-shot class-incremental learning. NEUCOM, 654, 131266. (www)

  @article{ , abstract = {Few-Shot Class-Incremental Learning (FSCIL) aims to learn knowledge about new classes continually with limited labeled data, where preserving knowledge about old classes and fast adaptation to new classes play important roles in FSCIL models. Traditional approaches usually learn discriminative global features among base classes and the feature extraction model is directly applied to novel categories. However, the global features, though effective for base class discrimination, are less generalized to novel categories. Therefore, we focus on the more generalizable local parts and propose the Latent Attribute Augmented Network (LAAN) to enhance the discriminative local features. Specifically, we automatically learn some latent attributes that are shared among different classes and utilize them to focus on the key local regions of the images. Furthermore, we construct a transformer-based knowledge interaction module, to fuse the information of latent attributes and the local features, which obtains more discriminative features to classify different classes. Our method has achieved superior performance across three benchmark datasets: CIFAR100, mini-ImageNet, and CUB200.}, archive = {J_NEUCOM}, author = {Yongli Hu and Jiasen Zhang and Huajie Jiang and Baocai Yin}, doi = {10.1016/j.neucom.2025.131266}, journal = {Neurocomputing}, month = {11}, pages = {131266}, shortjournal = {Neurocomputing}, title = {Latent attribute augmented network for few-shot class-incremental learning}, volume = {654}, year = {2025}, }
• (2025). SALA: Semantic alignment and localization alignment for visual grounding. NEUCOM, 654, 131265. (www)

  @article{ , abstract = {Visual grounding focuses on establishing fine-grained alignment between specific regions and query expressions, which is increasingly essential as a cornerstone of visual intelligence. Despite recent success, existing methods often struggle with two main issues. Firstly, using independently pre-trained uni-modal encoders to extract expressive feature embeddings leads to a significant semantic gap between uni-modal features, hindering the effective interaction of visual-linguistic contexts. Secondly, the supervision provided by box annotations is inherently sparse and often underexploited, which limits the model’s ability to capture the fine-grained visual cues necessary to distinguish referent objects from the background, thereby leading to localization ambiguity. In this paper, we propose a Semantic Alignment and Localization Alignment (SALA) framework for visual grounding, which effectively bridges the cross- and uni-modal semantic gap and improves localization performance through patch-level and pixel-level alignment. This contributes to enhancing the consistency of representation before multimodal fusion, thereby improving the localization performance. Extensive experiments show that the proposed method outperforms state-of-the-art methods on five widely used datasets. Codes will be made publicly available after acceptance.}, archive = {J_NEUCOM}, author = {Hongbing Li and Xinran Wang and Linyi Yang and Qi Li and Bo Xiao}, doi = {10.1016/j.neucom.2025.131265}, journal = {Neurocomputing}, month = {11}, pages = {131265}, shortjournal = {Neurocomputing}, title = {SALA: Semantic alignment and localization alignment for visual grounding}, volume = {654}, year = {2025}, }
• (2025). ImagePrompt: Image-enhanced prompt-tuning for multimodal relation extraction. NEUCOM, 654, 131264. (www)

  @article{ , abstract = {Multimodal Relation Extraction (MRE), which leverages images as auxiliary information to identify relationships between entities in social media texts, plays a vital role in information extraction and knowledge base construction. However, existing methods typically rely on pre-trained models to encode and fuse cross-modal features but often fail to fully utilize the potential of these models. To overcome this limitation, we propose the Image-Enhanced Prompt-Tuning Method (ImagePrompt) for MRE. Our approach addresses two key challenges in prompt construction: noisy visual information and the inherent modality gap between images and text. To mitigate these issues, we extract task-relevant objects and textual content from images using object detection (OD) and optical character recognition (OCR), convert them into textual form, and integrate them into prompt templates. Through carefully designed prompt templates for the MRE task, our method better aligns with the objectives of pre-trained language models while more effectively leveraging useful information from images. Extensive experiments on the MNRE dataset demonstrate the effectiveness of our approach.}, archive = {J_NEUCOM}, author = {Ruohan Zhang and Yanli Hu and Teng Liu and Jianwei Cao}, doi = {10.1016/j.neucom.2025.131264}, journal = {Neurocomputing}, month = {11}, pages = {131264}, shortjournal = {Neurocomputing}, title = {ImagePrompt: Image-enhanced prompt-tuning for multimodal relation extraction}, volume = {654}, year = {2025}, }
• (2025). Network collaborator: Knowledge transfer between network reconstruction and community detection. NEUCOM, 654, 131263. (www)

  @article{ , abstract = {The exact network structures and community partitions are valuable tools for modeling and analyzing complex systems. However, in complex natural systems, the network and community structures between entities are often unobservable. Instead, we may observe interdependent dynamics within these systems. While various approaches have been proposed to address network reconstruction (NR) from dynamics and community detection (CD) from networks independently, none of them explicitly considers the potential benefits of sharing knowledge across these two tasks. NR and CD from dynamics are natural synergistic tasks that motivate the proposed evolutionary multitasking NR and CD framework, called Network Collaborator . The NR task explicitly shares the network structure with the CD task, while the CD task explicitly transfers the community partitions to assist the NR task. Moreover, to share knowledge from the NR task to the CD task, Network Collaborator models the process of CD from dynamics to find communities in the dynamic network and then considers whether to transfer knowledge across tasks. To verify the performance of Network Collaborator , we design a test suite of multitasking NR and CD problems based on various synthetic and real-world networks. The experimental results have demonstrated that joint NR with CD has a synergistic effect. Specifically, community partitions can be used to improve the accuracy of network reconstruction. Conversely, the network structure also better assists in finding community partitions. The code can be accessed at https://github.com/xiaofangxd/EMTNRCD .}, archive = {J_NEUCOM}, author = {Tao Ma and Chao Wang and Kai Wu and Junyuan Chen and Jing Liu}, doi = {10.1016/j.neucom.2025.131263}, journal = {Neurocomputing}, month = {11}, pages = {131263}, shortjournal = {Neurocomputing}, title = {Network collaborator: Knowledge transfer between network reconstruction and community detection}, volume = {654}, year = {2025}, }
• (2025). Text-guided visual prompt learning with semantic prompt generation and feature fusion. NEUCOM, 654, 131253. (www)

  @article{ , abstract = {Recent advancements in prompt learning have significantly improved the flexibility and generalization of vision-language models (VLMs) in visual tasks. However, existing visual prompt learning methods often emphasize irrelevant visual details during the visual feature extraction process, which hinders semantic understanding and makes it difficult to achieve fine-grained feature capture. To address this, we propose Semantic Prompt Generation and Feature Fusion (SPGFF) , a novel method that dynamically generates visual prompts by integrating image–text interactions. SPGFF introduces two key modules: the Text-Guided Prompt Generation (TGPG) module, which leverages large language models to produce fine-grained textual descriptions and combines them with visual features to create semantically relevant prompts, and the Text-Guided Feature Fusion (TGFF) module, which employs self-attention and cross-attention mechanisms to enhance local-global feature integration under textual guidance. Extensive experiments on 11 benchmark datasets, including base-to-new generalization, cross-dataset transfer, and few-shot tasks, demonstrate SPGFF’s superiority. It achieves state-of-the-art average performance (79.69 % harmonic mean) in base-to-new settings, surpassing prior methods by effectively capturing task-critical details (e.g., texture, shape) and improving interpretability. These advancements highlight the critical role of semantic-textual guidance in refining visual representations, establishing a unified framework for high-precision cross-modal alignment while preserving the generalization capabilities of pre-trained VLMs.}, archive = {J_NEUCOM}, author = {Jiajia Zhang and Shengsheng Wang}, doi = {10.1016/j.neucom.2025.131253}, journal = {Neurocomputing}, month = {11}, pages = {131253}, shortjournal = {Neurocomputing}, title = {Text-guided visual prompt learning with semantic prompt generation and feature fusion}, volume = {654}, year = {2025}, }
• (2025). Dynamic obstacle avoidance for car-like mobile robots based on neurodynamic optimization with control barrier functions. NEUCOM, 654, 131252. (www)

  @article{ , abstract = {This paper studies the problem of dynamic obstacle avoidance for car-like mobile robots with physical limitations, including bounded steering angles and input saturation. By designing control barrier functions (CBFs), a novel collision-free module is proposed, which applies minimal adjustments to the velocity commands of a given controller when necessary to ensure dynamic obstacle avoidance motion under the physical limitations. Specifically, the module is implemented as a quadratic program that is both subject to constraints derived from the CBFs and solved in real-time through neurodynamic optimization. Compared with the existing results, the proposed scheme ensures real-time generation of physically realizable safe commands and enhances steering maneuverability for obstacle avoidance. Finally, the safety of the designed method is theoretically proven, and its effectiveness in dynamic obstacle avoidance is verified via numerical simulations.}, archive = {J_NEUCOM}, author = {Zheng Zhang and Guang-Hong Yang}, doi = {10.1016/j.neucom.2025.131252}, journal = {Neurocomputing}, month = {11}, pages = {131252}, shortjournal = {Neurocomputing}, title = {Dynamic obstacle avoidance for car-like mobile robots based on neurodynamic optimization with control barrier functions}, volume = {654}, year = {2025}, }
• (2025). End-to-end food condition system via wavelet attention enhanced network. NEUCOM, 654, 131248. (www)

  @article{ , abstract = {Automatic vegetable quality analysis systems can be useful in optimizing production processes and increasing harvest efficiency and effectiveness. For this reason, in this paper, we propose an end-to-end system that can be easily used in practical applications in production halls or even in mobile applications. The model operates on data obtained from the camera, which takes a photo of the vegetable and sends it to the convolutional neural network module. To improve classification quality, wavelet attention was proposed for better analysis of features during the processing of data in the network. For this purpose, we defined low and high frequencies that allow the creation of a score matrix showing important information. A proposed network allows for the recognition of objects with various textures and focuses on local and global feature extraction. Additionally, a system architecture that automates the addition of classes in the network to build an end-to-end solution was described. The modeled framework was tested on three publicly available datasets - PlantVillage, ATLDSD and Indigenous. The obtained results allowed for achieving higher evaluation metrics than the state-of-the-art.}, archive = {J_NEUCOM}, author = {Antoni Jaszcz and Dawid Połap}, doi = {10.1016/j.neucom.2025.131248}, journal = {Neurocomputing}, month = {11}, pages = {131248}, shortjournal = {Neurocomputing}, title = {End-to-end food condition system via wavelet attention enhanced network}, volume = {654}, year = {2025}, }
• (2025). Predictor-assisted evolutionary neural architecture search for spiking neural networks. NEUCOM, 654, 131244. (www)

  @article{ , abstract = {Spiking neural networks (SNNs) are a biologically inspired computing paradigm that transmit information through discrete sequences of spikes and have good biological interpretability and energy efficiency. The performance of SNNs primarily depends on their architectures. However, current SNN architectures are mainly designed manually, significantly rely on expert experience and lack flexibility. In this research, we propose the PESNN, a predictor-assisted evolutionary neural architecture search for SNNs, to automatically design SNN architectures in which the entire algorithm follows the evolutionary neural architecture search (ENAS) framework. Unlike the existing search methods for SNN architectures, which still involve some manual design, we propose a variable-length coding strategy based on spatial channel adaptation for fully automatic architecture design. Since the fitness evaluation stage in the search is very computationally expensive, we design a performance predictor to assist this stage to speed up the search process. Moreover, a multi-objective optimization method is introduced to reduce the number of spikes and power consumption. Compared with other methods, the experimental results with the static datasets (CIFAR-10, CIFAR-100) and neuromorphic dataset (DVS-CIFAR-10) demonstrate that the PESNN can find the optimal network architecture with high classification accuracy and fewer spikes in fewer timesteps.}, archive = {J_NEUCOM}, author = {Yuting Wang and Yu Xue and Wei Ding and Yiyu Tan and Peng Chen and Mohamed Wahib}, doi = {10.1016/j.neucom.2025.131244}, journal = {Neurocomputing}, month = {11}, pages = {131244}, shortjournal = {Neurocomputing}, title = {Predictor-assisted evolutionary neural architecture search for spiking neural networks}, volume = {654}, year = {2025}, }
• (2025). MADFlow: Multimodal difference compensation flow for multimodal anomaly detection. NEUCOM, 654, 131243. (www)

  @article{ , abstract = {Normalizing flow-based methods have been widely studied in image anomaly detection and have proven their effectiveness. However, relying solely on 2D images makes it difficult to capture anomalies of objects from different perspectives. To address this issue, we propose MADFlow , where a C ross-modal D ifference C ompensation F usion ( CDCF ) module is designed to utilize 3D information and avoid excessive domain gaps between multimodal features. Firstly, we perform consistent learning from features of different modalities by using a specific loss function. Based on this, the residual differences between the features of each modality are used to compensate for the missing 3D information in the 2D RGB data. In addition, we propose a F requency- S pace E nhancement ( FSE ) module, which models features from both frequency and space perspectives, and fuses them adaptively through a new gating mechanism, thus offering advantages over previous methods in modeling only spatial features. Finally, extensive experiments demonstrate the competitive results of the proposed method on two commonly used multimodal anomaly detection datasets, MVTec 3D-AD and Eyecandies.}, archive = {J_NEUCOM}, author = {Yao Li and Xinyuan Zhou and Shiyong Lan and Wenwu Wang and Xin Lai and Yixin Qiao}, doi = {10.1016/j.neucom.2025.131243}, journal = {Neurocomputing}, month = {11}, pages = {131243}, shortjournal = {Neurocomputing}, title = {MADFlow: Multimodal difference compensation flow for multimodal anomaly detection}, volume = {654}, year = {2025}, }
• (2025). Out-of-distribution detection: Sparsification meets subspace. NEUCOM, 654, 131240. (www)

  @article{ , abstract = {Large machine learning models are becoming prevalent in this era and ensuring their real-world deployment and production can be challenging due to the closed-world training setting and various unpredictable scenarios. Out-of-Distribution (OOD) detection has been proposed to address this dilemma and has become a hot and rapidly advancing research area. In this paper, we propose ASSL (Activation Sparsification meets Subspace Learning), an extremely simple and effective post-hoc OOD method that enhances pre-trained models’ outputs through activation adjustment to simplify feature representations. Our approach formulates a discriminative OOD score integrating both class-specific logits and class-agnostic information from feature subspaces derived from the modified model outputs. Experiments on a wide range of OOD benchmarks across different scales of ID (In-Distribution) datasets including ImageNet and CIFAR demonstrate superior detection performance as indicated by FPR95 and AUROC of our method. Additional ablation studies validate ASSL’s capability of maintaining competitive original ID classification accuracy which is equally vital for OOD detection.}, archive = {J_NEUCOM}, author = {Xun Zhang and Yang Liu and Jungong Han}, doi = {10.1016/j.neucom.2025.131240}, journal = {Neurocomputing}, month = {11}, pages = {131240}, shortjournal = {Neurocomputing}, title = {Out-of-distribution detection: Sparsification meets subspace}, volume = {654}, year = {2025}, }
• (2025). QIDLearningLib: A python library for quasi-identifier recognition and evaluation. NEUCOM, 654, 131239. (www)

  @article{ , abstract = {Quasi-identifiers (QIDs) are attributes in a dataset that are not directly unique identifiers of the users/entities themselves but can be used, often in conjunction with other datasets or information, to identify individuals and thus present a privacy risk in data sharing and analysis. Identifying QIDs is important in developing proper strategies for anonymization and data sanitization. This paper proposes QIDLearningLib , a Python library that offers a set of metrics and tools to measure the qualities of QIDs and identify them in data sets. It incorporates metrics from different domains – causality, privacy, data utility, and performance – to offer a holistic assessment of the properties of attributes in a given tabular dataset. Furthermore, QIDLearningLib offers visual analysis tools to present how these metrics shift over a dataset and implements an extensible framework that employs multiple optimization algorithms such as an evolutionary algorithm, simulated annealing, and greedy search using these metrics to identify a meaningful set of QIDs.}, archive = {J_NEUCOM}, author = {Sancho Amaral Simões and João P. Vilela and Miriam Seoane Santos and Pedro Henriques Abreu}, doi = {10.1016/j.neucom.2025.131239}, journal = {Neurocomputing}, month = {11}, pages = {131239}, shortjournal = {Neurocomputing}, title = {QIDLearningLib: A python library for quasi-identifier recognition and evaluation}, volume = {654}, year = {2025}, }
• (2025). Knowledge graph and large language model integration with focus on educational applications: A survey. NEUCOM, 654, 131230. (www)

  @article{ , abstract = {In recent years, artificial intelligence (AI) technology has made significant advancements, particularly in the areas of large language models (LLMs) and knowledge graphs (KGs). KGs excel at structured knowledge representation and reasoning, offering interpretability; however, they are costly to construct, have limited coverage, and lack natural language processing capabilities. Conversely, LLMs possess powerful language understanding and generation abilities, but they rely heavily on vast amounts of data, are prone to “hallucinations," and lack interpretability. The integration of these two approaches is an inevitable trend for achieving stronger and more reliable AI applications, and has become a hot topic of research. Simultaneously, the combination of LLMs and KGs perfectly aligns with the pressing needs of the education field for precise reasoning and personalized services, addressing the shortcomings of traditional teaching methods and providing support for intelligent education. In light of this, this paper undertakes work in the following three key areas. Firstly, the concepts and technologies of both LLMs and KGs, along with their applications in education, are introduced. On this basis, the paper then delves into a discussion of the methods for integrating LLMs and KGs, and reviews related research progress. Finally, the paper focuses on specific educational scenarios, such as intelligent tutoring systems, intelligent learning companions, and intelligent evaluation systems, to explore the collaborative application of LLMs and KGs. The aim of this paper is to provide researchers in the field with a systematic understanding of LLMs and KGs, and to offer valuable references for future AI-driven educational innovation.}, archive = {J_NEUCOM}, author = {Guanyu Chen and Tao Song and Quanyu Wang and Zheng Ma and Jun Hu and Qi Li and Chunming Wu}, doi = {10.1016/j.neucom.2025.131230}, journal = {Neurocomputing}, month = {11}, pages = {131230}, shortjournal = {Neurocomputing}, title = {Knowledge graph and large language model integration with focus on educational applications: A survey}, volume = {654}, year = {2025}, }
• (2025). Topology and semantic contrastive learning with joint distribution alignment for link prediction. NEUCOM, 654, 131228. (www)

  @article{ , abstract = {Graph contrastive learning (GCL) has gained significant attention for its superior performance in unsupervised node/graph representation learning. While excelling in node-level and graph-level classification tasks, GCL often underperforms in link prediction compared to other unsupervised methods. Current GCL approaches primarily rely on stochastic augmentation (e.g., edge masking, node dropping) and emphasize correlated node similarity across views in the embedding space, neglecting crucial topological and semantic graph information. This limitation hinders their potential for link prediction. Moreover, stochastic augmentation may introduce distributional discrepancies in node representations, challenging consistent graph learning. To address these issues, we propose a novel framework, T opology and S emantic C ontrastive L earning with Joint Distribution A lignment (TSCLA). TSCLA innovatively integrates topological and semantic relationships between nodes into GCL framework while aligning the representations of masked and original graphs at the topological and semantic distributions. Specifically, we first assign weights to positive nodes with two tailored metrics, i.e., topological mutual information and semantic soft clustering similarity. In this way, both topological and semantic information are well integrated into GCL. Furthermore, the targeted decoding strategies are designed to reconstruct the masked information, which helps the model capture topological and semantic information. Finally, we propose a joint distribution alignment module to strengthen the consistency between the corresponding node representations at the topological and semantic levels. Extensive experiments on 7 public datasets demonstrate that the proposed TSCLA achieves significantly better performance on most downstream tasks, especially for link prediction, compared with the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Gangzhe Zhang and Jun Huang and Shuzhen Rao}, doi = {10.1016/j.neucom.2025.131228}, journal = {Neurocomputing}, month = {11}, pages = {131228}, shortjournal = {Neurocomputing}, title = {Topology and semantic contrastive learning with joint distribution alignment for link prediction}, volume = {654}, year = {2025}, }
• (2025). Towards dynamic self-training for scalable semi-supervised learning on graphs. NEUCOM, 654, 131223. (www)

  @article{ , abstract = {In the realm of graph-based semi-supervised learning (GSSL), traditional methodologies often struggle to effectively handle labeled samples and scale to accommodate large datasets. To increase supervision information in semi-supervised learning, the self-training paradigm is often used, mainly in datasets with moderate sizes. On the other hand, the use of anchors was adopted with large datasets. In this research endeavor, we propose a novel framework for GSSL that leverages a novel self-training principle tailored for very large datasets, and introduces an advanced method for automatic graph construction using anchors. Our approach focuses on utilizing generated labels of random batches of unlabeled samples, subsequently incorporating these predictions into the training set to enhance the model’s accuracy. Pseudo-labeling, a specific instance of self-training, assigns pseudo-labels to the most confidently predicted unlabeled examples, treating them as ground truth during the training phase. By constructing anchor-to-anchor affinity graphs that incorporate both feature and label information, our method facilitates robust learning on large-scale datasets. Through comprehensive experimentation across diverse large datasets, our approach demonstrates its efficacy in achieving scalable and reliable semi-supervised learning outcomes. These findings represent a significant advancement in the field of GSSL, with wide-ranging implications for various applications across different domains. Our method not only addresses the scalability issue but also ensures the effective integration of both labeled and pseudo labeled data, thereby enhancing the overall learning process.}, archive = {J_NEUCOM}, author = {F. Dornaika and Z. Ibrahim and J. Charafeddine and A. Bosaghzadeh}, doi = {10.1016/j.neucom.2025.131223}, journal = {Neurocomputing}, month = {11}, pages = {131223}, shortjournal = {Neurocomputing}, title = {Towards dynamic self-training for scalable semi-supervised learning on graphs}, volume = {654}, year = {2025}, }
• (2025). Learning stabilizable symplectic ODE-net-based MPC for autonomous vehicle trajectory tracking. NEUCOM, 654, 131222. (www)

  @article{ , abstract = {Neural network (NN)-based Model predictive control (MPC) can model vehicle dynamics with uncertainty and provide higher tracking accuracy for autonomous vehicle trajectory tracking tasks. However, the existing NN-based MPCs employ fully connected layers with non-interpretable structures, leading to suboptimal modeling accuracy. Additionally, these methods often incur a high computational burden, and theoretical properties such as feasibility and stability are lost. To address these issues, we propose a deep stabilizing Symplectic ODE-Net (SymODEN)-based MPC approach for trajectory tracking. Firstly, the SymODEN is applied to learn the vehicle lateral dynamics and a SymODEN-based MPC is established, which achieves higher tracking accuracy. Then a deep stabilizing controller for MPC is proposed. By incorporating constraints into the control network and including the error bound between the learning system and the true system in the loss function, we conduct a stability analysis and identify the subset of the region of attraction. This subset is proven to be a subset of the feasible region of the learning MPC. Finally, the CarSim-Matlab co-simulations verify that our method significantly improves tracking performance compared to the existing works. Tracking error and computational time are reduced by at most 99.7 % and 99.8 % respectively in different manoeuvres.}, archive = {J_NEUCOM}, author = {Hengheng Gong and Tijin Yan and Huahui Xie and Runze Gao and Yufeng Zhan and Yuanqing Xia}, doi = {10.1016/j.neucom.2025.131222}, journal = {Neurocomputing}, month = {11}, pages = {131222}, shortjournal = {Neurocomputing}, title = {Learning stabilizable symplectic ODE-net-based MPC for autonomous vehicle trajectory tracking}, volume = {654}, year = {2025}, }
• (2025). Contextual information cross-feature transformer for image denoising applications. NEUCOM, 654, 131220. (www)

  @article{ , abstract = {It is significant to effectively restore detailed information under noise degradation using prior knowledge in real image denoising applications. To address the performance bottleneck of traditional convolutional neural networks caused by fixed weights and local receptive field limitations and the high computational complexity of the Transformer in global modeling, this paper proposes a contextual information cross-feature Transformer (CICFormer) for image denoising applications. Specifically, an effective receptive field module is designed to dynamically capture contextual information, significantly reducing computational overhead while enhancing spatial perception. A convolutional attention block is then constructed to combine local feature extraction with global dependency modeling, improving the robustness of the denoiser in complex scenes. Additionally, a cross-feature fusion mechanism is proposed, which further enhances the ability to retain image details through refined interactions of multi-scale features. Comprehensive quantitative and qualitative experiments demonstrate that this proposed method achieves a good balance between denoising accuracy and computational efficiency, outperforming many existing image denoising models on multiple benchmark datasets.}, archive = {J_NEUCOM}, author = {Yuxuan Hu and Shichao Zhang}, doi = {10.1016/j.neucom.2025.131220}, journal = {Neurocomputing}, month = {11}, pages = {131220}, shortjournal = {Neurocomputing}, title = {Contextual information cross-feature transformer for image denoising applications}, volume = {654}, year = {2025}, }
• (2025). Adaptive layer and token selection for efficient fine-tuning of vision transformers. NEUCOM, 654, 131216. (www)

  @article{ , abstract = {Foundation models for computer vision built on Vision Transformer (ViT) architectures have become increasingly widespread. However, their fine-tuning process is resource-intensive, slowing their adoption in edge or low-energy applications. We introduce ALaST ( Adaptive Layer Selection for ViT Fine-Tuning ), a novel approach that dynamically optimizes the fine-tuning process to significantly reduce computational cost, memory consumption, and training time. Our method is founded on the critical observation that during fine-tuning, the importance of individual layers and tokens varies substantially across training iterations and depends on the specific mini-batch being processed. ALaST leverages this insight by adaptively estimating layer importance at each fine-tuning step and allocating computational resources—or “compute budgets”—proportionally. Layers assigned lower budgets are either trained with a reduced token set or temporarily frozen. Through comprehensive empirical evaluation on standard benchmarks, we demonstrate that ALaST achieves substantial efficiency gains: up to 1.3 × reduction in training time, 1.5 × reduction in FLOPs, and 2 × decrease in memory requirements, all while maintaining model performance within 0.5 % of full fine-tuning. Notably, our approach provides an automatic schedule for distributing computational resources across layers and can be combined with existing parameter-efficient fine-tuning techniques, offering an orthogonal dimension of optimization for Vision Transformers.}, archive = {J_NEUCOM}, author = {Alessio Devoto and Federico Alvetreti and Jary Pomponi and Paolo Di Lorenzo and Pasquale Minervini and Simone Scardapane}, doi = {10.1016/j.neucom.2025.131216}, journal = {Neurocomputing}, month = {11}, pages = {131216}, shortjournal = {Neurocomputing}, title = {Adaptive layer and token selection for efficient fine-tuning of vision transformers}, volume = {654}, year = {2025}, }
• (2025). SAFCN: Self-enhancing attention fusion contrastive network for multimodal sentiment analysis. NEUCOM, 654, 131212. (www)

  @article{ , abstract = {Multimodal sentiment analysis (MSA) aims to analyze users’ sentiment information through heterogeneous data modalities. While contemporary MSA methodologies predominantly employ transformer-based architectures and related techniques for sentiment prediction, they neglect a comprehensive exploration of both unimodal feature refinement and cross-modal representation synthesis. To address these limitations, we propose the Self-Enhancing Attention Fusion Contrastive Network (SAFCN), a novel framework integrating hierarchical attention mechanisms with contrastive learning paradigms to achieve enhanced multimodal fusion. Specifically, three self-augmented encoders independently encode visual, textual, and auditory features, thereby generating optimized unimodal representations. Subsequently, a Cross-Modal Reinforced Fusion (CRF) module facilitates inter-modal interactions to derive enriched cross-modal representations. These representations are fused via a variety of fusion methods, with multimodal fusion further assisted by contrastive learning. Experimental results on three public datasets demonstrate the effectiveness of the proposed method. On the CH-SIMS dataset, SAFCN outperforms current state-of-the-art methods, achieving improvements in Acc-2, F1, Acc-3, MAE, and Corr of 0.74 % , 0.31 % , 0.7 % , 0.013 , and 0.041 , respectively. On CMU-MOSI and CMU-MOSEI, SAFCN delivers state-of-the-art comparable performance, with notable advancements: on CMU-MOSI, it surpasses previous methods by 3.8 % in Acc-7, 0.002 in MAE, and 0.04 in Corr; on CMU-MOSEI, it achieves a 0.09 % improvement in F1-score.}, archive = {J_NEUCOM}, author = {Yijun Chen and Lina Yang and Thomas Wu and Yifeng Tan and Xichun Li and Zhonghua Li and Wanxing Zhu}, doi = {10.1016/j.neucom.2025.131212}, journal = {Neurocomputing}, month = {11}, pages = {131212}, shortjournal = {Neurocomputing}, title = {SAFCN: Self-enhancing attention fusion contrastive network for multimodal sentiment analysis}, volume = {654}, year = {2025}, }
• (2025). DualAug: Exploiting additional heavy augmentation with OOD data rejection. NEUCOM, 654, 131209. (www)

  @article{ , abstract = {Data augmentation is a dominant method for reducing model overfitting and improving generalization. Most existing data augmentation methods tend to find a compromise in augmenting the data, i.e., increasing the amplitude of augmentation carefully to avoid degrading some data too much and doing harm to model performance. We delve into the relationship between data augmentation and model performance, revealing that the performance drop with heavy augmentation comes from the presence of out-of-distribution (OOD) data. Nonetheless, as the same data transformation has different effects on different training samples, even with heavy augmentation, there remains part of in-distribution data that is beneficial to model training. Based on this observation, we propose a novel data augmentation method, named DualAug , to keep the augmentation in distribution as much as possible at a reasonable time and computational cost. We design a data mixing strategy to fuse augmented data from both the basic- and the heavy-augmentation branches. Extensive experiments on supervised image classification benchmarks show that DualAug improves various automated data augmentation methods. Moreover, the experiments on semi-supervised learning and contrastive self-supervised learning demonstrate that our DualAug can also improve related methods. Code is available at https://github.com/shuguang99/DualAug .}, archive = {J_NEUCOM}, author = {Zehao Wang and Yiwen Guo and Qizhang Li and Guanglei Yang and Wangmeng Zuo}, doi = {10.1016/j.neucom.2025.131209}, journal = {Neurocomputing}, month = {11}, pages = {131209}, shortjournal = {Neurocomputing}, title = {DualAug: Exploiting additional heavy augmentation with OOD data rejection}, volume = {654}, year = {2025}, }
• (2025). ADP-GNN: A spectral graph neural network framework unifying homophilic and heterophilic patterns via adaptive dual-polynomial filters. NEUCOM, 654, 131206. (www)

  @article{ , abstract = {Spectral Graph Neural Networks (GNNs), also known as spectral-domain graph filters, have demonstrated state-of-the-art performance in graph-structured data processing tasks. Initially, these filters typically relied on the eigendecomposition of the graph Laplacian matrix to implement the graph Fourier transform, resulting in high computational complexity. To reduce this complexity, numerous polynomial filters have been proposed. However, existing polynomial approaches often utilize predefined, fixed polynomial forms that cannot adapt to varying degrees of graph heterogeneity encountered in practical applications. To address this limitation, this paper first theoretically examines the intrinsic relationship between graph heterogeneity and filter characteristics. Inspired by the principles of ensemble learning, we then propose an A daptive D ual- P olynomial spectral GNN model ( ADP-GNN ), which integrates Bernstein and Taylor bases through an adaptive filtering mechanism. Unlike conventional single-basis approaches, our formulation leverages the complementary strengths of Bernstein polynomials, which offer global stability and boundary robustness, and Taylor polynomials, known for their local approximation accuracy. This model effectively exploits initial node features and adaptively captures both homogeneous and heterogeneous structural properties of graphs with varying homophily levels. Experimental evaluations demonstrate that ADP-GNN can effectively learn diverse filter structures, significantly mitigate the oversmoothing problem, and consistently outperform current state-of-the-art methods across multiple benchmark datasets. The source code is publicly available at: https://github.com/zhu165/ADP-GNN .}, archive = {J_NEUCOM}, author = {Huaguang Zhu and Ran Chen and Liwen Xu}, doi = {10.1016/j.neucom.2025.131206}, journal = {Neurocomputing}, month = {11}, pages = {131206}, shortjournal = {Neurocomputing}, title = {ADP-GNN: A spectral graph neural network framework unifying homophilic and heterophilic patterns via adaptive dual-polynomial filters}, volume = {654}, year = {2025}, }
• (2025). Attention-guided multi-scale temporal interaction diffusion model for 3D human motion generation. NEUCOM, 654, 131205. (www)

  @article{ , abstract = {Text-to-motion generation has many important applications in animation, virtual reality and robotics. Pioneer works directly map another modality such as textual descriptions to human motions. State-of-the-art methods incorporate the powerful Denoising Diffusion Probabilistic Models (DDPMs) and the Transformer backbone into the motion generation framework. However, due to the insufficient interaction between textual features and temporal motion features, the motion sequences generated by these methods are not highly consistent with the textual descriptions. Moreover, since DDPMs require a large number of sampling steps, existing motion diffusion models are usually very slow. In this paper, we propose an Attention-Guided Multi-Scale Temporal Interaction Diffusion Model (AMTIDM). To address the inconsistency issue, we design a novel Multi-Scale Temporal Interaction Network (MTIN) and an attention-guided loss to fully interact textual features with temporal motion features, significantly enhancing motion generation results. Furthermore, we introduce a lightweight noise schedule predictor to reduce the number of inference steps. In the experiments, AMTIDM achieves the highest accuracy on both the HumanAct12 and UESTC datasets. Our method can achieve a similar R-Precision to real human motions (on the HumanML3D benchmark). Compared to most diffusion-based methods, our model uses fewer parameters, requires lower computational cost (in terms of FLOPs), and achieves superior performance.}, archive = {J_NEUCOM}, author = {Hongye Yang and Jianfeng Feng}, doi = {10.1016/j.neucom.2025.131205}, journal = {Neurocomputing}, month = {11}, pages = {131205}, shortjournal = {Neurocomputing}, title = {Attention-guided multi-scale temporal interaction diffusion model for 3D human motion generation}, volume = {654}, year = {2025}, }
• (2025). Semantic driven ViT and dual-level token fusion for weakly supervised semantic segmentation. NEUCOM, 654, 131204. (www)

  @article{ , abstract = {Inspired by the observation that existing pre-trained models possess the capability to locate discriminative features but lack the learning of those non-discriminative ones to exploit their complementary support, we propose to introduce additional semantic images derived from the input using pre-trained CAMs and segmentation to empower the visual transformer with a dual-branched expansion towards improved performance for weakly supervised semantic segmentation (WSSS). To facilitate the interaction between discriminative and non-discriminative feature attentions, we introduce two token fusion modules: Interchangeable Class Token Fusion (ICTF) and Patch Token Fusion (PTF). While ICTF operates at the class level to exchange class tokens between the branches, PTF enables a balanced integration at the patch level by incorporating additional learning parameters. To minimize complexity and computational cost, we implement a weight-sharing scheme, allowing the model to efficiently learn from both the segmented semantic features and the global features of the original images. Extensive experiments and ablation studies on PASCAL VOC 2012 and MS COCO 2014 demonstrate that our proposed method outperforms a number of representative existing benchmarks in the tasks of WSSS. To encourage open access and further research, we make our source codes publicly available via the link .}, archive = {J_NEUCOM}, author = {Xiang Zhong and Qiudan Zhang and Jianmin Jiang}, doi = {10.1016/j.neucom.2025.131204}, journal = {Neurocomputing}, month = {11}, pages = {131204}, shortjournal = {Neurocomputing}, title = {Semantic driven ViT and dual-level token fusion for weakly supervised semantic segmentation}, volume = {654}, year = {2025}, }
• (2025). Safe optimal control for multiplayer Stackelberg–Nash games of nonlinear time-delay systems via adaptive dynamic programming. NEUCOM, 654, 131203. (www)

  @article{ , abstract = {This paper investigates a safe optimal control method for time-delay nonlinear systems based on multiplayer Stackelberg-Nash games (SNGs). Initially, considering different roles of players in the multiplayer SNGs, a hierarchical decision-making process is described as the designed value functions for the leader and all followers. Next, this paper incorporates a control barrier function into the value function to ensure that the system states remain within a safe range, thereby guaranteeing safety while achieving the optimization. Meanwhile, new cost functions are constructed which include the Lyapunov-Krasovskii ( l -K) function to eliminate time-delay effects, and a single critic neural network is used to approximate the optimal controller for the leader and the N -followers. Based on the Lyapunov stability theory, it is proven that all signals in the closed-loop system are uniformly ultimately bounded (UUB). Finally, a simulation example is provided to demonstrate the effectiveness of the proposed optimal control method.}, archive = {J_NEUCOM}, author = {Junzheng Zhao and Hamid Reza Karimi and Ben Niu and Xudong Zhao}, doi = {10.1016/j.neucom.2025.131203}, journal = {Neurocomputing}, month = {11}, pages = {131203}, shortjournal = {Neurocomputing}, title = {Safe optimal control for multiplayer Stackelberg–Nash games of nonlinear time-delay systems via adaptive dynamic programming}, volume = {654}, year = {2025}, }
• (2025). Binary pooling: A novel approach for local feature extraction in convolutional neural networks. NEUCOM, 654, 131201. (www)

  @article{ , abstract = {Pooling operations are essential in Convolutional Neural Networks (CNNs) for reducing spatial dimensions and emphasizing relevant features. While Max and Average Pooling remain popular choices, they often overlook subtle local structures critical for certain tasks. To address this limitation, we propose a new binary pooling approach that captures fine-grained patterns by comparing each pixel to its neighbors and binarizing the local differences. A subsequent normalization step ensures stability by mitigating the influence of extreme values. Unlike conventional Local Binary Pattern (LBP) pooling, our method integrates local statistical variability to produce a more adaptive and robust representation. The proposed operation is differentiable and can seamlessly replace traditional pooling layers in standard CNN architectures, enabling end-to-end training without disrupting gradient flow. Experiments conducted on MNIST, CIFAR-10, and Chest X-ray datasets show that the proposed method consistently achieves competitive or slightly superior classification performance, particularly in scenarios where detailed local information is crucial.}, archive = {J_NEUCOM}, author = {Boussaad Leila}, doi = {10.1016/j.neucom.2025.131201}, journal = {Neurocomputing}, month = {11}, pages = {131201}, shortjournal = {Neurocomputing}, title = {Binary pooling: A novel approach for local feature extraction in convolutional neural networks}, volume = {654}, year = {2025}, }
• (2025). Stateless distributed stein variational gradient descent method for bayesian federated learning. NEUCOM, 654, 131198. (www)

  @article{ , abstract = {Traditional frequentist federated learning algorithms only learn point-estimated model parameters, which renders the obtained models lacking uncertainty measurements and susceptible to over-fitting problems. By inferring the posterior distribution of the model parameters, Bayesian federated learning (BFL) can provide uncertainty quantification to the inference result, making it a valuable tool for a wide range of mission-critical applications such as autonomous driving and medical diagnosis. In this paper, we propose a stateless distributed Stein variational gradient descent (SDSVGD) method to circumvent the issues inherent in existing BFL methods. After updating the local posterior particles in each scheduled client, the server updates the global posterior particles and the averaged likelihood particles subsequently. The SDSVGD method eliminates the need for the server to maintain local likelihood particles for any client, significantly reducing its storage requirements and enhancing the scalability of BFL. Meanwhile, the averaged likelihood particles are estimated in the server and used as a regularizer to avoid the adverse influence of outdated or inconsistent local likelihoods. Furthermore, we compute the density function from the gradient perspective and prove its rationality, which alleviates the influence of the curse of dimensionality when processing high-dimensional particles. Finally, we evaluated the performance of our proposed SDSVGD method by conducting extensive experiments on several regression and classification tasks. The experiment results demonstrate that the proposed SDSVGD method presents superior regression, classification, and calibration performance and excellent communication efficiency compared with other baseline algorithms.}, archive = {J_NEUCOM}, author = {Guorui Li and Jing Gan and Cong Wang and Sancheng Peng}, doi = {10.1016/j.neucom.2025.131198}, journal = {Neurocomputing}, month = {11}, pages = {131198}, shortjournal = {Neurocomputing}, title = {Stateless distributed stein variational gradient descent method for bayesian federated learning}, volume = {654}, year = {2025}, }
• (2025). Unbiased multimodal intent recognition with auxiliary rationale generation. NEUCOM, 654, 131197. (www)

  @article{ , abstract = {Multimodal Intent Recognition (MIR) is a crucial task in understanding user intentions from diverse input sources such as text, images, and audio. While recent advancements using Large Multimodal Models (LMMs) have yielded promising results in intent classification, these models often exhibit a strong modality bias—favoring textual inputs at the expense of visual and auditory cues. Our study empirically investigates this phenomenon across standard benchmark datasets, revealing a consistent over-reliance on textual information. We attribute this imbalance to the typically more explicit nature of intent cues in text, which can overshadow the more nuanced signals present in other modalities. To mitigate this bias, we propose a novel MIR framework that incorporates an auxiliary rationale generation task aimed at producing modality-grounded justifications for classification decisions. By jointly training intent classification and explanation generation through a reinforcement learning-based multi-task paradigm, our approach facilitates deeper multimodal integration. The classification output guides the explanation process via reward signals, while the generated rationales enhance the model’s interpretive capacity and provide complementary feedback for intent prediction. Experimental results on the MIR and EMIR datasets show that our method outperforms existing state-of-the-art approaches, achieving a relative improvement of approximately 3.24 % in classification accuracy over the leading CAGC model. Further ablation studies and qualitative analyses confirm the effectiveness of the auxiliary task in reducing modality bias and improving interpretability.}, archive = {J_NEUCOM}, author = {Qian Dong and Ruiting Dai and Guiduo Duan and Ke Qin and Yuyang Zhang and Tao He}, doi = {10.1016/j.neucom.2025.131197}, journal = {Neurocomputing}, month = {11}, pages = {131197}, shortjournal = {Neurocomputing}, title = {Unbiased multimodal intent recognition with auxiliary rationale generation}, volume = {654}, year = {2025}, }
• (2025). SEMPose: A single end-to-end network for multi-object pose estimation. NEUCOM, 654, 131194. (www)

  @article{ , abstract = {In computer vision, estimating the six-degree-of-freedom pose from an RGB image is a fundamental task. However, this task becomes highly challenging in multi-object scenes. Currently, the best methods typically employ an indirect strategy, which identifies 2D and 3D correspondences, and then solves with the Perspective-n-Point method. Yet, this approach cannot be trained end-to-end. Direct methods, on the other hand, suffer from lower accuracy due to challenges such as varying object sizes and occlusions. To address these issues, we propose SEMPose, an end-to-end multi-object pose estimation network. SEMPose utilizes a well-designed texture-shape guided feature pyramid network, effectively tackling the challenge of object size variations. Additionally, it employs an iterative refinement head structure, progressively regressing rotation and translation separately to enhance estimation accuracy. During training, we alleviate the impact of occlusion by selecting positive samples from visible parts. Experimental results demonstrate that SEMPose can perform inference at 32 FPS without requiring inputs other than the RGB image. It can accurately estimate the poses of multiple objects in real time, with inference time unaffected by the number of target objects. On the LM-O and YCB-V datasets, our method outperforms other RGB-based single-model methods, achieving higher accuracy. Even when compared with multi-model methods and approaches that use additional refinement, our results remain competitive.}, archive = {J_NEUCOM}, author = {Xin Liu and Hao Wang and Shibei Xue and Dezong Zhao}, doi = {10.1016/j.neucom.2025.131194}, journal = {Neurocomputing}, month = {11}, pages = {131194}, shortjournal = {Neurocomputing}, title = {SEMPose: A single end-to-end network for multi-object pose estimation}, volume = {654}, year = {2025}, }
• (2025). Time-enhanced compound geometric operations for temporal knowledge graph embedding. NEUCOM, 654, 131182. (www)

  @article{ , abstract = {Temporal knowledge graph embedding aims to represent entities, relations, and timestamps in a continuous vector space while preserving their temporal evolutionary patterns. TCompoundE is a recently proposed temporal knowledge graph embedding model that uses compound operations involving translation and scaling as the relation-specific and time-specific operations. However, this model has the following three problems: (1) The model only employs traditional geometric transformations and does not explore more complex transformation relationships. (2) The model fails to enhance the head entity with time-awareness, which weakens the model’s ability to model dynamic changes. (3) The model neglects the influence of the head entity and the relation in the modeling process. We propose the TTComE, a time-enhanced compound geometric temporal knowledge graph embedding model, to tackle these questions. Here are three improvement points: (1) We introduced the Spatiotemporal Rotation Operation for modeling, which enables the model to better capture complex relationships and temporal evolution patterns. (2) We performed the Temporal-aware Enhancement Operation on the head entities and relations respectively, to enhance the model’s ability to capture dynamic changes. (3) We introduced the Weight-adaptive Translation Operation, which assigns learnable weights to both the head entity and the relation, followed by a translation operation. This enables the model to adaptively adjust the contributions of the head entity and the relation in prediction tasks. Finally, we experimentally validated the model on multiple datasets, complemented by ablation studies, and the results demonstrated significant improvements over other baseline models.}, archive = {J_NEUCOM}, author = {Wenhao Li and Dong Zhang and Guanyu Li and Yingqi Zou}, doi = {10.1016/j.neucom.2025.131182}, journal = {Neurocomputing}, month = {11}, pages = {131182}, shortjournal = {Neurocomputing}, title = {Time-enhanced compound geometric operations for temporal knowledge graph embedding}, volume = {654}, year = {2025}, }
• (2025). Breaking the conventional forward-backward tie in neural networks: Activation functions. NEUCOM, 654, 131178. (www)

  @article{ , abstract = {Gradient-based neural network training traditionally enforces symmetry between forward and backward propagation, requiring activation functions to be differentiable (or sub-differentiable) and strictly monotonic in certain regions to prevent flat gradient areas. This symmetry, linking forward activations closely to backward gradients, significantly restricts the selection of activation functions, particularly excluding those with substantial flat or non-differentiable regions. In this paper, we challenge this assumption through mathematical analysis, demonstrating that precise gradient magnitudes derived from activation functions are largely redundant, provided the gradient direction is preserved. Empirical experiments conducted on foundational architectures—such as Multi-Layer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and Binary Neural Networks (BNNs)—confirm that relaxing forward-backward symmetry and substituting traditional gradients with simpler or stochastic alternatives does not impair learning and may even enhance training stability and efficiency. We explicitly demonstrate that neural networks with flat or non-differentiable activation functions, such as the Heaviside step function, can be effectively trained, thereby expanding design flexibility and computational efficiency. Further empirical validation with more complex architectures remains a valuable direction for future research.}, archive = {J_NEUCOM}, author = {Luigi Troiano and Francesco Gissi and Vincenzo Benedetto and Genny Tortora}, doi = {10.1016/j.neucom.2025.131178}, journal = {Neurocomputing}, month = {11}, pages = {131178}, shortjournal = {Neurocomputing}, title = {Breaking the conventional forward-backward tie in neural networks: Activation functions}, volume = {654}, year = {2025}, }
• (2025). Optimal control for singularly perturbed systems with modified stabilization: A novel two-stage online reinforcement learning approach. NEUCOM, 654, 131175. (www)

  @article{ , abstract = {This paper focuses on the optimal control problem for singularly perturbed systems based on reinforcement learning (RL). Firstly, we consider transforming the dynamics of the original system to attempt to adjust the speed of the system response and the cost of the performance metric. Subsequently, the transformed more general algebraic Riccati equation (GARE) with perturbed parameters is explored in the framework of policy iteration (PI). Furthermore, with the advantage of RL, a novel two-stage homotopic RL scheme is devised to solve the GARE online, while obviating ill-conditioned numerical problems and relaxing the requirement of the initial stabilizing matrix in PI, and proofs of stability, convergence and optimality are provided. Finally, the effectiveness of the proposed algorithm is verified by two illustrative examples: the quarter-car model of automotive suspension system and the RC ladder circuit system.}, archive = {J_NEUCOM}, author = {Hai Wang and Junmin Li and Chao He}, doi = {10.1016/j.neucom.2025.131175}, journal = {Neurocomputing}, month = {11}, pages = {131175}, shortjournal = {Neurocomputing}, title = {Optimal control for singularly perturbed systems with modified stabilization: A novel two-stage online reinforcement learning approach}, volume = {654}, year = {2025}, }
• (2025). Disease and personality information enhanced depression detection based on the TransGCL framework. NEUCOM, 654, 131160. (www)

  @article{ , abstract = {Depression, the most common psychological disorder, is a major contributor to suicide. Medical research reveals that physical illnesses and personality traits significantly influence depression risk by inducing varying levels of psychological stress and emotional response. However, existing depression detection approaches face three major limitations: first, they rely solely on behavioral features while neglecting the influence of disease and personality information; second, they fail to effectively capture the spatial relationships among features across different time points; and third, conventional multimodal fusion methods struggle to fully exploit the complementary information between modalities, often resulting in redundant data. To address these issues, this paper introduces the NEUD multimodal depression dataset, which includes facial video, audio, gait, personality data, and disease data from 118 individuals. Additionally, the correlation between disease and depression severity was examined with the Kruskal–Wallis H test, while personality traits’ association with depression severity was analyzed using multiple linear regression. Finally, a multimodal depression recognition framework, TransGCL, is proposed. This framework incorporates a GCN module consisting of fully connected graphs and a trainable attenuation matrix, along with a regularization-based orthogonal tri-axis multimodal fusion approach. In addition, the framework integrates disease types and personality traits to enhance the accuracy of depression recognition. Experimental results demonstrate that the classification task achieved an accuracy of 74 %, while the mean absolute error for the regression task reached 4.37, validating the effectiveness of the proposed model and method. This study offers novel insights and methodologies for depression detection, providing valuable references for future research.}, archive = {J_NEUCOM}, author = {Yuliang Zhao and Huawei Zhang and Tingting Sun and Zhicheng Zhou and Jian Li and Chao Lian and Kaixuan Tian and Yulin Wang and Changzeng Fu}, doi = {10.1016/j.neucom.2025.131160}, journal = {Neurocomputing}, month = {11}, pages = {131160}, shortjournal = {Neurocomputing}, title = {Disease and personality information enhanced depression detection based on the TransGCL framework}, volume = {654}, year = {2025}, }
• (2025). UCFN: Uncertainty-aware cross-granularity fusion network for visual intention understanding. NEUCOM, 654, 131117. (www)

  @article{ , abstract = {Visual intention understanding (VIU) is the task of understanding the inherent intention (such as emotion, opinion, and motivation) behind the image related to the human behavior, which is a combination of various objects, backgrounds, and implicit relationships. Existing methods suffer from the various magnitudes of label ambiguity at the different granularity levels, resulting in the weak performances for the ambiguous samples. For the reliable understanding of visual content, an uncertainty-aware cross-granularity fusion network (UCFN) is proposed for VIU, which utilizes the evidence-based uncertainty to guide the fusion of the multi-granularity information, mitigating the ambiguity of the different granularity levels. Specifically, we first model the evidence-based uncertainty of the different granularity levels via evidential deep learning (EDL), jointing the classification and uncertainty estimation into a unified paradigm. Then, a simplified Dempster-Shafer (DS) combination rule is adopted to fuse the evidence vectors across the multi-granularity label spaces, while a novel Fine-to-Coarse (F2C) alignment strategy is designed to align the dimensions of the evidence vectors at the multi-granularity levels. Extensive experimental results validate the superiority of the proposed method compared with the state-of-the-art methods for VIU. The source codes are available at https://github.com/WenboXu717/UFCN .}, archive = {J_NEUCOM}, author = {Wenbo Xu and Xinde Li and Chuanfei Hu and Huaping Liu}, doi = {10.1016/j.neucom.2025.131117}, journal = {Neurocomputing}, month = {11}, pages = {131117}, shortjournal = {Neurocomputing}, title = {UCFN: Uncertainty-aware cross-granularity fusion network for visual intention understanding}, volume = {654}, year = {2025}, }
• (2025). RegionDiffusion: Generative data augmentation for object detection with diffusion models. NEUCOM, 654, 131115. (www)

  @article{ , abstract = {Recently, research on generative data augmentation based on layout-to-image diffusion models for object detection has made some progress. However, there remain challenges in generated data, such as low visual quality, weak generalization and mismatches between generated images and annotations, which hinder the performance of object detectors. To alleviate these issues, we propose the RegionDiffusion, a novel layout-to-image model that uses an attention block called Region Self-Attention to divide the self-attention operation among visual tokens into several independent regions based on the bounding box of the layout to generate objects accurately in the given locations. Then, we design a pipeline utilizing a series of preprocessing and postprocessing steps to improve the quality of the generated data and then train detectors. With extensive experiments, we demonstrate that RegionDiffusion surpasses state-of-the-art models in layout-to-image generation and our pipeline significantly improves the performance of object detectors in both general domains and specific domains.}, archive = {J_NEUCOM}, author = {Kesong Li and Shang Luo and Kuo-Kun Tseng}, doi = {10.1016/j.neucom.2025.131115}, journal = {Neurocomputing}, month = {11}, pages = {131115}, shortjournal = {Neurocomputing}, title = {RegionDiffusion: Generative data augmentation for object detection with diffusion models}, volume = {654}, year = {2025}, }
• (2025). First/second-order norm-based gradient neural networks for dynamic linear matrix equation. NEUCOM, 654, 131072. (www)

  @article{ , abstract = {This paper develops two novel gradient neural networks (GNNs) for efficiently solving the dynamic linear matrix equation (DLME), namely first/second-order norm-based GNNs (FNGNN and SNGNN). Benefiting from a novel error function and gradient descent strategy, the proposed GNNs eliminate the reliance of conventional recurrent neural networks (RNNs) on inversion operations and time derivative information of the coefficient matrices, thereby reducing computational complexity. Moreover, the FNGNN and SNGNN feature scaling parameters that adaptively vary with the error function, further optimizing computational resource allocation. Theoretical results confirm the convergence properties of the proposed GNNs. Additionally, several numerical experiments are performed to validate their efficiency in solving the DLME. When compared to several classical or state-of-the-art RNNs, the FNGNN and SNGNN exhibit significant superiority in convergence rate and precision, respectively. Finally, applications of the proposed GNNs to dynamic alternating current computing and 6-link planar robot manipulator path-tracking tasks further highlight their effectiveness and practical feasibility.}, archive = {J_NEUCOM}, author = {Chunhao Han and Jiao Xu}, doi = {10.1016/j.neucom.2025.131072}, journal = {Neurocomputing}, month = {11}, pages = {131072}, shortjournal = {Neurocomputing}, title = {First/second-order norm-based gradient neural networks for dynamic linear matrix equation}, volume = {654}, year = {2025}, }
• (2025). Domain-unique group- and its subgroups-aware fault diagnostics for machine components: An open-set domain adaptation approach. NEUCOM, 653, 131291. (www)

  @article{ , abstract = {Open-set domain adaptation has advanced cross-domain fault diagnosis for machine components, yet critical challenges remain. Existing methods often fail when both domains contain private classes, and they struggle to differentiate unknown fault subgroups for fine-grained diagnostics. To bridge these gaps, this study proposes a novel open-set domain adaptation framework that enhances learning granularity and generalizability. The proposed model integrates a conditional variational autoencoder to learn noise-resilient feature representations, and a triple-parallel adversarial structure to jointly optimize alignment, group awareness, and subgroup separation. Crucially, a bi-directional filter is designed to prevent negative transfer by excluding private instances from both domains during alignment. The framework’s versatility and effectiveness are validated through extensive experiments on two fundamentally different industrial scenarios: an image-based diagnosis of ball screw surface defects, transferring knowledge from a related manufacturing context, and a vibration-based diagnosis of rolling bearing faults across different operating conditions. Results demonstrate that the proposed model outperforms state-of-the-art methods in both cases, achieving higher accuracy with improved awareness of the domain-unique group and its subgroups. The findings confirm the model’s high adaptability, suggesting its potential for robust diagnostic knowledge transfer in open-set industrial scenarios with diverse components or data modalities.}, archive = {J_NEUCOM}, author = {Feng Xue and Weimin Zhang and Shulian Xie and Alexander Puchta and Jürgen Fleischer}, doi = {10.1016/j.neucom.2025.131291}, journal = {Neurocomputing}, month = {11}, pages = {131291}, shortjournal = {Neurocomputing}, title = {Domain-unique group- and its subgroups-aware fault diagnostics for machine components: An open-set domain adaptation approach}, volume = {653}, year = {2025}, }
• (2025). Beyond dimensionality explosion: A latent diffusion framework for hyperspectral image classification. NEUCOM, 653, 131249. (www)

  @article{ , abstract = {In recent years, deep generative learning-based methods for hyperspectral image (HSI) classification have achieved significant progress. However, the inherently high dimensionality of hyperspectral data, combined with the modeling complexity of generative approaches, often leads these methods to suffer from a curse of dimensionality explosion during feature extraction. To address this issue, this paper proposes an HSI classification and generation framework based on a latent diffusion model. The framework simulates and reconstructs the spectral–spatial high-dimensional feature distribution through a generative diffusion mechanism in a low-dimensional latent space, and realizes feature extraction via latent features. This approach effectively avoids feature dimensionality explosion while achieving efficient spectral–spatial feature fusion. Specifically, the framework consists of a spectral–spatial diffusion module and an attention-based classification module. The spectral–spatial diffusion module first maps the high-dimensional HSI data to a low-dimensional latent space, where spectral–spatial diffusion modeling and noise injection are performed. Subsequently, the high-dimensional features are restored through a reverse denoising process, facilitating efficient fusion of spatial and spectral information. After training, low-dimensional features extracted from the latent space are fed directly into the classification module. The core advantage of this method lies in its latent space diffusion modeling and feature extraction, which effectively integrates spectral–spatial information while avoiding dimensionality explosion, thereby improving classification performance. Experimental results on multiple public datasets demonstrate that our framework outperforms previous state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Meiyun Lu and Ning Chen and Xia Yue and Yicong Li and Jun Yue and Leyuan Fang}, doi = {10.1016/j.neucom.2025.131249}, journal = {Neurocomputing}, month = {11}, pages = {131249}, shortjournal = {Neurocomputing}, title = {Beyond dimensionality explosion: A latent diffusion framework for hyperspectral image classification}, volume = {653}, year = {2025}, }
• (2025). Optimizing the temporal adjacency matrix for 3D human pose estimation through clustering. NEUCOM, 653, 131247. (www)

  @article{ , abstract = {With the booming development of multimedia technologies related to 3D images and videos, 3D human pose estimation has gained increasing attention. In 3D human pose estimation, exploring the temporal relationships between human joints is essential. Transformer-based methods for modeling temporal relationships leverage the temporal adjacency matrix within the self-attention mechanism as a key component for capturing these connections. We define temporal features that occupy a small portion of all temporal features and are distant from other temporal features in the high-dimensional space as noisy features. In the temporal adjacency matrix of self-attention, noisy features can interfere with the computation of other features, and thus their correlation with other features should be eliminated. However, existing methods overlook this issue. To address this issue, we propose a DBSCAN-based clustering module to detect noisy temporal features and an adjacency matrix masking mechanism to suppress their influence. First, we cluster the input temporal features to obtain noisy features and clustering results. Then, we eliminate the correlations of noisy features on other temporal features within the adjacency matrix and reduce the correlations between different classes. Extensive experiments on the Human3.6M and MPI-INF-3DHP datasets, using state-of-the-art methods as benchmarks, demonstrate that our approach achieves improvements of up to 6.94 % Mean Per Joint Position Error (MPJPE) compared to the original methods with ground–truth input.}, archive = {J_NEUCOM}, author = {Yingfeng Wang and Muyu Li and Nan Meng and Min Xu}, doi = {10.1016/j.neucom.2025.131247}, journal = {Neurocomputing}, month = {11}, pages = {131247}, shortjournal = {Neurocomputing}, title = {Optimizing the temporal adjacency matrix for 3D human pose estimation through clustering}, volume = {653}, year = {2025}, }
• (2025). A high-performance and lightweight five-class white blood cell classification network optimized by attention mechanisms. NEUCOM, 653, 131246. (www)

  @article{ , abstract = {The variation in the number and proportion of white blood cells (WBCs) in peripheral blood is a critical indicator for the diagnosis of diseases such as bacterial infections and inflammation, making WBC classification essential in clinical practice. However, current AI-based WBC classification approaches face challenges including class imbalance within datasets and the trade-off between model accuracy and lightweight design. Transfer learning offers the combined advantages of traditional deep learning and lightweight network architectures. To address these issues, this study constructs a private, balanced WBC dataset comprising 3726 images, effectively resolving class imbalance concerns. Based on this dataset, we propose a transfer learning method optimized with the Bottleneck Attention Module (BAM), aimed at enhancing the performance of a lightweight network for five-class WBC classification. This method eliminates the need for manual image segmentation, thereby reducing computational and time costs. Experimental results demonstrate that our approach achieves excellent overall performance across accuracy, precision, recall, and F1 score metrics, significantly outperforming the baseline model. Moreover, our approach exhibits strong generalizability, offering an efficient, cost-effective, and easily deployable solution for automated WBC classification.}, archive = {J_NEUCOM}, author = {Hongyu Li and Bianzheng Wang and Honghua Hu and Yusheng Fu and ZiYan Yin and Taiqiang Zhao and Fenglan Tang and Jinhong Guo and Jiuchuan Guo}, doi = {10.1016/j.neucom.2025.131246}, journal = {Neurocomputing}, month = {11}, pages = {131246}, shortjournal = {Neurocomputing}, title = {A high-performance and lightweight five-class white blood cell classification network optimized by attention mechanisms}, volume = {653}, year = {2025}, }
• (2025). IFSERec: A robust graph contrastive learning approach for implicit feedback recommendations. NEUCOM, 653, 131245. (www)

  @article{ , abstract = {Graph contrastive learning (GCL) has emerged as a promising approach for implicit feedback recommendation, leveraging the minimization of differences between diverse representations of the same node to learn effective embeddings. However, prevailing GCL-based methods often rely on intricate and heuristic adaptive view generation strategies, leading to challenges in adaptability and efficiency. This paper proposes IFSERec, a novel GCL-based recommendation model designed for simplicity and efficiency. IFSERec introduces a contextual interaction perception module to reconstruct and augment the original graph data, resulting in a comprehensive global graph structure. This process aims to uncover potential user-item implicit semantic interactions, thereby enriching the interaction information available for data augmentation. Furthermore, to enhance the model’s ability to discern inherent semantic information, we employ controlled Excitation Gumbel noise for the structured enhancement of interactive information. This strategic noise injection bolsters the model’s robustness against noisy data. Finally, a straightforward neighborhood aggregation mechanism is employed for message passing, effectively enriching node semantic information and improving adaptability. Extensive evaluations conducted on three benchmark datasets demonstrate that IFSERec not only achieves performance comparable to state-of-the-art baselines but also benefits from reduced space complexity and enhanced adaptability.}, archive = {J_NEUCOM}, author = {Yi Yang and Shaopeng Guan and Xiaoyang Wen}, doi = {10.1016/j.neucom.2025.131245}, journal = {Neurocomputing}, month = {11}, pages = {131245}, shortjournal = {Neurocomputing}, title = {IFSERec: A robust graph contrastive learning approach for implicit feedback recommendations}, volume = {653}, year = {2025}, }
• (2025). A landslide identification method based on integrated segmentation network and transfer learning. NEUCOM, 653, 131242. (www)

  @article{ , abstract = {Acquiring information on landslide locations and affected areas is crucial for geological disaster monitoring. This paper proposes a novel landslide identification method based on integrated segmentation network and transfer learning. Specifically, an integrated network architecture for landslide area segmentation is constructed using Kolmogorov-Arnold Network and Transformer-based U-Net, and the fusion of features at different scales is improved to accurately extract and restore the detailed features of landslides. To overcome the difficulty of landslide identification with indistinct optical features, a multi-modal image fusion network integrating Digital Elevation Model is proposed. A branch for extracting elevation features is added and a topographic-guided loss function is constructed. Furthermore, transfer learning is applied to improve the accuracy of landslide identification for small sample sizes. Experimental results validate the effectiveness of the proposed method in both single-modal identification of new landslides and multi-modal identification of old landslides.}, archive = {J_NEUCOM}, author = {Sijing Chen and Hanqi Qu and Yunyan Shao and Yuxuan Zeng and Zikang Wu and Chengda Lu and Min Wu}, doi = {10.1016/j.neucom.2025.131242}, journal = {Neurocomputing}, month = {11}, pages = {131242}, shortjournal = {Neurocomputing}, title = {A landslide identification method based on integrated segmentation network and transfer learning}, volume = {653}, year = {2025}, }
• (2025). Deep metric learning for end-to-end document classification. NEUCOM, 653, 131241. (www)

  @article{ , abstract = {Document classification systems become more and more complex with the need to deal with new document formats or categories while obtaining a low error rate when classifying more and more documents. Such systems need to have important features including (1) the ability to eliminate ambiguity to improve precision or reduce the error rate, (2) the capability to detect and reject documents belonging to new categories or new variations. Previous studies often focused on closed datasets or solely on the problem of novelty detection, without evaluating the ability to reject ambiguous results after the novelty detection. In this paper, we propose an end-to-end document classification algorithm including both novelty and ambiguity rejection. The proposed algorithm utilizes deep metric learning to compact the knowledge space, and then uses the last hidden layer’s features as input for an unsupervised KNN-based method for novelty and ambiguity rejection. Extensive experiments and analysis on private and public benchmark datasets demonstrate the effectiveness of our proposed algorithm. The algorithm provides the capability to handle new documents while effectively rejecting ambiguity to enhance the precision, recall of known categories, and coverage rate of the end-to-end document classification system.}, archive = {J_NEUCOM}, author = {Tri-Cong Pham and Mickaël Coustaty and Antoine Doucet and Aurélie Joseph and Vincent Poulain d’Andecy}, doi = {10.1016/j.neucom.2025.131241}, journal = {Neurocomputing}, month = {11}, pages = {131241}, shortjournal = {Neurocomputing}, title = {Deep metric learning for end-to-end document classification}, volume = {653}, year = {2025}, }
• (2025). Learning consistent representation for incomplete multi-view weak multi-label classification. NEUCOM, 653, 131238. (www)

  @article{ , abstract = {In recent years, multi-view multi-label learning has become a popular interdisciplinary research topic. Building on this, this paper further explores the complex real-world problem of incomplete multi-view weak multi-label learning. In this problem, how to effectively utilize interrelationships across views to embed consistent learning objectives into models has been a key challenge. To address this, we propose a Learning Consistent Representation Network (LCRN) based on the Information Noise-Contrastive Estimation (InfoNCE) loss. This method transforms consistency learning between multiple views into a pseudo dictionary look-up problem. By matching the query with a positive key, it treats different views from the same sample as positive pairs, thereby promoting consistency learning across multiple views. Benefiting from the flexibility of splitting a single view into query and key representations, we can achieve mutual alignment in consistency learning between two views. Finally, we use missing view and missing label indicators to mitigate the negative performance impact caused by missing multi-views and weak multi-labels. After extensive experiments on five incomplete multi-view weak multi-label datasets, our model demonstrated competitive results.}, archive = {J_NEUCOM}, author = {Kaibo Duan and Shaodong Cui and Hiroyuki Shinnou and Shi Bao}, doi = {10.1016/j.neucom.2025.131238}, journal = {Neurocomputing}, month = {11}, pages = {131238}, shortjournal = {Neurocomputing}, title = {Learning consistent representation for incomplete multi-view weak multi-label classification}, volume = {653}, year = {2025}, }
• (2025). Illumination guided domain adaptation object detection in thermal imagery. NEUCOM, 653, 131237. (www)

  @article{ , abstract = {Large-scale RGB datasets have played vital role in the huge success of training deep learning models. As a result, in many image applications, such as object detection, they have helped deliver superior performance in processing images from traditional cameras. However, under adverse illumination conditions, the visible-light cameras tend to capture poor and low-quality images. This may severely deteriorate the final performance in object detection. In contrast, thermal infrared cameras are very robust in various lighting conditions and thus can well serve as an efficient replacement or supplement to the regular cameras. However, training a robust thermal object detector faces its own limitations caused by the lack of large-scale labeled thermal data. To rectify this problem, we propose to supply the small-scale thermal dataset by leveraging the large-scale visible dataset, so as to enhance the performance of object detection models in the thermal domain. This leads to a novel domain adaptation method for thermal object detection, which integrates the illumination-guided attention mechanism module and the Prototype-based Adaptation module. The first module takes advantage of illumination information to selectively focus on domain-invariant features, which are important for robust adaptation across different domains. The second module, on the other hand, performs both category-agnostic and category-specific adaptation to further enhance the model’s adaptation ability. By leveraging the output of the attention mechanism module, the adaptation module can more effectively align the distribution between the visible and thermal domains, leading to better performance of detector in the thermal domain. Through extensive experiments, we have demonstrated that our approach achieves superior performance compared to state-of-the-art methods on both the KAIST and FLIR datasets. Our code is available at https://github.com/ZLab540/IATDA .}, archive = {J_NEUCOM}, author = {Yuan Zhou and Hongyang Zhang and Ye Sun and Yu Liu and Sun-Yuan Kung}, doi = {10.1016/j.neucom.2025.131237}, journal = {Neurocomputing}, month = {11}, pages = {131237}, shortjournal = {Neurocomputing}, title = {Illumination guided domain adaptation object detection in thermal imagery}, volume = {653}, year = {2025}, }
• (2025). A comprehensive survey of multi-agent deep reinforcement learning for wireless spectrum management. NEUCOM, 653, 131236. (www)

  @article{ , abstract = {The increasing demand for wireless communications and the rapid evolution of next-generation networks have significantly challenged traditional wireless spectrum management approaches. Deep reinforcement learning (DRL), especially multi-agent deep reinforcement learning (MADRL), has emerged as a powerful solution to address the dynamic and complex issues in wireless networks. This paper provides a comprehensive analysis of the core elements of DRL, including state representation, reward mechanism design, and learning paradigms, and outlines the development of MADRL. We review classic DRL algorithms and introduce current mainstream MADRL algorithms, focusing on research progress in enhancing their scalability, transferability, and security. Additionally, we present an in-depth review of the latest applications of DRL and MADRL in wireless spectrum management, with a particular emphasis on practical applications such as power control, spectrum sensing, allocation, channel access, and joint resource optimization in cognitive radio networks, the Internet of Things (IoT), and vehicular networks. We also discuss the major challenges faced by MADRL in the wireless domain, including limited generalization capabilities, adaptability constraints, and deployment complexity. Future research directions are proposed, highlighting the need for highly generalizable models, energy-efficient algorithms, and real-world deployment strategies. By synthesizing recent research findings, this paper aims to offer valuable insights and innovative ideas for researchers and practitioners working towards intelligent spectrum management solutions.}, archive = {J_NEUCOM}, author = {Ying Wang and Jianjun Lei and Fengjun Shang and Yuanyuan Li}, doi = {10.1016/j.neucom.2025.131236}, journal = {Neurocomputing}, month = {11}, pages = {131236}, shortjournal = {Neurocomputing}, title = {A comprehensive survey of multi-agent deep reinforcement learning for wireless spectrum management}, volume = {653}, year = {2025}, }
• (2025). Multi-level feature fusion for generalized face forgery detection. NEUCOM, 653, 131235. (www)

  @article{ , abstract = {With the meteoric rise of AI-based face forgery techniques, colloquially known as ‘Deepfakes’, the research community faces a huge challenge in designing detection tools that offer high detection accuracy, strong generalization and compact architectures for deployment on resource-constrained devices. In this paper, we introduce a novel Multi-Level Feature Fusion (MLFF) framework that employs a lightweight yet effective strategy for capturing artifacts across multiple resolutions, using a backbone network, enhanced by a Spatial Feature Selection Module (SFSM), and integrating these representations via a Feature Fusion Network (FFN), finally followed by a classification head for classification. We explore two pre-trained backbones—a CNN-based EfficientNet and a lightweight Swin Transformer, and demonstrate that our method not only surpasses SOTA approaches on the FaceForensics++ and FaceShifter benchmarks in terms of classification accuracy and AUC, but also exhibits superior generalization capabilities for cross-data set and cross-manipulation scenarios, while achieving more than 50 % model parameters reduction compared to current best-performing models, making it ideal for deployment in resource-limited environments.}, archive = {J_NEUCOM}, author = {Tanusree Ghosh and Ruchira Naskar}, doi = {10.1016/j.neucom.2025.131235}, journal = {Neurocomputing}, month = {11}, pages = {131235}, shortjournal = {Neurocomputing}, title = {Multi-level feature fusion for generalized face forgery detection}, volume = {653}, year = {2025}, }
• (2025). GGSTrack: Geometric graph with spatio-temporal convolution for multi-object tracking. NEUCOM, 653, 131234. (www)

  @article{ , abstract = {Despite significant advancements in Multi-Object Tracking (MOT) with CNNs and Transformers, current algorithms still struggle with two critical challenges : first, severe occlusion and visual similarities between objects in dense scenes leading to target loss or identity switches; second, motion blur phenomena and the inability of traditional algorithms to fully utilize temporal features when relying solely on current and previous frames. Inspired by the success of geometric deep learning in modeling complex spatial relationships and the effectiveness of temporal networks in sequence modeling, we propose GGSTrack , a Multi-Object Tracking algorithm based on Geometric Graph with Spatio-Temporal Convolutional Networks . Our approach transforms video frames into undirected graphs through Geometric Graph Convolutional Networks, effectively sparsifying dense scenes and optimizing node aggregation mechanisms to enhance target feature representation. Furthermore, we innovatively utilize the past T frames as input and model temporal dependencies using Temporal Convolutional Networks , capturing motion patterns and evolutionary trends along the temporal dimension and significantly improving tracking robustness. To enhance computational efficiency while maintaining accuracy, we design a four-stage cascade matching framework —comprising initial filtering, uncertain object matching, low-score processing, and occlusion handling—that progressively reduces the number of targets while preserving high-precision matching results. Extensive experiments on multiple public datasets demonstrate that our method achieves superior performance, particularly in handling dense scenes and complex occlusion scenarios.}, archive = {J_NEUCOM}, author = {Shun Yan and Zhenghao Wang and Youhao Huang and Yufan Liu and Zekai Liu and Fei Yang and Weigang Lu and Dong Li}, doi = {10.1016/j.neucom.2025.131234}, journal = {Neurocomputing}, month = {11}, pages = {131234}, shortjournal = {Neurocomputing}, title = {GGSTrack: Geometric graph with spatio-temporal convolution for multi-object tracking}, volume = {653}, year = {2025}, }
• (2025). A multimodal fusion framework for semantic segmentation of remote sensing based on multilevel feature fusion learning. NEUCOM, 653, 131233. (www)

  @article{ , abstract = {Significant advancements in remote sensing semantic segmentation using many image modalities have been accomplished since the multimodal fusion technique was developed. Nevertheless, existing methods have struggled to consider information from small objects effectively and have been prone to introducing feature noise into the fusion process due to the high number of small objects and the complex architectural textures and color range present in Very High-Resolution (VHR) images. To address these challenges, we have proposed an end-to-end Multi-Modal Fusion Network (MMFNet), which is based on a three-branch fusion framework for leveraging VHR remote sensing images and Digital Surface Model (DSM) data. Recognizing that remote sensing data exhibit a wide variety of colors and intricate textures and that different modalities convey different color and texture information, we have proposed a Local-Global Feature Fusion (LGFF) module to integrate information from different modalities. We have tackled the difficulty of segmenting easily overlooked small objects with limited pixel coverage by devising an Multi-Level Feature Learning (MLFL) module to capture different-level feature representations. Ultimately, the single-modal and multimodal segmentation information are fused in MMFNet to determine the prediction information jointly. Our proposed model attains excellent miou scores of 84.12 % and 86.13 % on the Vaihingen and Potsdam datasets, respectively, effectively establishing a new state-of-the-art benchmark in multimodal image segmentation tasks. These results underscore not only the model's superior performance but also its significant competitiveness in this domain, marking a notable advancement in the field.}, archive = {J_NEUCOM}, author = {Pingzhu Liu}, doi = {10.1016/j.neucom.2025.131233}, journal = {Neurocomputing}, month = {11}, pages = {131233}, shortjournal = {Neurocomputing}, title = {A multimodal fusion framework for semantic segmentation of remote sensing based on multilevel feature fusion learning}, volume = {653}, year = {2025}, }
• (2025). A generative AI model for modelling and creating an emotion-aware database: Design, and analytics. NEUCOM, 653, 131232. (www)

  @article{ , abstract = {In this study, we illustrate an ongoing work regarding the building of an ad hoc textual dataset for dialogues between a user and an agent of a customer care center. The user is required to express a given emotion and to use a specific level of the Common European Framework of Reference for Languages (CEFR). Once the criteria had been defined, we asked ChatGPT 3.5 to generate dialogues. We analyzed them to ensure an acceptable level of quality for the dataset. Furthermore, we measured the language complexity used both by the user and the agent through different readability measures. Moreover, we extracted from each turn of each dialogue the attitude that ChatGPT exhibited for each dialogue turn; the interaction patterns were then analyzed with algorithm Sequential Pattern Mining (SPM) to extract typical interaction mechanisms occurring in specific emotional contexts. The generated dataset of dialogues, enriched with the readability measures and the extracted patterns, can be exploited as a system that facilitates the interaction between humans and artificial agents in specific contexts and it can constitute the basis of a reinforcement learning system for improving the effectiveness of the dialogues.}, archive = {J_NEUCOM}, author = {Alfredo Cuzzocrea and Giovanni Pilato and Pablo G. Bringas}, doi = {10.1016/j.neucom.2025.131232}, journal = {Neurocomputing}, month = {11}, pages = {131232}, shortjournal = {Neurocomputing}, title = {A generative AI model for modelling and creating an emotion-aware database: Design, and analytics}, volume = {653}, year = {2025}, }
• (2025). A meta-survey of adversarial attacks against artificial intelligence algorithms, including diffusion models. NEUCOM, 653, 131231. (www)

  @article{ , abstract = {Deep neural networks have revolutionized artificial intelligence, solving complex issues in areas like healthcare or law enforcement and security. However, they are susceptible to adversarial attacks where small data manipulations can compromise system reliability and security. This paper conducts an umbrella review of the literature on these attacks, synthesizing results from various systematic reviews to assess attack strategies, defense effectiveness, and research gaps. Guided by the PICO framework, this review categorizes and examines adversarial attacks, identifying key challenges in the field. The review finds that even though adversarial vulnerabilities were first explored in computer vision, analogous threats have expanded to domains like graph neural networks, natural language processing, federated learning, and text-to-image models. Despite varied attack surfaces, commonalities can be found.}, archive = {J_NEUCOM}, author = {Marek Pawlicki and Aleksandra Pawlicka and Rafał Kozik and Michał Choraś}, doi = {10.1016/j.neucom.2025.131231}, journal = {Neurocomputing}, month = {11}, pages = {131231}, shortjournal = {Neurocomputing}, title = {A meta-survey of adversarial attacks against artificial intelligence algorithms, including diffusion models}, volume = {653}, year = {2025}, }
• (2025). Zero-shot neural architecture search with weighted response correlation. NEUCOM, 653, 131229. (www)

  @article{ , abstract = {Neural architecture search (NAS) is a promising approach for automatically designing neural network architectures. However, the architecture estimation of NAS is computationally expensive and time-consuming because of training multiple architectures from scratch. Although existing zero-shot NAS methods use training-free proxies to accelerate the architecture estimation, their effectiveness, stability, and generality are still lacking. We present a novel training-free estimation proxy called weighted response correlation (WRCor). WRCor utilizes correlation coefficient matrices of responses across different input samples to calculate the proxy scores of estimated architectures, which can measure their expressivity and generalizability. Experimental results on proxy evaluation demonstrate that WRCor and its voting proxies are more efficient estimation strategies than existing proxies. We also apply them with different search strategies in architecture search. Experimental results on architecture search show that our zero-shot NAS algorithm outperforms most existing NAS algorithms in different search spaces. Our NAS algorithm can discover an architecture with a 22.1 % test error on the ImageNet-1k dataset within 4 GPU hours. All codes are publicly available at https://github.com/kunjing96/ZSNAS-WRCor.git .}, archive = {J_NEUCOM}, author = {Kun Jing and Luoyu Chen and Jungang Xu and Jianwei Tai and Yiyu Wang and Shuaimin Li}, doi = {10.1016/j.neucom.2025.131229}, journal = {Neurocomputing}, month = {11}, pages = {131229}, shortjournal = {Neurocomputing}, title = {Zero-shot neural architecture search with weighted response correlation}, volume = {653}, year = {2025}, }
• (2025). Dynamical behaviors of the computational model in parkinsonian state under the effect of electromagnetic induction. NEUCOM, 653, 131227. (www)

  @article{ , abstract = {This study introduces a computational model of Parkinson’s disease that incorporates electromagnetic induction. The model has been enhanced to analyze firing rates, power spectral density, relay fidelity error, box dimension, and periodic point graphs. Indicator changes are evaluated in response to parameter variations in both healthy and diseased states. A novel fractal stimulation approach is proposed for treating Parkinson’s disease. Compared to traditional open-loop stimulation and PID control, fractal stimulation significantly impacts the subthalamic nucleus (STN), external segment of globus pallidus (GPe), and internal segment of globus pallidus (GPi), resulting in a distinct separation that facilitates parameter selection. Optimal stimulation, as determined by the specified objective function, depends on the weights assigned to different parameters. Each of control methods exhibits unique advantages under various conditions. Hence, it is imperative to adopt a personalized treatment approach for patients with Parkinson’s disease, tailored to their individual circumstances through a comprehensive analysis of their specific conditions.}, archive = {J_NEUCOM}, author = {Zhong Dai and Shutang Liu and Changan Liu}, doi = {10.1016/j.neucom.2025.131227}, journal = {Neurocomputing}, month = {11}, pages = {131227}, shortjournal = {Neurocomputing}, title = {Dynamical behaviors of the computational model in parkinsonian state under the effect of electromagnetic induction}, volume = {653}, year = {2025}, }
• (2025). SDENK: Unbiased subspace density-k-clustering. NEUCOM, 653, 131225. (www)

  @article{ , abstract = {Clustering is one of the most important data analysis techniques, as it extracts knowledge without requiring data labels, making it crucial in many unsupervised application scenarios. However, conventional k -clustering struggles to detect irregularly distributed clusters owing to its inherent preference for convex clusters, while density-based clustering often lacks the ability to customise an appropriate metric space for different clustering tasks owing to the lack of a task-oriented optimisation process. To simultaneously address these biases in cluster shape and metric space, this paper proposes an unbiased hybrid framework to perform density-based clustering in subspaces. These subspaces are constructed through a newly developed attribute-weighted k -clustering paradigm. To exploit the irregularly distributed clusters obtained via density-based clustering for subspace learning, a novel strategy is designed to subdivide the clusters into compact sub-clusters, which are more suitable for evaluating attribute importance through k -clustering. As a result, the proposed subspace density K -clustering algorithm inherits the shape flexibility of density-based clustering and the metric adaptiveness of k -clustering. Moreover, the learnable design enables mutual optimisation between density clusters and subspaces, yielding robust and superior clustering performance across various datasets. Comprehensive evaluations, including comparative clustering performance evaluation, ablation studies, significance tests, noise-robustness evaluation, and hyper-parameter sensitivity studies, are conducted. Among 10 compared methods, the proposed Subspace DENsity-K-clustering (SDENK) achieves an average rank of 1.75 on 12 datasets in terms of the clustering accuracy metric ARI.}, archive = {J_NEUCOM}, author = {Rong Zou and Yunfan Zhang and Mingjie Zhao and Zexi Tan and Yiqun Zhang and Yiu-Ming Cheung}, doi = {10.1016/j.neucom.2025.131225}, journal = {Neurocomputing}, month = {11}, pages = {131225}, shortjournal = {Neurocomputing}, title = {SDENK: Unbiased subspace density-k-clustering}, volume = {653}, year = {2025}, }
• (2025). Joint clustering and feature extraction with adaptive outlier penalty. NEUCOM, 653, 131224. (www)

  @article{ , abstract = {High-dimensional data clustering remains a critical challenge due to inherent sparsity, redundant features, and outliers. Conventional approaches typically separate feature extraction and clustering into sequential steps, failing to ensure that the projected data retains discriminative clustering structures. Furthermore, during the iterative clustering process, current methods lack adaptive mechanisms to dynamically identify and suppress outliers. To address these limitations, we propose a unified framework, Joint Clustering and Feature Extraction with Adaptive Outlier Penalty (JCFEAOP), which integrates clustering, feature extraction, and outlier detection into a single optimization process. Unlike separate strategies, JCFEAOP iteratively optimizes both tasks with cluster structures guiding feature extraction and the projected space refining clustering. Simultaneously, we design an adaptive outlier penalty method that identifies outliers by combining reconstruction errors from feature extraction and distances from cluster centers, thereby dynamically reducing their interference. We solve the proposed model using an alternating optimization approach, analyze its time complexity, and theoretically prove its convergence. Extensive experiments on eight datasets demonstrate the superiority of our algorithm in clustering performance and robustness.}, archive = {J_NEUCOM}, author = {Yiwen Wu and Xianneng Li and Fanqi Lin}, doi = {10.1016/j.neucom.2025.131224}, journal = {Neurocomputing}, month = {11}, pages = {131224}, shortjournal = {Neurocomputing}, title = {Joint clustering and feature extraction with adaptive outlier penalty}, volume = {653}, year = {2025}, }
• (2025). Advancements in neuromorphic computing for bio-inspired artificial vision: A review. NEUCOM, 653, 131221. (www)

  @article{ , abstract = {Neuromorphic computing is revolutionising artificial vision by emulating the human brain’s remarkable efficiency, adaptability, and spatio-temporal processing. This review synthesises recent advances in neuromorphic vision, with a special focus on wave-based dynamics; particularly the role of cortical travelling waves and neural oscillations in coordinating activity across the visual cortex. We examine how these biological mechanisms inspire cutting-edge computational models, including Physics-Informed Neural Networks, reservoir computing, and spiking neural networks, each enabling real-time, energy-efficient visual processing. The review also highlights breakthroughs in hardware, from memristive devices and photonic circuits to optoelectronic polymers, which support in-sensor and event-based processing while dramatically reducing power consumption. By integrating insights from computational neuroscience, materials science, and machine intelligence, we identify persistent challenges; such as scalable training, robust hardware integration, and biologically plausible modelling and outline actionable directions for future research. Our synthesis provides a comprehensive roadmap for the next generation of neuromorphic vision systems, paving the way toward artificial perception that rivals the efficiency and adaptability of biological vision.}, archive = {J_NEUCOM}, author = {Sharmarke A. Gabayre and Mindula Illeperuma and Varuna D. De-Silva and Xiyu Shi and Sergey E. Savel’ev}, doi = {10.1016/j.neucom.2025.131221}, journal = {Neurocomputing}, month = {11}, pages = {131221}, shortjournal = {Neurocomputing}, title = {Advancements in neuromorphic computing for bio-inspired artificial vision: A review}, volume = {653}, year = {2025}, }
• (2025). Hypergraph network with multi-granularity temporal representation for traffic accident prediction. NEUCOM, 653, 131219. (www)

  @article{ , abstract = {Traffic accidents have become a major global concern with the rapid pace of urbanization, posing significant challenges to urban safety and traffic management. Accurately predicting traffic accidents requires effectively modeling both spatial dependencies and temporal patterns. However, existing methods struggle to capture the high-order spatial relationships and complex interactions among heterogeneous traffic-related features. To address these challenges, this paper proposes the Hypergraph network with multi-granularity temporal representation for traffic accident prediction framework (HMGTAP). By leveraging hypergraph convolution, HMGTAP captures high-order spatial dependencies and facilitates the integration of diverse urban data. Simultaneously, a multi-granularity temporal modeling mechanism is designed to learn both short-term fluctuations and long-term trends, improving the model’s capacity to represent diverse temporal dynamics. Furthermore, an attention-based fusion strategy dynamically refines spatiotemporal representations, enhancing model adaptability and robustness. Extensive experiments conducted on real-world traffic datasets demonstrate that HMGTAP consistently outperforms state-of-the-art models, achieving performance improvements of 0.75 %–6.65 % across two datasets. These results validate the effectiveness of the proposed framework in advancing traffic accident prediction and supporting data-driven traffic management strategies.}, archive = {J_NEUCOM}, author = {Weibin Deng and Wei Wang and Sen Liu and Hong Yu and Jin Liu}, doi = {10.1016/j.neucom.2025.131219}, journal = {Neurocomputing}, month = {11}, pages = {131219}, shortjournal = {Neurocomputing}, title = {Hypergraph network with multi-granularity temporal representation for traffic accident prediction}, volume = {653}, year = {2025}, }
• (2025). Enhancing effective connection with link prediction for session-based recommendation. NEUCOM, 653, 131218. (www)

  @article{ , abstract = {The core objective of session-based recommendation (SBR) is to precisely capture user interest preferences from massive candidate items by modeling short-term behavioral patterns. Current approaches predominantly rely on global features or heuristic rules to restructure session graphs for enhanced data representation, yet face a critical limitation: ineffective modeling of implicit item transition patterns within individual sessions. Compounding this issue, while link prediction techniques have demonstrated proven effectiveness in pattern mining, their systematic integration into session recommendation remains underexplored. To address the above issues, this paper proposes the Enhancing effective connection with link prediction for session-based recommendation (ECLP-SR) framework, which introduces link prediction as an information enhancer for session recommendation. Our method employs a Variational Graph Autoencoder (VGAE) to disentangle higher-order item interaction patterns from historical sessions, generating probabilistic transition distributions through latent space inference. This enables quantitative identification of high-frequency, semantically meaningful connections. Unlike the traditional static graph construction method, we design a dynamic weight adjustment mechanism to reconstruct the edge weight distribution of the session graph based on edge importance, thus enhancing the data-driven semantic characterization capability. Experiments on three benchmark datasets show that ECLP-SR improves the semantic distinctiveness of the session graph structure and enhances the model recommendation performance. The source code is available at: https://github.com/Typejunjie/code2 .}, archive = {J_NEUCOM}, author = {Hui Wang and Junjie You and Jianbing Yi and Junyan Liu and Runjie Huang}, doi = {10.1016/j.neucom.2025.131218}, journal = {Neurocomputing}, month = {11}, pages = {131218}, shortjournal = {Neurocomputing}, title = {Enhancing effective connection with link prediction for session-based recommendation}, volume = {653}, year = {2025}, }
• (2025). Alzheimer’s disease detection via human activity recognition and temporal segmentation in untrimmed videos. NEUCOM, 653, 131215. (www)

  @article{ , abstract = {Alzheimer’s disease (AD) is a well-known type of dementia, characterized by neurodegeneration that interferes with daily activities. It affects the cells in the hippocampus region of the brain, leading to an irreversible brain disorder. People developing AD often struggle to remember small details and gradually lose the ability to perform complex daily living tasks. AD is a progressive disease that shows no symptoms in its early stages, making diagnosis challenging through periodic medical exams. Proposing a test that could be fully performed remotely at home by a family member could have a significant impact on diagnosing AD in its early stages. In this study, we propose a novel approach for detecting Alzheimer’s disease by analyzing daily living activities from untrimmed videos (such as visual data from the Dem@care dataset). The proposed approach leverages deep learning techniques to determine whether an individual is affected by Alzheimer’s or not, through a three-step pipeline: The first step (a) employs a Human action recognition approach, where an I3D model as Spatio-Temporal Convolutional Neural Network (ST-CNN) is trained and applied to differentiate between various daily activities based on RGB video data. The second step (b) is presented as Temporal action segmentation approach, using a sliding window technique to extract temporal information by generating a sequence of predicted actions from the untrimmed videos, this is achieved by applying the trained I3D model over the entire video. The third step (c) involves final classification, where a Long Short-Term Memory (LSTM) network is used to analyze the sequence of predicted actions and classify individuals as either cognitively healthy or affected by dementia.Our approach achieves an accuracy of 92.98 % on the Dem@care dataset, which, to the best of our knowledge, is the first result of this kind for such a methodology. Furthermore, we developed and tested an application for its deployment in real-world settings.}, archive = {J_NEUCOM}, author = {Mohamed Amine Zayene and Hend Basly and Annalisa Franco and Bachra Dridi and Fatma Ezahra Sayadi}, doi = {10.1016/j.neucom.2025.131215}, journal = {Neurocomputing}, month = {11}, pages = {131215}, shortjournal = {Neurocomputing}, title = {Alzheimer’s disease detection via human activity recognition and temporal segmentation in untrimmed videos}, volume = {653}, year = {2025}, }
• (2025). EC-fake: A fake news detection model based on external knowledge and contrast-driven feature augmentation. NEUCOM, 653, 131214. (www)

  @article{ , abstract = {Fake news has spurt with the rapid development of Internet technology and social media, causing a number of serious negative impacts on individuals and society. Aiming at the shortcomings of existing multi-modal fake news detection on cross-modal feature fusion and external knowledge features, we proposed a multi-modal fake news detection model based on external knowledge and contrast-driven feature augmentation named EC-Fake. EC-Fake employs a Semantic Information Decoupling (SID) module to separate the foreground subject and background context of news images. SID can be utilized to guide multi-level feature fusion and key feature augmentation. This enables EC-Fake to reduce uncertainty in image features and allocate more attention to foreground and background features. Then, we used Large Language Models (LLMs) to extract external knowledge from news and fused it with multi-modal features, thereby effectively improving the interpretability and accuracy of EC-Fake. The accuracy on three public datasets is improved by 1.9 %, 1.6 %, and 2.7 % over the state-of-the-art model, respectively. EC-Fake can provide technical support to social media platforms and government regulators in detecting fake news.}, archive = {J_NEUCOM}, author = {Mingxing Han and Jiaxuan Li and Yu Chen and Liwei Xu and Lingling Tao}, doi = {10.1016/j.neucom.2025.131214}, journal = {Neurocomputing}, month = {11}, pages = {131214}, shortjournal = {Neurocomputing}, title = {EC-fake: A fake news detection model based on external knowledge and contrast-driven feature augmentation}, volume = {653}, year = {2025}, }
• (2025). Revived transformation-based transfer learning safety control: A geometric interpretation. NEUCOM, 653, 131213. (www)

  @article{ , abstract = {Ensuring safety during the learning process of reinforcement learning (RL) control poses a significant challenge. To address this issue, our previous work proposed a novel transfer learning approach, allowing controllers to be trained in a non-dangerous source system and then transferred to the dangerous target system. This approach is built upon a system transformation–termed the revived transformation –that maps the source system to the target system. In this work, we develop a rigorous geometric framework to characterize the intrinsic manifold structures induced by the revived transformation, which link the source and target systems. These geometric insights provide a precise theoretical foundation for analyzing how control policies can be reliably transferred between systems. Specifically, we show that the revived transformation induces a diffeomorphic mapping between the state manifolds of the source and target systems, thereby yielding a structured relationship between their respective vector fields. Furthermore, this relationship establishes a connection between the one-parameter transformation groups acting on the two state manifolds. These geometric structures ultimately ensure the safety and stability of the target system under the transferred controller. Finally, simulation results validate the correctness of the geometric framework as well as the transfer learning approach.}, archive = {J_NEUCOM}, author = {Quanqi Zhang and Chengwei Wu and Yupeng Zhu}, doi = {10.1016/j.neucom.2025.131213}, journal = {Neurocomputing}, month = {11}, pages = {131213}, shortjournal = {Neurocomputing}, title = {Revived transformation-based transfer learning safety control: A geometric interpretation}, volume = {653}, year = {2025}, }
• (2025). Search and rescue operations in wildfires using unmanned aerial vehicles: A multi-agent deep reinforcement learning approach. NEUCOM, 653, 131211. (www)

  @article{ , abstract = {Wildfires pose major challenges to natural ecosystems and smart living due to their destructive nature. Unmanned Aerial Vehicles (UAVs) or drones have been used to support fire fighter in identifying vulnerable areas and detecting people who need assistance. Most of the current solutions use path planning approaches under simple and deterministic environments that fail to model the dynamic nature of fire. Furthermore, the localisation of victims is assumed to be known which is unrealistic in disaster-like scenarios. To alleviate this issue, this paper proposes a novel search and rescue (SAR) application using drones. A multi-agent deep Q-network is designed to train a fleet of UAVs to search for people and evacuate them in a wildfire scenario. A realistic forest environment is designed that considers variations in vegetation and fire propagation. This helps to challenge RL algorithms to be more adaptive to changes in the environment due to the evolution of fire. Extensive simulation experiments are conducted to show the advantages and effectiveness of the proposed approach.}, archive = {J_NEUCOM}, author = {Maxime Collignon and Adolfo Perrusquía and Antonios Tsourdos and Weisi Guo}, doi = {10.1016/j.neucom.2025.131211}, journal = {Neurocomputing}, month = {11}, pages = {131211}, shortjournal = {Neurocomputing}, title = {Search and rescue operations in wildfires using unmanned aerial vehicles: A multi-agent deep reinforcement learning approach}, volume = {653}, year = {2025}, }
• (2025). PAMG:Position-aware and multi-grained context fusion for optimized table-to-text generation. NEUCOM, 653, 131210. (www)

  @article{ , abstract = {Table-to-text generation is a central task in the field of natural language processing, automating the transformation of structured data such as spreadsheets into natural language descriptions. However, existing models often face issues with weak controllability and poor structural understanding when dealing with complex tables. To address these challenges, we propose Position-Aware and Multi-Grained Context Fusion for Optimized Table-to-Text Generation, namely PAMG. This approach enhances the T5 model through several key improvements. First, we introduce two prefixes to the T5 encoder to enhance the model’s interpretability and controllability. Second, we propose explicit position encoding and multi-grained context modeling to improve the model’s capability to capture structural and contextual information from tables. Finally, during the decoding time step, we incorporate a copying mechanism to ensure the accuracy of the generated text. Evaluation on the WikiBio dataset across various domains (Humans, Songs, and Books) and the ToTTo dataset, using both automatic and human evaluation, demonstrates substantial improvements over baseline methods. This offers a novel solution to the table-to-text generation task.}, archive = {J_NEUCOM}, author = {Siwei Wang and Xiaohong Li and Jiahao Liu and Yang Han}, doi = {10.1016/j.neucom.2025.131210}, journal = {Neurocomputing}, month = {11}, pages = {131210}, shortjournal = {Neurocomputing}, title = {PAMG:Position-aware and multi-grained context fusion for optimized table-to-text generation}, volume = {653}, year = {2025}, }
• (2025). A distance for mixed-variable and hierarchical domains with meta variables. NEUCOM, 653, 131208. (www)

  @article{ , abstract = {Heterogeneous datasets emerge in various machine learning and optimization applications that feature different input sources, types or formats. Most models or methods do not natively tackle heterogeneity. Hence, such datasets are often partitioned into smaller and simpler ones, which may limit the generalizability or performance, especially when data is limited. The first main contribution of this work is a modeling framework that generalizes hierarchical, tree-structured, variable-size or conditional search frameworks. The framework models mixed-variable and hierarchical domains in which variables may be continuous, integer, or categorical, with some identified as meta when they influence the structure of the problem. The second main contribution is a novel distance that compares any pair of mixed-variable points that do not share the same variables, allowing to use of whole heterogeneous datasets that reside in mixed-variable and hierarchical domains with meta variables. The contributions are verified empirically through regression and classification experiments using simple distance-based models applied to datasets of hyperparameters with corresponding performance scores.}, archive = {J_NEUCOM}, author = {Edward Hallé-Hannan and Charles Audet and Youssef Diouane and Sébastien Le Digabel and Paul Saves}, doi = {10.1016/j.neucom.2025.131208}, journal = {Neurocomputing}, month = {11}, pages = {131208}, shortjournal = {Neurocomputing}, title = {A distance for mixed-variable and hierarchical domains with meta variables}, volume = {653}, year = {2025}, }
• (2025). NSGHG: Neural surface guided generalizable human gaussian splatting for sparse view synthesis. NEUCOM, 653, 131207. (www)

  @article{ , abstract = {We propose Neural Surface Guided Generalizable Human Gaussian Splatting (NSGHG), a framework for synthesizing high-fidelity novel views of unseen human subjects from highly sparse input images, without tedious optimization or fine-tuning. Existing generalizable methods either struggle with sparsely sampled views with limited overlap or perform poorly when facing complex human geometry. To address these issues, we propose to employ a neural surface estimation module that predicts occupancy to guide the adaptive displacement of a human template toward the true surface, generating surface-aligned multi-shells and regressing human Gaussians from their surface to more efficiently capture complex human geometry. Furthermore, we propose a double-geometry MLP architecture that shares geometric feature embeddings between the neural surface estimator and the Gaussian parameter regressor, facilitating finer-grained geometry and appearance modeling of the human Gaussians. Moreover, NSGHG supports end-to-end joint training of both the neural surface estimation and the Gaussian parameter regression modules. Extensive experiments on multiple benchmark datasets show that our method significantly outperforms state-of-the-art techniques in both novel view synthesis quality and cross-dataset generalization.}, archive = {J_NEUCOM}, author = {Hong Xie and Xiaoyan Zhang and Yingying Zhu}, doi = {10.1016/j.neucom.2025.131207}, journal = {Neurocomputing}, month = {11}, pages = {131207}, shortjournal = {Neurocomputing}, title = {NSGHG: Neural surface guided generalizable human gaussian splatting for sparse view synthesis}, volume = {653}, year = {2025}, }
• (2025). Nesterov-accelerated non-negative matrix factorization unrolling network for hyperspectral unmixing. NEUCOM, 653, 131202. (www)

  @article{ , abstract = {Deep learning is making notable progress in hyperspectral unmixing (HU), with algorithm unrolling networks providing advantages in interpretability compared to traditional networks. These networks mainly focused on leveraging prior knowledge from conventional algorithms to enhance their performance. However, research on improving network learning efficiency remains limited, specifically in terms of accuracy of endmember and abundance estimation, as well as the number of iterations and training time. In this paper, we propose a Nesterov-accelerated non-negative matrix factorization unrolling network (NANMF-Net) for HU, which not only improves unmixing accuracy but also reduces the number of training iterations. Specifically, NANMF-Net adopts a two-stage network structure. The first stage uses an unrolling NMF structure, which focuses on learning relevant information. The second stage applies the Nesterov’s method to optimize the network, reinforcing the connections among layers. These two stages work together to achieve more accurate endmember and abundance estimates with fewer training iterations. NANMF-Net demonstrated superior unmixing performance and fewer iterations on both synthetic and real datasets compared to some traditional models and networks.}, archive = {J_NEUCOM}, author = {Sheng Shu and Ting-Zhu Huang and Jie Huang and Gemine Vivone}, doi = {10.1016/j.neucom.2025.131202}, journal = {Neurocomputing}, month = {11}, pages = {131202}, shortjournal = {Neurocomputing}, title = {Nesterov-accelerated non-negative matrix factorization unrolling network for hyperspectral unmixing}, volume = {653}, year = {2025}, }
• (2025). DFCT-net for automatic modulation recognition in UAV communication systems. NEUCOM, 653, 131200. (www)

  @article{ , abstract = {With the increasing frequency range of broadband communication signals and more complex electromagnetic environments, detecting and recognizing low signal-to-noise ratio signals poses greater challenges to existing technologies. This article proposes a Dual Flow Convolutional Transformer Network (DFCT Net) for modulation recognition in unmanned aerial vehicle communication systems, which extracts and combines features from both time and frequency domains. Firstly, we propose the Convolutional LSTM and Channel Enhancement Network (CLCE-Net) to adaptively enhance and extract temporal and spatial features in the time domain. Secondly, we combine convolutional neural networks with Transformers to efficiently recognize low signal-to-noise ratio signals in the frequency domain. Furthermore, the signal spectrum is denoised using an unsupervised convolutional autoencoder. Experiments on the Radio-Mod-5.8G-160K and HisarMod2019.1 datasets demonstrate that DFCT Net achieves state-of-the-art performance with improved parameter efficiency, highlighting its potential for practical unmanned aerial vehicle communication systems.}, archive = {J_NEUCOM}, author = {Yantian Shen and Yang Yang and Yunxia Liu and Na Wang and Liang Ma and Hongjun Wang and Fang Li}, doi = {10.1016/j.neucom.2025.131200}, journal = {Neurocomputing}, month = {11}, pages = {131200}, shortjournal = {Neurocomputing}, title = {DFCT-net for automatic modulation recognition in UAV communication systems}, volume = {653}, year = {2025}, }
• (2025). CorAFR: Arbitrary-time frame reconstruction from motion-blurred and correlation images. NEUCOM, 653, 131199. (www)

  @article{ , abstract = {This paper proposes a deep-learning solution based on a three-phase correlation image sensor (3PCIS) to recover consecutive sharp frames from a single motion-blurred image at any given moment within the exposure time. The correlation image generated by 3PCIS is the temporal correlation between incident light and reference signals, carrying motion information about the speed and direction of moving objects during the exposure time. This is crucial for restoring clear texture details. Therefore, we combine correlation images and motion-blurred images as inputs, and build a two-stream network with temporal encoding to restore sharp frames at any given moment. Two key designs in our model are 1) a Double-gated Feature Fusion (DGFF), facilitating effective fusion of complementary information for high-quality outputs; 2) a Gaussian-based Temporal Encoding (GBTE), transforming arbitrary time within exposure time into feature-like representations for establishing correspondence between time and restored frames. Our model can produce any number of sharp frames by changing time. Experimental results demonstrate that our solution significantly outperforms the state-of-the-art methods on public datasets and real-world scenarios.}, archive = {J_NEUCOM}, author = {Pan Wang and Toru Kurihara and Jun Yu}, doi = {10.1016/j.neucom.2025.131199}, journal = {Neurocomputing}, month = {11}, pages = {131199}, shortjournal = {Neurocomputing}, title = {CorAFR: Arbitrary-time frame reconstruction from motion-blurred and correlation images}, volume = {653}, year = {2025}, }
• (2025). Reinforcement learning-based affine formation maneuver control of autonomous surface vehicles with safety constraints under time-varying actuator faults. NEUCOM, 653, 131196. (www)

  @article{ , abstract = {As a novel strategy, the affine formation maneuver control (AFMC) employs stress matrices and affine transformation theory to simultaneously realize various leader–follower formation maneuvers including collinearity, rotation, scaling, shearing, and translation. So far, only a few publications have reported the AFMC of multiple autonomous surface vehicles (ASVs). This study aims to explore the leader–follower reinforcement learning-based fractional-order sliding mode AFMC (RL-based FOSMAFMC) approach for ASVs with safety constraints subjected to time-varying actuator faults. A barrier Lyapunov function (BLF) is utilized to keep the ASVs operating within a constrained safety range. The problem of loss of control direction due to the time-varying actuator effectiveness faults is addressed using the Nussbaum function. The RL is formulated using the actor–critic framework. The actor approximates the uncertain nonlinearities and time-varying bias faults, whereas the critic system assesses the control action while minimizing a long-term cost function. The leader–follower AFMC uses a fractional-order sliding mode surface to enhance tracking accuracy and robustness. A Lyapunov function is used to guarantee the boundedness of the closed-loop system. Simulation experiments and performance comparisons are carried out to demonstrate the capability of the proposed controller.}, archive = {J_NEUCOM}, author = {Muhammad Maaruf and Abdulrahman Aliyu and Luai M. Alhems and Muhammad Majid Gulzar}, doi = {10.1016/j.neucom.2025.131196}, journal = {Neurocomputing}, month = {11}, pages = {131196}, shortjournal = {Neurocomputing}, title = {Reinforcement learning-based affine formation maneuver control of autonomous surface vehicles with safety constraints under time-varying actuator faults}, volume = {653}, year = {2025}, }
• (2025). New generalized integral inequality and its application in exponential stability of impulsive neutral neural networks driven by G-brownian motion. NEUCOM, 653, 131195. (www)

  @article{ , abstract = {Accompanying with the comment on the article [ Inequalities and stability of stochastic Hopfield neural networks with discrete and distributed delays, D. Ruan, Z. Huang and X. Guo, Neurocomputing, 407 (2020) 281–291 ], we provide a new proof of the Lemma 1 as correction or modification to the proof of Lemma 3.1 in the mentioned article. Then, in this work, we have also demonstrated a new generalized multidimensional inequality and a new generalized Halanay inequality to be applied in the case of the model with impulse effects and neutral type terms. Moreover, we have investigated the global mean-square exponential stability of an impulsive neural net system driven by the G-Brownian motion to show the usefulness of our generalized inequality. Two examples and their simulations are illustrated to verify our theoretical results.}, archive = {J_NEUCOM}, author = {Puchen Liu and Guoyan Cao and Haidong Liu and Yan Jin}, doi = {10.1016/j.neucom.2025.131195}, journal = {Neurocomputing}, month = {11}, pages = {131195}, shortjournal = {Neurocomputing}, title = {New generalized integral inequality and its application in exponential stability of impulsive neutral neural networks driven by G-brownian motion}, volume = {653}, year = {2025}, }
• (2025). VDMPO: Policy optimization for cooperative multi-agent reinforcement learning based on joint value decomposition. NEUCOM, 653, 131193. (www)

  @article{ , abstract = {Low collaboration efficiency and poor task performance have long been significant challenges in the field of cooperative Multi-Agent Reinforcement Learning (MARL). This paper proposes a Value Decomposition Multi-Policy Optimization (VDMPO) method. This method designs a new network structure that incorporates agent networks based on the actor-critic framework, along with an upper-level mixing network, supporting both discrete and continuous action space cooperative tasks. The upper-level mixing network decomposes the joint state value into the value of each agent and implements a team reward allocation mechanism based on value fitting. This effectively addresses the difficulty of designing rewards for individual agents in cooperative MARL and improves the collaboration efficiency among agents. Additionally, a special target value network is designed in this method to enhance the stability of network training. In typical experimental scenarios, VDMPO significantly outperforms state-of-the-art (SOTA) MARL methods (e.g., MAPPO, QMIX, MADDPG, MATD3) in terms of average performance, demonstrating that VDMPO can better achieve cooperation and coordination among multiple agents in complex tasks.}, archive = {J_NEUCOM}, author = {Ze Wang and Ni Li and Guanghong Gong}, doi = {10.1016/j.neucom.2025.131193}, journal = {Neurocomputing}, month = {11}, pages = {131193}, shortjournal = {Neurocomputing}, title = {VDMPO: Policy optimization for cooperative multi-agent reinforcement learning based on joint value decomposition}, volume = {653}, year = {2025}, }
• (2025). Modeling events and interactions through temporal processes: A survey. NEUCOM, 653, 131191. (www)

  @article{ , abstract = {In real-world scenarios, numerous phenomena generate a series of events that occur in continuous time. Point processes provide a natural mathematical framework for modeling these event sequences. In this comprehensive survey, we aim to explore probabilistic models that capture the dynamics of event sequences through temporal processes. We revise the notion of event modeling and provide the mathematical foundations that underpin the existing literature on this topic. To structure our survey effectively, we introduce an ontology that categorizes the existing approaches considering three horizontal axes: modeling , inference and estimation , and application . We conduct a systematic review of the existing approaches, with a particular focus on those leveraging deep learning techniques. Finally, we delve into the practical applications where these proposed techniques can be harnessed to address real-world problems related to event modeling. Additionally, we provide a selection of benchmark datasets that can be employed to validate the approaches for point processes.}, archive = {J_NEUCOM}, author = {Angelica Liguori and Luciano Caroprese and Marco Minici and Bruno Veloso and Francesco Spinnato and Mirco Nanni and Giuseppe Manco and João Gama}, doi = {10.1016/j.neucom.2025.131191}, journal = {Neurocomputing}, month = {11}, pages = {131191}, shortjournal = {Neurocomputing}, title = {Modeling events and interactions through temporal processes: A survey}, volume = {653}, year = {2025}, }
• (2025). Concept drift detection based on radial distance. NEUCOM, 653, 131190. (www)

  @article{ , abstract = {With the advancement of powerful computational hardware in recent years, neural networks, once popular for machine learning on pre-collected datasets, are becoming applicable for streaming data processing. Stream data have the characteristics of high velocity, variety and volume. In stream data mining a common challenge is concept drift, which refers to the phenomenon where the statistical properties of the target variable in predictive tasks change over time. For instance, in image prediction problems, the input facilities may be altered by the environment or device flaws. The materials generated from them could be distorted, such as blurring, discoloring or part-missing. Concept drift problem is considered a root cause of performance degradation in machine learning models on stream data. Traditional concept drift detection methods usually require large amount of historical data, which leads to substantial memory footprint and high computational cost, and tend to be overly sensitive to arbitrary distribution changes not related to prediction results. Such limitations are particularly evident in settings using neural network models, where input data are usually high dimensional images or videos. Aiming to improve the accuracy and efficiency of concept drift detection in neural network models, we propose a new concept drift detection method that specifically addresses these limitations and is applicable to general neural network models. Our method represents the original data with a distance-based statistic, extracted from the layer outputs of the neural network models, and is able to adjust its sensitivity to input distribution changes based on their relevance to neural network features. We evaluated our method with popular neural network architectures on both synthetic and real-world data sets. The results showed our method not only outperforms existing concept drift methods in accuracy, but is also significantly faster and consumes less resources.}, archive = {J_NEUCOM}, author = {Dan Shang and Guangquan Zhang and Jie Lu}, doi = {10.1016/j.neucom.2025.131190}, journal = {Neurocomputing}, month = {11}, pages = {131190}, shortjournal = {Neurocomputing}, title = {Concept drift detection based on radial distance}, volume = {653}, year = {2025}, }
• (2025). A universal sampling method based on feature and structural comprehensive proximity measure. NEUCOM, 653, 131189. (www)

  @article{ , abstract = {Sampling is a widely employed technique to tackle the neighbor explosion problem in Graph Neural Networks (GNNs). However, existing sampling methods fail to adequately consider both the feature and structural information of the graph, limiting their adaptability across different GNN models. To address these issues, we introduce a universal Feature-Structure Sampling (FSS) method based on a comprehensive proximity measure that effectively integrates both feature and structural information. Specifically, FSS utilizes a graph attention network to extract feature information from the graph and introduces the Hodge score, derived from the Hodge decomposition, to capture structural information. Then, FSS samples the top k most important nodes using the comprehensive proximity measure. Moreover, we provide theoretical evidence that the Hodge score effectively represents the global importance of the graph’s structure. We provide both theoretical and experimental evidence that FSS offers the following advantages: (1) Efficient Sampling : Sampling the top k nodes in the neighborhood mitigates the neighbor explosion problem in deep Graph Neural Networks and dense graphs, resulting in accelerated computations; (2) Enhanced Information Retention : The integration of feature and structural information enables FSS to preserve most significant information within the neighborhood while discarding irrelevant data, thereby enhancing model performance; (3) Plug-and-Play Compatibility : FSS is designed to be plug-and-play, allowing for efficient integration with existing GNN models ( e.g. , GCN, GAT) to improve both training speed and overall performance. Extensive experimental results demonstrate the effectiveness of our sampling method.}, archive = {J_NEUCOM}, author = {Jinhui Pang and Cheng Shang and Ziyu Jia and Peng Hao and Weilong Wang and Xiaoshuai Hao}, doi = {10.1016/j.neucom.2025.131189}, journal = {Neurocomputing}, month = {11}, pages = {131189}, shortjournal = {Neurocomputing}, title = {A universal sampling method based on feature and structural comprehensive proximity measure}, volume = {653}, year = {2025}, }
• (2025). MVF-PointCLIP: Training-free multi-view fusion PointCLIP for zero-shot 3D classification. NEUCOM, 653, 131188. (www)

  @article{ , abstract = {The remarkable success of Contrastive Language-Image Pretraining (CLIP) in zero-shot 2D vision classification inspires researchers to explore its potential application to zero-shot 3D classification. Some researchers project 3D point clouds into 2D images from multiple views to leverage CLIP. However, by doing experiments, we find that this method suffers from two critical drawbacks: (1) noise views existing in multiview depth maps, which provide limited information that may mislead classification; (2) covariance inconsistencies between sample views, which can lead to misclassification when using cosine similarity. To address these issues, we propose a training-free MultiView Fusion PointCLIP (MVF-PointCLIP). It contains a Spatial and Frequency Attention (SFA) module and a Mahalanobis Distance module designed by us. The SFA module automatically assigns importance weights to views, effectively filtering out noisy information. The Mahalanobis Distance module models the distribution of views to tackle covariance inconsistencies. Experimental results verify the superiority of MVF-PointCLIP to SOTA models in zero-shot classification across ModelNet10, ModelNet40, and ScanObjectNN.}, archive = {J_NEUCOM}, author = {Jiuqian Dai and Zhenyan Ji and Zechang Xiong and Guiping Zhu and Hui Liu and Shen Yin and Jose Enrique Armendariz-Inigo}, doi = {10.1016/j.neucom.2025.131188}, journal = {Neurocomputing}, month = {11}, pages = {131188}, shortjournal = {Neurocomputing}, title = {MVF-PointCLIP: Training-free multi-view fusion PointCLIP for zero-shot 3D classification}, volume = {653}, year = {2025}, }
• (2025). Hybrid serial-parallel photonic reservoir computing based on single semiconductor laser with delayed optical feedback for multivariable task processing. NEUCOM, 653, 131187. (www)

  @article{ , abstract = {Photonic reservoir computing (RC) is an emerging neural network approach with advantages of low training power consumption, minimal hardware requirements, high processing speed, and potential for physical implementation. However, existing research lacks a systematic solution for the common challenge of multivariable collaborative processing in practical applications. In this study, we propose a hybrid serial-parallel photonic RC scheme based on a semiconductor laser (SL) with delayed optical feedback to address multivariable tasks. This approach constructs a global data input stream and explores its computational potential in handling complex multivariable tasks through a nonlinear central node with an internal multi-feedback mode. To validate the proposed method, we designed a verification system for gas type recognition and concentration prediction, focusing on a representative multivariable scenario of mixed gas detection. We have verified the effectiveness of the proposed method through numerical simulations and conducted a detailed analysis of the effects of system parameters on performance. The superior performance of the proposed method was finally verified as compared with the existing methods. By integrating the physical characteristics of optical devices with the high-speed computational capabilities of RC, the proposed method effectively compensates for the slow response kinetics of chemical sensor arrays. Its simple structure makes it highly adaptable, allowing for further extension to other multivariable processing tasks.}, archive = {J_NEUCOM}, author = {Guizheng Guan and Bin Liu}, doi = {10.1016/j.neucom.2025.131187}, journal = {Neurocomputing}, month = {11}, pages = {131187}, shortjournal = {Neurocomputing}, title = {Hybrid serial-parallel photonic reservoir computing based on single semiconductor laser with delayed optical feedback for multivariable task processing}, volume = {653}, year = {2025}, }
• (2025). Distributed impulsive control for mean-square exponential synchronization of coupled stochastic delayed neural networks under intermittent couplings. NEUCOM, 653, 131186. (www)

  @article{ , abstract = {This paper focuses on addressing the problem of mean-square exponential synchronization (MSES) for coupled stochastic delayed neural networks (CSDNNs) under intermittent coupling, employing a distributed impulsive control approach. Unlike conventional continuous coupling schemes, intermittent coupling reflects scenarios where network nodes interact only during discrete time intervals, a phenomenon commonly observed in practical systems. The study formulates models of CSDNNs incorporating both delayed and non-delayed intermittent couplings along with impulsive effects. To tackle the complexities introduced by these intermittent interactions and impulsive events, two types of piecewise impulsive time-delay differential inequalities are developed to describe the system dynamics across coupling intervals, non-coupling periods, and impulsive instants. A distributed impulsive control protocol is subsequently designed for the intermittently coupled CSDNNs. By integrating the derived inequalities with graph theory, the paper rigorously investigates the MSES problem and establishes synchronization criteria. Finally, the theoretical results are corroborated through two numerical examples, demonstrating the proposed method’s effectiveness and practicality.}, archive = {J_NEUCOM}, author = {Siyi Xiong and Yinfang Song and Tao Zhang and Shiping Wen}, doi = {10.1016/j.neucom.2025.131186}, journal = {Neurocomputing}, month = {11}, pages = {131186}, shortjournal = {Neurocomputing}, title = {Distributed impulsive control for mean-square exponential synchronization of coupled stochastic delayed neural networks under intermittent couplings}, volume = {653}, year = {2025}, }
• (2025). Pseudo-intervention based local-to-global causal structure learning. NEUCOM, 653, 131185. (www)

  @article{ , abstract = {Causal structure learning (CSL) often encounters complex computational challenges, whereas local-to-global approaches facilitate making it more scalable to high-dimensional data. Traditional causal structure learning algorithms often obtain more detailed causal relationships by designing complex intervention experiments, but these are computationally expensive and may even involve ethical issues. Moreover, the existing data may usually lead to local skeleton asymmetry due to noise or small sample problems, which can affect the final causal structure. To address these issues, a local-to-global causal structure learning algorithm based on pseudo-intervention is proposed. Firstly, a pseudo-intervention scheme and corresponding selection strategies are presented, which simulate different intervention scenarios by generating multiple datasets from original data and selecting appropriate sub-datasets to learn local skeletons without relying on a large number of costly intervention experiments. Then, a comprehensive scoring function is designed to weight the causal skeletons obtained from different sub-datasets for constructing a weighted global skeleton, which involves similarity weight and quantitative weight to comprehensively evaluate the representativeness and sufficiency of each pseudo-intervention sub-dataset. Thereby, the learning accuracy and robustness of the causal graph are enhanced. Finally, weighted edges are filtered based on different thresholds to overcome the problem of missed or redundant edges caused by using a single rule. Ultimately, the global causal structure is constructed following a score-search strategy for orientation. The experimental results show that the proposed method outperforms existing advanced algorithms on multiple datasets.}, archive = {J_NEUCOM}, author = {Dingyuan Liu and Yaling Xun and Haifeng Yang and Jianghui Cai and Yanfeng Li}, doi = {10.1016/j.neucom.2025.131185}, journal = {Neurocomputing}, month = {11}, pages = {131185}, shortjournal = {Neurocomputing}, title = {Pseudo-intervention based local-to-global causal structure learning}, volume = {653}, year = {2025}, }
• (2025). A shapelet-driven distillation generation method for generalized zero-shot learning in compound fault diagnosis. NEUCOM, 653, 131184. (www)

  @article{ , abstract = {Acquiring comprehensive compound fault data for rotating machinery is costly and limited, posing significant challenges for deep learning-based intelligent diagnosis. Zero-shot learning (ZSL) addresses this issue by leveraging seen-class data to construct class-level semantics, enabling the recognition of unseen compound faults without target samples. However, existing ZSL methods suffer from semantic deficiencies, limited robustness, and weak orthogonality, reducing their effectiveness when single and compound faults coexist. Therefore, this paper proposes a shapelet-driven distillation generation (SDDG) method within the generalized zero-shot learning (GZSL) framework to enhance fault recognition and improve generalization to unseen fault classes. SDDG enhances semantic discrimination and robustness, while mitigating inter-class interference, by incorporating sampling order optimization and location sensitivity metrics into an information gain-driven shapelet screening process. Furthermore, a co-optimization strategy for feature generation and distribution refinement is introduced, ensuring representation consistency and improved generalization through feature alignment loss, logical representation space mapping, and classification loss. Experimental results on public and lab-built datasets demonstrate that SDDG outperforms state-of-the-art methods, achieving superior performance in compound fault diagnosis under the GZSL framework.}, archive = {J_NEUCOM}, author = {Shuangyan Yin and Jingli Yang and Yongqi Chang and Ye Li and Changdong Wang}, doi = {10.1016/j.neucom.2025.131184}, journal = {Neurocomputing}, month = {11}, pages = {131184}, shortjournal = {Neurocomputing}, title = {A shapelet-driven distillation generation method for generalized zero-shot learning in compound fault diagnosis}, volume = {653}, year = {2025}, }
• (2025). Multi-view clustering for localized global time series forecasting. NEUCOM, 653, 131183. (www)

  @article{ , abstract = {Recently, clustering has been widely used to localize global forecasting models and address data heterogeneity. Most prior work has focused on single-view clustering approaches, where constructing effective representations of time series data is critical. Inspired by advances in multi-view clustering and time series imaging techniques, and aiming to leverage the complementary information captured from multiple views of time series data, this study proposes a global forecasting methodology based on multi-view clustering (MVC) to enhance clustering robustness and improve forecasting accuracy. Specifically, we introduce an optimal fuzzy multi-view clustering algorithm (OFMVC) as part of the proposed forecasting framework. OFMVC includes an intelligent view selection mechanism that leverages diversity and manifold learning to identify views that are both complementary and representative of the dataset’s inherent geometric structure. Experiments on the six M3 monthly datasets show that forecasting models built using OFMVC outperform all single-view clustering methods. In terms of mean symmetric Mean Absolute Percentage Error (sMAPE), the proposed model also exceeds a baseline multi-view clustering method, with a performance gain of 0.21. Moreover, the proposed approach outperforms several state-of-the-art methods reported in the literature, including univariate and deep learning-based time series forecasting models. The source code and data for this article are located at: https://github.com/AliReza000J/MVC-Time-series-Forecasting .}, archive = {J_NEUCOM}, author = {Hossein Abbasimehr and Alireza Abri and Amin Golzari Oskouei and Ali Noshad}, doi = {10.1016/j.neucom.2025.131183}, journal = {Neurocomputing}, month = {11}, pages = {131183}, shortjournal = {Neurocomputing}, title = {Multi-view clustering for localized global time series forecasting}, volume = {653}, year = {2025}, }
• (2025). A novel supervised framework for multi-video summarization: Addressing dataset biases and enhancing feature representation. NEUCOM, 653, 131180. (www)

  @article{ , abstract = {The exponential growth of video content necessitates effective summarization and information retrieval strategies, particularly in multi-video contexts involving heterogeneous sources. However, existing video summarization datasets often suffer from limitations such as class imbalance, content redundancy, and subjectivity in annotations, which hinder the development and evaluation of robust summarization models. In this work, we propose a novel supervised framework for extractive multi-video summarization designed to address these challenges. The framework integrates clustering, classification, and scoring mechanisms to improve data quality and generate summaries that are both relevant and diverse. Clustering is utilized to minimize redundancy, and two distinct labeling strategies are introduced to construct four curated datasets that strike a balance between annotation strictness and redundancy reduction. To effectively capture the salient characteristics of video segments, we extract a rich set of features encompassing spatio-temporal descriptors, color histograms, and Gray Level Co-occurrence Matrix (GLCM) attributes. Comprehensive evaluations conducted on the Tour20 dataset demonstrate the effectiveness of the proposed approach, achieving a mean F1-score of 0.774. This represents a 5 % improvement over the best-performing baseline, a graph-based centrality method (F1-score: 0.741), and an average relative improvement of 21 % over all compared methods.}, archive = {J_NEUCOM}, author = {Aziz M. Qaroush and Mohammad Jubran and Qutaiba Olayyan}, doi = {10.1016/j.neucom.2025.131180}, journal = {Neurocomputing}, month = {11}, pages = {131180}, shortjournal = {Neurocomputing}, title = {A novel supervised framework for multi-video summarization: Addressing dataset biases and enhancing feature representation}, volume = {653}, year = {2025}, }
• (2025). Dynamic event-triggered regional synchronization for discrete-time delayed dynamical networks: Dealing with saturating actuators. NEUCOM, 653, 131179. (www)

  @article{ , abstract = {This paper considers the regional synchronization problem for discrete-time dynamical networks containing node delays and disturbance inputs subject to saturating actuators. In order to decrease the usage of communication resources, an event-triggered scheme with dynamic mechanism is adopted to regulate the data transmission. By incorporating the modified sector condition, the Lyapunov–Krasovskii functional and the discrete Jensen inequality, a sufficient condition is first established such that the synchronization error dynamics has the desirable properties including boundedness, H ∞ performance and regional stability. Then, the design parameters are dedicatedly characterized based on the solvability of linear matrix inequalities. Furthermore, certain optimization problems concerning the event-triggered rate and the initial condition set are formulated. Finally, two numerical examples are presented to illustrate the feasibility and advantages of the proposed results.}, archive = {J_NEUCOM}, author = {Yonggang Chen and Hongjian Liu and Jun Hu and Hong Lin}, doi = {10.1016/j.neucom.2025.131179}, journal = {Neurocomputing}, month = {11}, pages = {131179}, shortjournal = {Neurocomputing}, title = {Dynamic event-triggered regional synchronization for discrete-time delayed dynamical networks: Dealing with saturating actuators}, volume = {653}, year = {2025}, }
• (2025). Decoupled likelihood modeling: A scalable approach for incremental generalized category discovery. NEUCOM, 653, 131177. (www)

  @article{ , abstract = {Incremental Generalized Category Discovery (IGCD) aims to enable intelligent systems to adaptively identify and learn new categories from unsupervised data streams while preserving prior knowledge. Such data streams often exhibit significant class imbalance, where new class samples gradually dominate while data from previously learned classes are scarce or unavailable. This imbalance poses two critical challenges for conventional joint posterior modeling: (1) the decision boundaries established by old classes impede the discovery of new classes, and (2) the addition of new classes and the inaccessibility of old classes result in catastrophic forgetting. To tackle these issues, we propose a Decoupled Likelihood Modeling (DLM) framework that leverages Variational Autoencoders (VAEs) to independently model the likelihood distributions of each class, demonstrating strong scalability under large-scale class settings. This decoupling isolates the modeling of new classes from previously learned ones without interference to prior knowledge. Furthermore, the introduction of a class-balanced prior ensures fair decision boundaries, thereby mitigating misclassification caused by class imbalance. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that DLM achieves state-of-the-art performance, substantially improving new class discovery accuracy and classification fairness. 1}, archive = {J_NEUCOM}, author = {Siyu Zhang and Wenju Sun and Yangli-ao Geng and Qingyong Li and Wen Wang}, doi = {10.1016/j.neucom.2025.131177}, journal = {Neurocomputing}, month = {11}, pages = {131177}, shortjournal = {Neurocomputing}, title = {Decoupled likelihood modeling: A scalable approach for incremental generalized category discovery}, volume = {653}, year = {2025}, }
• (2025). Unsupervised meta-learning for few-shot classification by progressively enhancing task diversity. NEUCOM, 653, 131176. (www)

  @article{ , abstract = {Unsupervised meta-learning (UML) aims to perform meta-learning using a large, unlabeled auxiliary dataset to address the few-shot classification problem. Existing UML methods are mainly based on augmentation-based and clustering-based strategies. However, these methods struggle to enhance task diversity while preserving task reliability throughout the meta-training process. In this study, we propose a UML method, Progressively Enhancing Task Diversity (PETD), which integrates both strategies into an end-to-end model to construct more diverse and reliable tasks for training. Furthermore, considering the insufficient capability of the model in the early training stage, inspired by the human learning process from simple to complex, we introduce a progressive diversification strategy to progressively enhance task diversity while ensuring reliability by introducing an extra increasing probability during the training. We conduct extensive in-domain experiments on three public datasets: mini-ImageNet, CIFAR-FS, tiered-ImageNet, and more challenging cross-domain experiments on the CUB dataset to validate the effectiveness of PETD. Experimental results demonstrate that PETD consistently outperforms previous UML methods. For example, PETD achieves a significant improvement of 6.88 % and 4.34 % respectively for the 5-way 1-shot and 5-shot tasks of CIFAR-FS over the prior best method. The code will be available at https://github.com/ZWB-1224171056/PETD .}, archive = {J_NEUCOM}, author = {WanBing Zhao and JunHua Liao and KangHui Feng and LiangYin Chen and YanRu Chen}, doi = {10.1016/j.neucom.2025.131176}, journal = {Neurocomputing}, month = {11}, pages = {131176}, shortjournal = {Neurocomputing}, title = {Unsupervised meta-learning for few-shot classification by progressively enhancing task diversity}, volume = {653}, year = {2025}, }
• (2025). Path planning based on improved deep Q-network algorithm for bionic robotic fish with ocean currents. NEUCOM, 653, 131173. (www)

  @article{ , abstract = {Path planning is a fundamental issue for bionic robotic fish (BRF) to complete the missions such as underwater resource exploration, cable inspection, water quality detection and so on. The existing challenge is that how to design an efficient optimal path with uncertainty of ocean currents for the BRF working in the complex underwater three-dimensional (3D) environment. This paper proposes an improved deep Q-network (DQN) algorithm to improve the ability of BRF path planning in the unknown complex environment with dynamic ocean currents. A dynamic integrated reward mechanism is proposed based on the environmental information of terrains, obstacles, ocean currents and so on, which is used to balance multiple objective functions for the path planning problem. It presents a dynamic two-step action-selection strategy for the improved DQN algorithm. For avoiding premature convergence to the suboptimal path in the early stage, Boltzmann strategy is used to balance high-value actions and new actions in the first step; considering the factors such as ocean currents in the later stage, ϵ -greedy strategy is employed to exploit known information efficiently and design the path in the second step. A double dynamic learning rate method is proposed in light of the meta-gradient descent and Adam optimization methods, which is used to optimize the learning rate at two levels of hyperparameter and model parameter. Simulation results demonstrate the effectiveness of the proposed algorithm for the BRF path planning in the complex environment with dynamic ocean currents.}, archive = {J_NEUCOM}, author = {Qunhong Tian and Jialin Li and Guangtao Ran and Hongyu Li and Weizhuang Ma}, doi = {10.1016/j.neucom.2025.131173}, journal = {Neurocomputing}, month = {11}, pages = {131173}, shortjournal = {Neurocomputing}, title = {Path planning based on improved deep Q-network algorithm for bionic robotic fish with ocean currents}, volume = {653}, year = {2025}, }
• (2025). LAttE: A label-free and multimodal framework for context-aware person re-identification. NEUCOM, 653, 131172. (www)

  @article{ , abstract = {Person re-identification (Re-ID) refers to the task of identifying the same individual across multiple non-overlapping camera views, and it is considered a critical task in surveillance, security, and smart city applications. We introduce label-free attributes and pose embedding (LAttE), a novel Re-ID framework designed for intelligent surveillance systems that eliminates the need for manual attribute annotation. LAttE constructs a rich attribute bank by using GPT-4o to synthesize diverse human-centric descriptors, which are then embedded using a contrastive language–image pretraining encoder. These textual attributes are fused with visual and pose features through a cross-modal attention mechanism, resulting in a comprehensive representation of pedestrian appearance and structure. To further enhance robustness, we incorporate a feature alignment strategy based on the maximum mean discrepancy, improving consistency across varying viewpoints and sensor conditions. Experiments on benchmark datasets show that LAttE achieves state-of-the-art performance in both mean average precision and Rank-1 accuracy. These findings highlight its potential as a scalable and label-free solution for Re-ID tasks in intelligent vehicle applications, including pedestrian detection, in-cabin monitoring, and cooperative driving systems.}, archive = {J_NEUCOM}, author = {Dasom Ahn and Sangwon Kim and Kwang-Ju Kim and Byoung Chul Ko}, doi = {10.1016/j.neucom.2025.131172}, journal = {Neurocomputing}, month = {11}, pages = {131172}, shortjournal = {Neurocomputing}, title = {LAttE: A label-free and multimodal framework for context-aware person re-identification}, volume = {653}, year = {2025}, }
• (2025). Weak-edge sample extension for enhancing unsupervised feature learning. NEUCOM, 653, 131171. (www)

  @article{ , abstract = {Unsupervised person re-identification (Re-ID) methods aim to learn discriminative features from unlabeled datasets for person retrieval. Most existing methods rely on clustering to generate pseudo-labels. However,due to inaccurate feature representations and inherent limitations of clustering algorithms,clustering may sometimes merge different identities into the same group. Recent methods mitigate the issue by improving feature representations or refining pseudo-labels,but they overlook critical samples located on the sparse edges of clusters,which we refer to as weak-edge samples. In this paper,we propose a Weak-edge Sample Extension (WSE) framework to learn a more accurate decision boundary by performing sample extension specifically on weak-edge samples. To achieve this,we introduce an edge strength scoring mechanism based on the neighborhood relations of individual samples within their clusters. The neighborhood structure of each sample is represented by its k -nearest neighbor clusters,and a global neighborhood frequency distribution is constructed by aggregating these local neighborhoods across the entire cluster. Based on this distribution,we compute an edge strength score to automatically detect weak-edge samples and apply sample extension,providing structural support that enhances discriminative feature learning and improves model performance. Extensive experimental results on Market-1501 and MSMT17 demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Yuehai Chen and Shanying Chen and Jing Yang and Yifei Xu and Badong Chen and Shaoyi Du and Yuewen Liu}, doi = {10.1016/j.neucom.2025.131171}, journal = {Neurocomputing}, month = {11}, pages = {131171}, shortjournal = {Neurocomputing}, title = {Weak-edge sample extension for enhancing unsupervised feature learning}, volume = {653}, year = {2025}, }
• (2025). LightSpikformer: A compression framework for spiking transformer via low-rank decomposition of tensor networks. NEUCOM, 653, 131169. (www)

  @article{ , abstract = {Bioinspired spiking neural networks (SNNs) transmit information through sparse and discrete spikes, making them suitable for low-energy environments. However, since spiking neurons receive zero inputs at most time steps, using full-rank matrices for computation leads to wasted storage resources. Furthermore, SNNs expand spatial tasks to spatio-temporal domains through temporal coding, which increases computational complexity, model parameters, and memory requirements, especially in large-scale spiking transformer architectures. To address these issues, this paper proposes a compression framework via Low-rank decomposition of tensor networks. Specifically, the paper discusses the performance of two canonical tensor network frameworks Matrix Product State (MPS) and Matrix Product Operator (MPO), and constructs low-rank layers (LRLayer) based on their low-rank representations. Then, LRlayers are integrated into the Spiking Self-Attention (SSA) and Multilayer Perceptron (MLP) modules to construct LightSpikformer. Additionally, to mitigate noise spikes caused by the spatio-temporal invariance of SNNs, LightSpikformer adopts a channel-independent training strategy. Experimental results demonstrate the compression ratios and accuracy losses of MPS and MPO. The proposed framework achieves a parameter compression ratio of 84.49 % on the large-scale ImageNet dataset and reduces memory consumption to 42.29 % of the original model across various spatio-temporal spiking transformer architectures. Meanwhile, it maintains the accuracy advantage of the spiking transformer, achieving 94.72 % on the static dataset CIFAR-10 and 97.8 % on the neuromorphic dataset DVS128 Gesture.}, archive = {J_NEUCOM}, author = {Siyu Chen and Zhenyu Chen and Jianli Zhao and Liang Li and Tianheng Zhang and Xingzhao Feng}, doi = {10.1016/j.neucom.2025.131169}, journal = {Neurocomputing}, month = {11}, pages = {131169}, shortjournal = {Neurocomputing}, title = {LightSpikformer: A compression framework for spiking transformer via low-rank decomposition of tensor networks}, volume = {653}, year = {2025}, }
• (2025). Drone-guard: A self-supervised deep learning framework for real-time spatiotemporal anomaly detection in UAV surveillance systems. NEUCOM, 653, 131168. (www)

  @article{ , abstract = {Anomaly detection is a cornerstone of intelligent video surveillance, facilitating the early identification of irregular or potentially dangerous events. While deep learning has significantly advanced this domain, current approaches often face a fundamental trade-off between detection accuracy and computational efficiency. High-accuracy models are typically computationally intensive and unsuitable for real-time deployment on resource-constrained platforms such as Unmanned Aerial Vehicles (UAVs). Conversely, lightweight alternatives often lack the representational capacity to model complex spatiotemporal patterns in dynamic environments. To address this challenge, we present Drone-Guard, a self-supervised deep learning framework designed for real-time spatiotemporal anomaly detection in UAV surveillance systems. Drone-Guard introduces three core contributions: (1) a lightweight encoder-decoder architecture with multi-stage feature extraction to jointly capture fine-grained and high-level spatial structures; (2) a novel Multi-Scale Grouped Query Attention (MS-GQA) mechanism for efficient fusion of hierarchical spatial and temporal features, enabling context-aware anomaly modeling; and (3) a Residual Vector Quantization (RVQ) module that enhances latent representation compactness and reconstruction fidelity, crucial for discriminative anomaly detection. To overcome the lack of labeled anomalies—a central limitation in self-supervised learning—we further propose a latent-space pseudo-anomaly synthesizer. This component perturbs the learned representations of normal samples to generate synthetic anomalies, thereby facilitating effective decision boundary learning without requiring manual annotations. Extensive experiments and ablation studies on multiple benchmark datasets demonstrate that Drone-Guard outperforms state-of-the-art methods in both accuracy and efficiency. Its low computational footprint and robust anomaly localization capabilities make it well-suited for real-time deployment in edge-aware, IoT-enabled UAV surveillance applications. The source code and pre-trained models of the proposed framework are publicly available at: https://github.com/slitiWassim/Drone-Guard .}, archive = {J_NEUCOM}, author = {Wassim Sliti and Olfa Besbes}, doi = {10.1016/j.neucom.2025.131168}, journal = {Neurocomputing}, month = {11}, pages = {131168}, shortjournal = {Neurocomputing}, title = {Drone-guard: A self-supervised deep learning framework for real-time spatiotemporal anomaly detection in UAV surveillance systems}, volume = {653}, year = {2025}, }
• (2025). Small sample hyperspectral image classification based on spiking re-parameterization and multi-discriminator adversarial. NEUCOM, 653, 131165. (www)

  @article{ , abstract = {Spiking neural networks (SNNs) offer low energy consumption and have shown promising performance in hyperspectral image (HSI) classification, making them suitable for deployment on airborne and satellite platforms. However, their performance significantly degrades when training samples are scarce, particularly for minority classes. To address this limitation, this paper proposes a method called SR-MDA, which combines spiking re-parameterization with multi-discriminator adversarial (SR-MDA) for small-sample HSI classification. Firstly, the spiking re-parameterization network (SRepNet) was constructed. This network adopts a multi-branch structure during training and can be converted into a single-branch structure through spiking re-parameterization during deployment, thereby effectively improving the classification performance of the network. Secondly, a multi-discriminator loss feedback module was proposed, which guides the learning of the generator through multiple discriminators and adopts a partial loss feedback strategy, effectively alleviating the mode collapse problem of generative adversarial network (GAN). Finally, a hybrid network architecture combining SNN and GAN was constructed by using SRepNet as the strong classifier and the discriminator in GAN as the weak classifier. The model not only leverages the energy efficiency of SNNs, but also significantly enhances classification accuracy under limited training data. Experiments on widely used PU, IP, SA, and WHLK datasets have shown that the proposed method exhibits excellent performance in small sample HSI classification tasks, providing strong technical support for promoting the deployment of HSI classification algorithms on unmanned autonomous equipment such as airborne and satellite platforms.}, archive = {J_NEUCOM}, author = {Yahui Li and Yang Liu and Huaxu He and Yi Xie}, doi = {10.1016/j.neucom.2025.131165}, journal = {Neurocomputing}, month = {11}, pages = {131165}, shortjournal = {Neurocomputing}, title = {Small sample hyperspectral image classification based on spiking re-parameterization and multi-discriminator adversarial}, volume = {653}, year = {2025}, }
• (2025). MM-instruct: Generated visual instructions for large multimodal model alignment. NEUCOM, 653, 131164. (www)

  @article{ , abstract = {This paper presents MM-Instruct, an automated pipeline for generating diverse and high-quality visual instruction data to better align large multimodal models (LMMs) with real-world use cases. While previous works have focused on question-answering data, their generated instruction datasets pose challenges for broader application scenarios. Additionally, manually collecting diverse instruction data at scale from users is prohibitively costly. To mitigate these issues, MM-Instruct leverages ChatGPT to automatically generate diverse instructions from a limited set of seed instructions through augmentation and summarization. It then uses an open-sourced large language model (LLM) to construct instruction-following answers and builds a large-scale visual instruction dataset. Evaluating the LLaVA-Instruct models trained with the generated data shows significant improvements in instruction-following capabilities compared to LLaVA-1.5 models. MM-Instruct releases its synthetic dataset containing diverse instructions and high-quality instruction-answer pairs to support training LMMs for real-world applications.}, archive = {J_NEUCOM}, author = {Xin Huang and Jihao Liu and Jinliang Zheng and Boxiao Liu and Jia Wang and Yu Liu and Hongsheng Li and Osamu Yoshie}, doi = {10.1016/j.neucom.2025.131164}, journal = {Neurocomputing}, month = {11}, pages = {131164}, shortjournal = {Neurocomputing}, title = {MM-instruct: Generated visual instructions for large multimodal model alignment}, volume = {653}, year = {2025}, }
• (2025). Variational inference-aided neural architecture search for secure deep learning implementation. NEUCOM, 653, 131163. (www)

  @article{ , abstract = {With the proliferation of AI in day-to-day life, serious concerns arise regarding the secure implementation of these algorithms in real-time image classification tasks. This work implements adversarially robust deep networks by augmenting the neural architecture search (NAS) technique that focuses on improving generalization (i.e., the ability of a model to predict well on adversarial inputs). The structure of the DL model and the learning objective have been augmented to encourage the model to learn through low-rank features and adapt to the shared domain of natural and adversarial examples. The proposed methodology is computationally efficient compared to the prominent defensive strategy, i.e., adversarial training (AT), because of having a negligible perturbation budget. Furthermore, the on-chip implementation of the well-generalized and adaptive(and hence robust) models has been done using high-level synthesis (HLS). The efficacy of the proposed methodology towards adversarial attacks is evaluated on multiple datasets, showing an average enhancement in clean accuracy of 12.43 % compared to AT on both Adam and SGD optimizers. Also, the experimental results indicate that the proposed method can enable different state-of-the-art DNN model implementations, with a significant 1.5 to 3-fold relative enhancement in throughput (frames per second), over the widely popular model MobileNetV2 for image classification tasks.}, archive = {J_NEUCOM}, author = {Abhishek Yadav and Vyom Kumar Gupta and Gaurav Singh Bhati and Binod Kumar}, doi = {10.1016/j.neucom.2025.131163}, journal = {Neurocomputing}, month = {11}, pages = {131163}, shortjournal = {Neurocomputing}, title = {Variational inference-aided neural architecture search for secure deep learning implementation}, volume = {653}, year = {2025}, }
• (2025). Dimension-driven feature complementation for visible-infrared person re-identification. NEUCOM, 653, 131162. (www)

  @article{ , abstract = {Visible-infrared person re-identification (VI-ReID) seeks to search visible and infrared images of a pedestrian captured by different cameras. Its primary challenge stems from the substantial modality discrepancy between visible and infrared images due to their distinct spectral properties. A crucial approach to reducing the cross-modality discrepancy is to capture comprehensive modality-shared features while filtering out modality-specific information. In this paper, we present a Dimension-Driven Feature Complementer (DDFC), which performs local perception and global regulation across different dimensions to capture comprehensive modality-shared information. Specifically, we design a Partitioned Local Perceptor (PLP) to enhance the perception of pedestrian details by capturing directional information granularly. Furthermore, we propose a Global Attention Regulator (GAR) which leverages attention mechanisms to adaptively enhance features across different dimensions, enabling the model to focus effectively on critical identity information. Finally, we integrate these two components to form the Dimension-Driven Feature Complementer (DDFC), which optimizes both local and global features across dimensions and leverages the complementarity between them to improve the comprehensiveness and diversity of feature representations, thereby remarkably reducing modality discrepancies. Extensive experiments on public SYSU-MM01, RegDB and LLCM datasets demonstrate the superiority of our DDFC over state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Ran Wang and Dechang Pi and Rong Yu and Guoqing Zhang}, doi = {10.1016/j.neucom.2025.131162}, journal = {Neurocomputing}, month = {11}, pages = {131162}, shortjournal = {Neurocomputing}, title = {Dimension-driven feature complementation for visible-infrared person re-identification}, volume = {653}, year = {2025}, }
• (2025). Diverse object placement with dual interaction. NEUCOM, 653, 131161. (www)

  @article{ , abstract = {Object placement aims to find the appropriate positions and sizes of the foreground in the background image, thereby achieving natural composition. Existing methods struggle to simultaneously achieve reasonable and diverse placement, which hinders large-scale automatic image composition. To address this problem, this paper proposes a novel placement method named RDPlace. The core of RDPlace is the placement search network, which efficiently utilizes the correlation information between the foreground and the background and the background semantics through its dual interaction module, thereby enhancing the ability to represent information related to placement. In addition, we design a novel sphere sampling method and adopt the loss constraint on its sampling and output space, enabling the model to generate placement results with different positions and sizes. Experimental results on the well-known OPA and OPAZ datasets show that the proposed RDPlace performs well in visual rationality and generation diversity, and the generated placements are more natural and coordinated.}, archive = {J_NEUCOM}, author = {Xianhe Cheng and Peng Zhai and Dingkang Yang and Xiangrui Meng and Yang Xia and Lihua Zhang}, doi = {10.1016/j.neucom.2025.131161}, journal = {Neurocomputing}, month = {11}, pages = {131161}, shortjournal = {Neurocomputing}, title = {Diverse object placement with dual interaction}, volume = {653}, year = {2025}, }
• (2025). Joint image upsampling with affinity learning. NEUCOM, 653, 131159. (www)

  @article{ , abstract = {Joint image upsampling is of significance to various fields of image processing and computer vision. Traditional joint upsampling methods mostly attempt to model the relations between the low-resolution guidance and target images, then the relations are upsampled and applied to the high-resolution guidance image. However, we argue that the upsampling of the relations could also be subject to loss of information. In this paper, we propose to discover the task-dependent affinities between the low-resolution and high-resolution guidance images, then map the affinities to the low-resolution target image. We show that the proposed scheme is effective for joint image upsampling. Based on this principle, we propose a deep neural network that learns the affinities based on the convolutional spatial propagation network and applies them to the low-resolution target images for upsampling. Compared to the state-of-the-art joint upsampling networks, our model is more interpretable. We have experimented with the proposed upsampling method on a variety of low-level vision and image processing tasks, including depth map upsampling, image smoothing upsampling, style transfer upsampling, and image dehazing upsampling. The results indicate the superiority of the proposed joint upsampling method over the state-of-the-art methods. Furthermore, the proposed method is highly efficient, rendering real time upsampling into 720P color images on a modern GPU. Therefore, it is practical for real usage. Our code and trained models are available at https://github.com/LiQingCode/JUAL .}, archive = {J_NEUCOM}, author = {Yang Yang and Shuai Luo and Shuailong Qiu and Lanling Zeng and Xinyu Wang and Xinsheng Wang}, doi = {10.1016/j.neucom.2025.131159}, journal = {Neurocomputing}, month = {11}, pages = {131159}, shortjournal = {Neurocomputing}, title = {Joint image upsampling with affinity learning}, volume = {653}, year = {2025}, }
• (2025). Fully-distributed dynamic event-triggered optimized backstepping bipartite consensus control based on neural network observer and reinforcement learning. NEUCOM, 653, 131157. (www)

  @article{ , abstract = {A fully-distributed dynamic event-triggered optimized bipartite consensus control is investigated for nonlinear strict-feedback multiagent systems (MASs), utilizing a neural network (NN) observer and reinforcement learning (RL) technique. First, an NN-based output-feedback state observer is proposed to estimate the immeasurable system states, eliminating the need to design a constant matrix for ensuring a Hurwitz characteristic polynomial. Second, an optimized backstepping (OB) control based on the NN observer and the critic-actor structure of RL is proposed to achieve bipartite consensus, in which gradient descent is derived from a simple positive function that greatly reduces the complexity of the algorithm. Moreover, a dynamic event-triggered control is designed to reduce the execution frequency of practical controllers. Finally, simulations are ultimately carried out to confirm the effectiveness of the proposed NN observer and dynamic event-triggered bipartite consensus control.}, archive = {J_NEUCOM}, author = {Ao Teng and Guoguang Wen and Zhaoxia Peng and Yidi Wang and Bofan Wu and Tingwen Huang}, doi = {10.1016/j.neucom.2025.131157}, journal = {Neurocomputing}, month = {11}, pages = {131157}, shortjournal = {Neurocomputing}, title = {Fully-distributed dynamic event-triggered optimized backstepping bipartite consensus control based on neural network observer and reinforcement learning}, volume = {653}, year = {2025}, }
• (2025). RD-DIP: Rician denoising deep image prior. NEUCOM, 653, 131156. (www)

  @article{ , abstract = {Rician denoising of magnetic resonance images (MRI) is a fundamental problem in medical image processing. Although variational methods can address Rician noise, the mathematical complexity of the underlying models makes them challenging to implement. Neural networks offer a compelling alternative but require extensive labelled datasets and are prone to introducing artifacts. This study proposes advanced tailored models for MRI Rician denoising within the unsupervised Deep Image Prior (DIP) framework, eliminating the need for additional data and reducing training-related artifacts. To our knowledge, this is the first application of the Rician maximum likelihood in the DIP framework. Our experiments, conducted on the Brainweb dataset and a set of real MRI scans, show the superiority of the proposed models over traditional and recent state-of-the-art unsupervised approaches. Additionally, we provide a detailed pipeline to ensure the reproducibility of our experiments. The code is available at https://github.com/heqro/rd-dip .}, archive = {J_NEUCOM}, author = {H.R. Iglesias-Goldaracena and I. Ramírez and E. Schiavi}, doi = {10.1016/j.neucom.2025.131156}, journal = {Neurocomputing}, month = {11}, pages = {131156}, shortjournal = {Neurocomputing}, title = {RD-DIP: Rician denoising deep image prior}, volume = {653}, year = {2025}, }
• (2025). PMN: A prototype network based metric framework for solving aspect-based sentiment analysis tasks. NEUCOM, 653, 131155. (www)

  @article{ , abstract = {Aspect-based Sentiment Analysis (ABSA) aims to identify aspect terms and analyze sentiment by associating them with corresponding sentiment polarity and opinion words. Existing approaches primarily involve four key components: aspect term extraction, aspect category classification, opinion term extraction, and sentiment polarity classification, each giving rise to multiple fine-grained tasks. These tasks complicate the construction and annotation of foundational datasets. However, due to the heavy reliance on manual annotations and the limited applicability of domain-specific datasets, this study introduces a few-shot learning approach. It constructs a prototype network based on these four tasks. Despite the interdependencies among various ABSA tasks, exploring their potential relationships remains limited. This research centers on a quadruple (QUAD) cognitive task framework, evaluating the complexity of ABSA tasks from this perspective. We propose a Prototype Metric Network that enhances the mapping capabilities of the prototype network through three distinct distance metrics, effectively addressing challenges arising from data scarcity and the constrained scope of multiple ABSA tasks. Experimental results demonstrate that our model outperforms state-of-the-art algorithms, providing a more complete and accurate extraction of significant sentiment targets.}, archive = {J_NEUCOM}, author = {Chenhao Li and Wenti Huang and Yunfei Chen and Tingxuan Chen and Zhan Yang and Jun Long}, doi = {10.1016/j.neucom.2025.131155}, journal = {Neurocomputing}, month = {11}, pages = {131155}, shortjournal = {Neurocomputing}, title = {PMN: A prototype network based metric framework for solving aspect-based sentiment analysis tasks}, volume = {653}, year = {2025}, }
• (2025). Multi-resolution learning with DeepONets and long short-term memory neural networks. NEUCOM, 653, 131154. (www)

  @article{ , abstract = {Deep operator networks (DeepONets, DONs) offer a distinct advantage over traditional neural networks in their ability to learn from data collected at heterogeneous resolutions. This property, known as discretization invariance , can be particularly valuable when modeling fine-scale temporal dynamics with limited high-resolution data, provided that lower-resolution data are available. Nevertheless, DeepONets alone often struggle to capture and maintain dependencies over long sequences compared to other state-of-the-art algorithms. We propose a novel framework that leverages multi-resolution data in training and provides precise models when dealing with limited high-resolution data. We achieve this through extending the DeepONet architecture with a long short-term memory network (LSTM), and training it in a three-step procedure that utilizes data of different levels of granularity. Combining these two architectures, we equip the network with explicit mechanisms to leverage multi-resolution data, as well as capture temporal dependencies in long sequences. We test our method on long-time-evolution modeling of multiple non-linear systems and show that the proposed multi-resolution DON-LSTM achieves significantly lower generalization error and requires fewer high-resolution samples compared to its vanilla counterparts.}, archive = {J_NEUCOM}, author = {Katarzyna Michałowska and Somdatta Goswami and George Em Karniadakis and Signe Riemer-Sørensen}, doi = {10.1016/j.neucom.2025.131154}, journal = {Neurocomputing}, month = {11}, pages = {131154}, shortjournal = {Neurocomputing}, title = {Multi-resolution learning with DeepONets and long short-term memory neural networks}, volume = {653}, year = {2025}, }
• (2025). “Less but more efficient ”: Using less data to achieve better logical reasoning. NEUCOM, 653, 131153. (www)

  @article{ , abstract = {Logical reasoning is a fundamental aspect of natural language processing tasks and has been a focal point for researchers. In a previous work, MERIt introduced meta-path guided contrastive learning to enhance models’ logical reasoning capabilities. However, MERIt relies on extensive Wikipedia data, leading to challenges such as variable data quality and low data utilization rates. This paper presents a solution to these challenges through a novel data filtering approach, termed TAP ( t -SNE a lgorithm and p erplexity ranking). The t-SNE algorithm is employed for data classification, followed by perplexity ranking within each class. Subsequently, data with lower perplexity from each class are selected to construct a refined dataset. Experimental results demonstrate that our proposed method achieves comparable or superior performance to MERIt while utilizing less than 10 % of the data volume.}, archive = {J_NEUCOM}, author = {Ruixue Duan and Xin Liu and Zhigang Ding and Yangsen Zhang}, doi = {10.1016/j.neucom.2025.131153}, journal = {Neurocomputing}, month = {11}, pages = {131153}, shortjournal = {Neurocomputing}, title = {“Less but more efficient ”: Using less data to achieve better logical reasoning}, volume = {653}, year = {2025}, }
• (2025). Density peaks clustering algorithm via fusing natural neighbor and fuzzy information. NEUCOM, 653, 131152. (www)

  @article{ , abstract = {Although density peaks clustering (DPC) is an efficient clustering algorithm, it performs poorly on manifold data, density-imbalanced data, and cluster boundary overlapping data. Moreover, DPC is highly sensitive to the cutoff distance. To address these issues, a novel density peak clustering algorithm via fusing natural neighbor and fuzzy information (FNaN-DPC) is proposed. Firstly, the adaptive local density metric based on natural neighbor is introduced to eliminate the high density differences among different clusters to accurately recognize cluster centers. Secondly, a label propagation method based on natural neighbor is proposed to adapt to manifold data. Thirdly, the termination condition for label propagation is set to distinguish the core and boundary regions of clusters. Finally, the designed fuzzy membership function is utilized to determine the attribution of data points in the boundary region, which enables the FNaN-DPC algorithm to effectively process data with overlapping cluster boundaries. The experimental results on both synthetic and real-world datasets show the effectiveness and superiority of FNaN-DPC.}, archive = {J_NEUCOM}, author = {Xufei Guo and Qiong Liu and Chaoqun Huang and Mingjie Cai}, doi = {10.1016/j.neucom.2025.131152}, journal = {Neurocomputing}, month = {11}, pages = {131152}, shortjournal = {Neurocomputing}, title = {Density peaks clustering algorithm via fusing natural neighbor and fuzzy information}, volume = {653}, year = {2025}, }
• (2025). Occluded facial expression recognition algorithm integrating logical reasoning on small-sample datasets. NEUCOM, 653, 131151. (www)

  @article{ , abstract = {To solve the problem of low prediction accuracy of deep learning based expression recognition algorithms when facing small-sample occluded expression datasets, this paper proposes a novel approach that integrates logical reasoning with a deep learning model combining local action and global feature extraction. The algorithm consists of three steps: first, using a local action recognition module (LARM) to obtain key local actions for the eyes and mouth & nose respectively; secondly, facial image inpainting module is used to repair the occluded parts to obtain global features and semantic information of the occluded parts. At the same time, spatial attention mechanism is introduced to obtain the spatial attention weight features of the original expression image, and the two are fused for feature fusion; finally, using logic reasoning, the final prediction result is obtained from local actions and global features that integrate spatial attention weight features. In summary, the proposed model introduces logical reasoning as supplementary guidance, mitigating the problem of insufficient features caused by small datasets and improving accuracy when dealing with limited samples. Furthermore, by incorporating global feature extraction and a spatial attention mechanism, the model reduces the negative impact of inaccurate reconstruction results from occluded regions, while preserving global and semantic information. Experimental results on small-sample occluded facial expression classification datasets based on the KDEF, RAFD and RAF-DB datasets demonstrate that the proposed algorithm achieves a significant accuracy improvement compared to existing methods.}, archive = {J_NEUCOM}, author = {Ning Zhou and Bin Liu and Jiawei Cao}, doi = {10.1016/j.neucom.2025.131151}, journal = {Neurocomputing}, month = {11}, pages = {131151}, shortjournal = {Neurocomputing}, title = {Occluded facial expression recognition algorithm integrating logical reasoning on small-sample datasets}, volume = {653}, year = {2025}, }
• (2025). Development and application of HDA-ABi-GRU-based on the interpersonal problems scale of personality disorders in acromegaly patients. NEUCOM, 653, 131138. (www)

  @article{ , abstract = {Acromegaly, a chronic condition caused by excessive Growth Hormone (GH) production, leads to significant physical changes. These physical manifestations can severely affect patients' social interactions, potentially leading to interpersonal difficulties and the development of personality disorders. The Inventory of Interpersonal Problems (IIP) is a widely used tool for assessing interpersonal difficulties, which are often linked to personality disorders. However, applying the IIP to acromegaly patients presents unique challenges due to the complex interplay between physical appearance and psychological distress. To address these limitations, this study developed a new model that combines the IIP with additional psychological evaluations specifically tailored to acromegaly patients. The novelty of the proposed model lies in its robust and clinically informed approach to detecting interpersonal problems in acromegaly patients. By combining psychological evaluation with clinical parameters and deep learning, it provides a more holistic and accurate understanding of patient well-being. It confirms its superior predictive capabilities, offering a practical solution for both clinicians and mental health professionals. This work not only advances the field of medicine but also addresses an important unmet need in the comprehensive management of acromegaly. Here, a dimension reduction procedure was performed using Multiscale Weighted Principal Component Analysis (MWPCA) and T-Distributed Stochastic Neighbor Embedding (t-SNE). The relationship between IIP values and various parameters, such as age, gender, and GH values, is analyzed with the help of MWPCA and t-SNE. The results from MWPCA and t-SNE are concatenated and fed into the Hybrid Dilation and Attention-based Adaptive Bi-Directional Gated Recurrent Unit (HDA-ABiGRU) to detect the IIP in acromegaly patients based on the IIP values. Here, the hyperparameters of the Bi-Directional Gated Recurrent Unit (Bi-GRU) are tuned using the Refined and Efficient Fossa Optimization Algorithm (RE-FOA). The suggested HDA-ABiGRU effectively predicted the IIP in Acromegaly patients to enhance the quality of life. The proposed HDA-ABiGRU model effectively predicts the IIP in acromegaly patients, thereby enhancing their quality of life. Finally, extensive experimental validation is performed to confirm the effectiveness of the developed model in IIP detection. The findings indicated that the developed model achieved high accuracy in detecting acromegaly from facial images. This is particularly valuable as facial changes in acromegaly can be gradual and difficult to notice early on.}, archive = {J_NEUCOM}, author = {Shashi Kiran Tagilapalli and Lei Shi and Zhentian Wu and Canhui Tian and Yanmei Tie and Dongsheng He and Haijun Wang and Wenli Chen}, doi = {10.1016/j.neucom.2025.131138}, journal = {Neurocomputing}, month = {11}, pages = {131138}, shortjournal = {Neurocomputing}, title = {Development and application of HDA-ABi-GRU-based on the interpersonal problems scale of personality disorders in acromegaly patients}, volume = {653}, year = {2025}, }
• (2025). Spatio-temporal memory-driven CT organ segmentation with hybrid CNN-mamba encoding and frequency-domain decoding. NEUCOM, 653, 131137. (www)

  @article{ , abstract = {Accurate segmentation of organs from CT scan slice sequences is essential for effective diagnosis and treatment planning. However, many recent deep learning-based methods overlook the rich temporal context across slices, resulting in poor performance when organs exhibit large deformations or positional shifts. To address this, we propose HCMFDS-Net, a novel architecture that integrates a spatio temporal memory module to fully exploit inter slice relationships. By storing spatial and temporal features of the target organ from preceding slices and retrieving them during decoding, our model dynamically guides segmentation of the current slice, thereby enhancing robustness to deformation and misalignment. CT images are often corrupted by high noise and suffer from low organ-background contrast. To overcome these challenges, we design a Hybrid CNN-Mamba Encoder that effectively extracts rich semantic features from noisy inputs, and a Frequency Domain Decoder that reinforces boundary details in the decoded representations. Code: https://github.com/Z0721Z/HCMFDS-Net .}, archive = {J_NEUCOM}, author = {Yongpeng Zhao and Guangyuan Zhang and Kefeng Li and Zhenfang Zhu and Jiayi Yu and Xiaotong Li and Yongshuo Zhang and Zhiming Fan}, doi = {10.1016/j.neucom.2025.131137}, journal = {Neurocomputing}, month = {11}, pages = {131137}, shortjournal = {Neurocomputing}, title = {Spatio-temporal memory-driven CT organ segmentation with hybrid CNN-mamba encoding and frequency-domain decoding}, volume = {653}, year = {2025}, }
• (2025). Privacy-preserving online medical image exchange via hyperchaotic memristive neural networks and DNA encoding. NEUCOM, 653, 131132. (www)

  @article{ , abstract = {In this paper, we introduce an innovative hyperchaotic memristive Hopfield neural network (MHNN) architecture that synergistically embeds dual memristive components within conventional Hopfield neural circuitry. The proposed MHNN demonstrates a unique capability of producing intricate multidimensional hidden hyperchaotic attractors with multi-wing configurations. Through comprehensive dynamical investigations employing bifurcation diagrams, Lyapunov exponent spectra, phase portrait reconstruction, and spectral entropy analysis, we systematically elucidate the formation mechanisms and nonlinear characteristics governing these hidden multiwing attractor structures. A digital verification platform is engineered to experimentally validate the numerical predictions, achieving hardware–software consistency in hyperchaotic behavior reproduction. Building upon these theoretical advancements, we design a cryptographic framework integrating the hidden hyperchaotic multi-wing MHNN (HHM-MHNN) dynamics with DNA encoding principles. This hybrid encryption architecture can be successfully deployed in telemedicine infrastructures to safeguard diagnostic imaging data, demonstrating clinical-grade security through multilayered protection mechanisms. Comprehensive security evaluations demonstrate the algorithm’s capacity to withstand statistical cryptanalysis and differential attacks, leveraging the HHM-MHNN’s nonlinear unpredictability to ensure ciphertext indistinguishability. Experiments demonstrate that this encryption scheme can achieve an encryption entropy of 7.9997 and an key space of 2 372 . It has been verified that the scheme possesses sufficient performance and robustness to adequately address the challenges in remote online healthcare, including real-time encryption requirements, lightweight computation demands, and large-scale data processing. The solution maintains medical image integrity while preventing unauthorized data reconstruction, establishing a reliable paradigm for secure health data exchange in distributed healthcare ecosystems.}, archive = {J_NEUCOM}, author = {Xiaoheng Deng and Sirui Ding and Hairong Lin and Lin Jiang and Hong Sun and Jie Jin}, doi = {10.1016/j.neucom.2025.131132}, journal = {Neurocomputing}, month = {11}, pages = {131132}, shortjournal = {Neurocomputing}, title = {Privacy-preserving online medical image exchange via hyperchaotic memristive neural networks and DNA encoding}, volume = {653}, year = {2025}, }
• (2025). MixNet: Efficient global modeling for ultra-high-definition image restoration. NEUCOM, 653, 131130. (www)

  @article{ , abstract = {Recent advancements in image restoration methods employing global modeling have shown promising results. However, these approaches often incur substantial memory requirements, particularly when processing ultra-high-definition (UHD) images. In this paper, we propose a novel image restoration method called MixNet, which introduces an alternative approach to global modeling approaches and is more effective for UHD image restoration. To capture the long-range dependency of features without introducing excessive computational complexity, we present the Global Feature Modulation Layer (GFML). GFML associates features from different views by permuting the feature maps, enabling efficient modeling of long-range dependencies. In addition, we also design the Local Feature Modulation Layer (LFML) and Feed-forward Layer (FFL) to capture local features and transform features into a compact representation. This way, our MixNet achieves effective restoration with low inference time overhead and computational complexity. We conduct extensive experiments on four UHD image restoration tasks, including low-light image enhancement, underwater image enhancement, image deblurring and image demoiréing, and the comprehensive results demonstrate that our proposed method surpasses the performance of current state-of-the-art methods. The code will be available at https://github.com/5chen/MixNet .}, archive = {J_NEUCOM}, author = {Chen Wu and Shuning Sun and Yu Zhang and Zhuoran Zheng}, doi = {10.1016/j.neucom.2025.131130}, journal = {Neurocomputing}, month = {11}, pages = {131130}, shortjournal = {Neurocomputing}, title = {MixNet: Efficient global modeling for ultra-high-definition image restoration}, volume = {653}, year = {2025}, }
• (2025). Multi-objective optimization collaborating with deep reinforcement learning: Adversarial audio attacks. NEUCOM, 653, 131124. (www)

  @article{ , abstract = {With the increasing prevalence of human-computer speech interaction, the security of Automatic Speech Recognition (ASR) models in Speech-to-Text (STT) systems has become a critical concern. Generating targeted and concealed adversarial speech examples poses significant challenges due to noise disturbances, inefficient querying processes, and limited consideration of speech timing characteristics. This paper introduces a novel black-box generative audio attack method, Deep Q-Network-Driven Multi-objective Optimization in Adversarial Audio Attack (DQMOA). For the first time, deep reinforcement learning is integrated into adversarial audio attacks. By leveraging deep Q-network capabilities to map solution spaces to attack behaviors, along with a reward and punishment mechanism and experience pool, DQMOA delivers optimal decision-making for multi-objective genetic algorithm. This approach substantially reduces query counts, achieves high attack success rates, and ensures low word error rates with superior speech naturalness. Extensive evaluations are conducted on the Mozilla Common Voice and LibriSpeech datasets across three commercial ASR platforms, where DQMOA consistently outperforms five state-of-the-art baselines. Further experiments on Whisper large-v2/v3 and real-world over-the-air scenarios demonstrate its strong generalization ability and practical applicability. DQMOA achieves high attack success rates with over 10 % fewer queries, while preserving low word error rates and high perceptual audio quality. These findings highlight the potential of reinforcement learning-guided evolutionary optimization for robust and stealthy adversarial audio attack generation.}, archive = {J_NEUCOM}, author = {Pengchuan Wang and Wen Cui and Deqiang Li and Qianmu Li}, doi = {10.1016/j.neucom.2025.131124}, journal = {Neurocomputing}, month = {11}, pages = {131124}, shortjournal = {Neurocomputing}, title = {Multi-objective optimization collaborating with deep reinforcement learning: Adversarial audio attacks}, volume = {653}, year = {2025}, }
• (2025). Temporal constrained joint nonnegative matrix factorization for imaging genomic study of early mild cognitive impairment. NEUCOM, 653, 131121. (www)

  @article{ , abstract = {The early mild cognitive impairment (EMCI) has hindered the early diagnosis and intervention of Alzheimer’s disease. Imaging genetics models based on joint nonnegative matrix factorization have offered an opportunity to effectively explore the relationship between quantitative brain features and epigenetic sites in patients with EMCI. However, previous studies fail to fully utilize the temporal dimension information, resulting in image genetic associations being mined solely at a single time point. In this study, the temporal prior information was extracted based on the similarity network fusion method using quantitative brain feature data at four time points. Then, the temporal constraints were then regularized and integrated into the joint non-negative matrix decomposition model, resulting in the proposal of a temporal constrained joint non-negative matrix (TCJNMF) algorithm. Besides, the proposed algorithm utilizes a semi-supervised learning model to mitigate result randomness. The results revealed that the enhancement significantly improves the capability of the JNMF algorithm in uncovering the association patterns between neuroimages and epigenetic loci in patients with EMCI. A set of brain region-epigenetic locus association pairs was also obtained, demonstrating statistical significance, and offering valuable insights regarding biological interpretation.}, archive = {J_NEUCOM}, author = {Jin Deng and Kaihan Huang and Jinfeng Wang and Wenjian Zhong and Yufang Xu and Wei Kong}, doi = {10.1016/j.neucom.2025.131121}, journal = {Neurocomputing}, month = {11}, pages = {131121}, shortjournal = {Neurocomputing}, title = {Temporal constrained joint nonnegative matrix factorization for imaging genomic study of early mild cognitive impairment}, volume = {653}, year = {2025}, }
• (2025). TD-IVDM: A multi-scale concept drift detection method for time series forecasting tasks. NEUCOM, 653, 131120. (www)

  @article{ , abstract = {Concept drift refers to unpredictable changes in the statistical properties of target variables over time, which degrade model performance in time series forecasting. Therefore, it is crucial to detect concept drift accurately. Existing concept drift detection methods based on data distribution focus on changes in the overall data distribution of all variables within specific windows, which are inadequate for tackling local temporal and joint distribution drift of smaller time frames and subsets of variables. Therefore, a multi-scale concept drift detection method called TD-IVDM (Time Dependency-Inter Variable Dependency Method) is proposed, which captures both temporal and inter-variable dependencies. This method consists of a data packaging module, a feature extraction module and a threshold comparison module. In the feature extraction module, the representation learning network TS2Vec is improved to extract time dependencies and a multi-dimensional Kernel Density Estimation (KDE) method is used to capture inter-variable dependencies. The changes in the distribution are quantified by calculating the difference in weighted features. In the threshold comparison module, a task-adaptive triple-threshold comparison module is proposed to compare the calculated differential features with the thresholds and output the severity level of concept drift. The results of experiments on artificial and real-world datasets validate the effectiveness of our method in detecting concept drift in time series forecasting tasks.}, archive = {J_NEUCOM}, author = {Xiao-li Wang and Shang-lin Zhu and Li-yang Qin and Jie Han and Yu-long Wang and Feng Yan}, doi = {10.1016/j.neucom.2025.131120}, journal = {Neurocomputing}, month = {11}, pages = {131120}, shortjournal = {Neurocomputing}, title = {TD-IVDM: A multi-scale concept drift detection method for time series forecasting tasks}, volume = {653}, year = {2025}, }
• (2025). A distributed inference algorithm for dirichlet process mixture models with exponential family components. NEUCOM, 653, 131119. (www)

  @article{ , abstract = {This article extends our earlier work published in [24] . Dirichlet Process Mixture Models (DPMMs) are widely used for clustering due to their ability to infer the number of clusters via a Bayesian non-parametric framework. However, MCMC inference with DPMMs becomes computationally intensive on large datasets. To address this, we propose a distributed inference algorithm, DisCGS, which approximates the collapsed Gibbs sampler using sufficient statistics and is specifically designed for horizontally distributed data across independent and potentially heterogeneous machines, making it well-suited for federated learning scenarios. Our contributions are threefold: first, we develop and evaluate DisCGS for the multivariate Gaussian DPMM, demonstrating significant computational gains—for example, reducing runtime from approximately 12 h to 3 min for 100 iterations on 100 K data points, a 200 × speedup—while maintaining comparable clustering quality as measured by Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and Accuracy. Second, we introduce a multinomial DPMM formulation within the DisCGS framework tailored for discrete data such as text documents, showcasing its applicability to categorical and count-based data common in natural language processing and genomics. Third, we prove that DisCGS generalizes to all exponential family distributions, thereby extending its flexibility and applicability beyond Gaussian and multinomial models. The source code for the continuous (Gaussian) model is available at https://github.com/redakhoufache/DisCGS , and the discrete (multinomial) version for text clustering is provided at https://github.com/redakhoufache/DisCGS-for-Discrete-data .}, archive = {J_NEUCOM}, author = {Reda Khoufache and Mustapha Lebbah and Hanene Azzag and Etienne Goffinet and Djamel Bouchaffra}, doi = {10.1016/j.neucom.2025.131119}, journal = {Neurocomputing}, month = {11}, pages = {131119}, shortjournal = {Neurocomputing}, title = {A distributed inference algorithm for dirichlet process mixture models with exponential family components}, volume = {653}, year = {2025}, }
• (2025). Decoding fake news fabrications and trends: A comprehensive survey. NEUCOM, 653, 131118. (www)

  @article{ , abstract = {Increased internet access has led to a surge in online content across blogs, websites, news portals, and social media, where people actively share personal ideas, opinions, ideologies while seeking information of their interest. However, relying on individual sources can lead to information overload and the spread of unverified data, often shaped by personal biases. This lack of fact-based reliability fuelled the generation and spread of fake news, undermining trust in digital information ecosystems. To tackle these challenges, Fake News Detection (FND) has become a crucial research area, drawing significant attention of experts to develop solutions to combat misinformation and restore trust in online information. This paper provides a comprehensive review of the changing patterns of fake news trends over time, tracing its shift from text to visual and eventually hybrid formats over the past decade. It reviews the generation and propagation of fake news, explores detection methods and highlights the challenges for efficient detection, including how human and algorithmic bias unknowingly contributes to its spread. The paper discusses key research questions and their implications, emphasizing why multimodal sentiment analysis outperforms other methods for detecting complex, malicious intent. It also provides an overview of popular datasets and resources, along with a bibliometric analysis highlighting key authors and leading institutes in the research area. Finally, it discusses the future direction of fake news detection, underscoring the need for continuous advancements in this rapidly evolving domain.}, archive = {J_NEUCOM}, author = {Gaurav Kumar and Chhavi Dhiman}, doi = {10.1016/j.neucom.2025.131118}, journal = {Neurocomputing}, month = {11}, pages = {131118}, shortjournal = {Neurocomputing}, title = {Decoding fake news fabrications and trends: A comprehensive survey}, volume = {653}, year = {2025}, }
• (2025). CMCNet:Cross-directional morphology-aware convolution network for chest X-ray anatomy segmentation. NEUCOM, 653, 131112. (www)

  @article{ , abstract = {Chest X-ray (CXR) images pose significant segmentation challenges for neural networks due to their characteristically soft tissue contrasts, ambiguous structural boundaries, and complex anatomical superimpositions. To address the challenges of anatomical variability in Chest X-ray segmentation, we propose the Cross-directional Morphology-aware Convolution Network (CMCNet), which adaptively perceives anatomical structures of varying shapes and sizes and enhances the model’s focus on structures. Specifically, we design a Cross-directional Morphology-Aware Module (CMAM) that introduces learnable offsets into transverse and longitudinal convolutions, allowing the kernels to adjust sampling positions adaptively. By extending these convolutions to three different scales and performing cross-scale pyramid feature fusion, the module effectively captures morphological variations across anatomical structures of different sizes. Additionally, to further enhance the model’s representational capacity across both axial directions, we propose a Dual-Cross Attention Module (DCAM), which computes dual attention across two parallel axial mechanisms by leveraging features from the CMAM, thereby effectively capturing directional information while modeling global context. Extensive experiments conducted on two large-scale medical datasets, CXRS and VinDr-RibCXR, demonstrate that CMCNet achieves superior accuracy and efficiency compared to existing methods in anatomical structure segmentation, delivering competitive performance.}, archive = {J_NEUCOM}, author = {Lili Huang and Yuhan Feng and Xiaowei Zhao and Chenglong Li and Jin Tang}, doi = {10.1016/j.neucom.2025.131112}, journal = {Neurocomputing}, month = {11}, pages = {131112}, shortjournal = {Neurocomputing}, title = {CMCNet:Cross-directional morphology-aware convolution network for chest X-ray anatomy segmentation}, volume = {653}, year = {2025}, }
• (2025). Game-based distributed formation control by multiple relative distances over switching topologies. NEUCOM, 653, 131108. (www)

  @article{ , abstract = {This paper develops a game-based distributed formation control problem over switching topologies based on the relative distances among different agents and between agents and specific targets. With the help of theoretical analyses, it is shown that when the Nash equilibrium (NE) of the formulated game is derived, the desired formation is achieved. To address this formation problem, we propose a distributed controller based on the location estimation technique and gradient descent. By constructing the topology-dependent time-varying Lyapunov function, it is proven that the desired formation is achieved by this controller. Furthermore, concerning the game-based distributed formation control problem subject to local and coupled constraints, another distributed controller is presented based on the projected gradient descent and the Lagrange duality theory, and the relevant theoretical analysis is also provided using topology-dependent time-varying Lyapunov functions. The application to the unmanned surface vehicle (USV) swarm formation is provided to verify the effectiveness of these controllers.}, archive = {J_NEUCOM}, author = {Jiaxun Liu and Shuai Liu and Zhu Wang and Lewei Dong and Xin Hu and Wuquan Li}, doi = {10.1016/j.neucom.2025.131108}, journal = {Neurocomputing}, month = {11}, pages = {131108}, shortjournal = {Neurocomputing}, title = {Game-based distributed formation control by multiple relative distances over switching topologies}, volume = {653}, year = {2025}, }
• (2025). BLAST: Towards robust detection of sleep spindles across clinical settings. NEUCOM, 653, 131107. (www)

  @article{ , abstract = {Detecting sleep spindles in Electroencephalography (EEG) is crucial for assessing cognitive function and neurological disorders. How to provide robust detection across diverse datasets and clinical settings has long remained an open problem. Aiming at this challenge, this study introduces B i L STM- A ttention S pindle ne T ( BLAST ), a deep learning framework focusing on the localization of individual spindle events operating in three phases: (1) Representation : BLAST first enhances morphological and temporal representations through the U-Net to learn hierarchical features, followed by BiLSTM (Bidirectional Long Short-Term Memory) to model the intricate temporal dynamics within EEG; (2) Highlighting : It then enforces attention to spindle bodies with an attention mechanism, which diminishes irrelevant background signals and magnifies salient features that discriminate spindles from other activities; and (3) Detection : These refined features are efficiently integrated through skip connections, preserving both abstract representations and fine-grained details for precise localization of each spindle event based on predictions at each sampling point. BLAST has been extensively evaluated against seven state-of-the-art counterparts (e.g., SEED and CDTSD) using three distinct datasets. The results demonstrate the following: (1) BLAST achieves the best overall performance, with precision, recall, and F1-score of 86.4 %, 86.9 %, and 85.9 %, respectively, on the publicly available MASS dataset; (2) It exhibits the highest consistency with expert evaluation on downstream tasks, such as spindle density and average duration, with correlation coefficients of 0.828 and 0.821, respectively; and (3) Transfer learning on the open-access DREAMS and private BCH-Children datasets further highlights BLAST ’s strong generalization potential across various clinical settings.}, archive = {J_NEUCOM}, author = {Chenyun Guo and Yifeng Ji and Dan Chen and Xiaoli Li and Tengfei Gao and Wei Yu and Zhifei Xu and Mingqi Dong}, doi = {10.1016/j.neucom.2025.131107}, journal = {Neurocomputing}, month = {11}, pages = {131107}, shortjournal = {Neurocomputing}, title = {BLAST: Towards robust detection of sleep spindles across clinical settings}, volume = {653}, year = {2025}, }
• (2025). A cross-feature mutual learning framework integrating multiple features for brain disorder diagnosis. NEUCOM, 653, 131105. (www)

  @article{ , abstract = {The time courses (TC) and functional network connectivity (FNC) features extracted from functional magnetic resonance imaging (fMRI) data demonstrate substantial potential in the investigation of brain disorders. Nevertheless, prevailing diagnostic approaches for brain disorders utilizing fMRI predominantly rely on either single-feature analyses or simplistic feature concatenation strategies, thereby neglecting the complementary information inherent in the interplay between temporal dynamics and spatial connectivity patterns. To address this problem, we propose Cross-Feature Mutual Learning (CFML), a novel framework that enables collaborative learning between feature-specific models through mutual knowledge transfer during training. Unlike traditional data-level fusion or decision-level ensemble methods, CFML introduces three key innovations: (1) a cross-feature mutual learning paradigm where TC and FNC encoders collaboratively exchange knowledge to unlock complementary information, (2) an end-to-end framework that jointly optimizes feature extraction, cross-modal knowledge transfer, and classification without requiring manual fusion design, and (3) a contrastive learning-based feature exchange mechanism that maximizes mutual information between heterogeneous brain feature networks. Extensive experiments on large multi-site datasets demonstrate CFML's superior performance, achieving 85.1 % and 72.3 % accuracy for schizophrenia and autism classification respectively, outperforming 16 state-of-the-art methods by 3.0–9.2 %. Attention analysis highlights biologically relevant areas, confirming the framework's clinical value. In summary, CFML represents a significant advance in neuroimaging-based mental disorder diagnosis with its innovative mutual learning approach.}, archive = {J_NEUCOM}, author = {Xiangxiang Cui and Min Zhao and Dongmei Zhi and Aichen Feng and Yuhui Du and Vince D. Calhoun and Jing Sui}, doi = {10.1016/j.neucom.2025.131105}, journal = {Neurocomputing}, month = {11}, pages = {131105}, shortjournal = {Neurocomputing}, title = {A cross-feature mutual learning framework integrating multiple features for brain disorder diagnosis}, volume = {653}, year = {2025}, }
• (2025). Synthetic short-wave infrared images generated by pixel-to-perceptual multilevel constraint-based GAN architecture for marine object detection. NEUCOM, 653, 131100. (www)

  @article{ , abstract = {Short-wave infrared (SWIR) imagery has already exhibited great power in marine ship detection due to its all-weather monitoring capability, strong penetration of fog, and rich representational details in the imagery, etc. However, the shortage of labeled SWIR datasets limits the usage of deep learning-based methods in marine ship detection, which have powerful feature extraction capabilities, high accuracy and robustness in marine ship detection, but rely on large-scale hand-labeled data for training. To address this problem, we develop an image-to-image framework, which can convert abundant labeled visible imagery data to synthetic SWIR data. We can train end-to-end optimal features for marine ship detection on this massive labeled dataset of synthetic SWIR images. This is, as far as we know, the first study on visible-to-SWIR image translation for marine ship detection. Specifically, the proposed method follows the image-to-image paradigm and the model is optimised with a multi-level loss function: 1) Adversarial losses are introduced to optimise the generator and discriminator, generating synthetic images that closely resemble real ones. 2) Pixel-wise losses optimises the network from the pixel level, measuring the differences between the input and output images on a pixel-by-pixel basis. 3) Patch-wise losses preserve the content in the input and output corresponding patches, facilitating the retention of the main information about the target object in the source domain image. 4) Perceptual losses enable the network to extract the high-level semantic and perceptual information of cross-spectrum images, allowing the generator to produce images that are closer to human perception and the target domain. The proposed method adopts a multi-level feature embedding process to preserve object identities and achieve cross-domain semantic consistency. Experimental results demonstrate the effectiveness of the proposed approach in generating high-quality synthetic SWIR images for marine object detection. The proposed method paves the way for the application of SWIR imaging in marine object detection, and also provides a reference for the generation of image data in other spectral bands (e.g., Terahertz, UV) imaging and other related problems.}, archive = {J_NEUCOM}, author = {Liqian Wang and Yakui Dong and Cheng Fei and Junliang Liu and Shuzhen Fan and Yunxia Liu and Zhaojun Liu and Xian Zhao and Yongfu Li}, doi = {10.1016/j.neucom.2025.131100}, journal = {Neurocomputing}, month = {11}, pages = {131100}, shortjournal = {Neurocomputing}, title = {Synthetic short-wave infrared images generated by pixel-to-perceptual multilevel constraint-based GAN architecture for marine object detection}, volume = {653}, year = {2025}, }
• (2025). Diverse route-driven deep reinforcement learning for vehicle routing problems. NEUCOM, 653, 131099. (www)

  @article{ , abstract = {Recently, there has been a growing propensity to utilize deep reinforcement learning (DRL) for solving a broad range of vehicle routing problems, owing to its data-driven and prior-independent characteristics. However, existing constructive DRL approaches do not adhere to Bellman’s Principle of Optimality completely, since the previously visited nodes that are irrelevant to future decisions still impact subsequent sub-tasks, thereby restricting further improvement of the routing performance. To address this issue, we formulate the DRL-based routing process as a Markov Decision Process (MDP) and propose a novel encoder–decoder framework with three branches of targeted modifications. First, in the phase of encoding, we devise a ranking-based self-attention block for node embedding, which achieves lightweight computational complexity and enables the model to tackle larger-scale problems. Second, during decoding, we explicitly remove the already visited nodes at each step in accordance with the property of MDP (i.e., the future decision is independent of the past given the present), and depict a dynamic-aware context token to capture the state transition. Third, in the training phase, we develop a Jensen–Shannon divergence loss to regularize multiple decoders to output diverse routes. Extensive experiments on both synthetic and real-world datasets demonstrate that our diverse route-driven deep reinforcement learning (DR-DRL) approach achieves state-of-the-art performance, surpassing heuristic and learning-based baselines with tour length reductions of up to 8.94 %. Notably, our DR-DRL also exhibits strong generalization capabilities, underscoring its potential for addressing complex, real-world routing scenarios.}, archive = {J_NEUCOM}, author = {Qingshu Guan and Hui Cao and Tiansen Niu and Lixin Jia and Shuangsi Xue and Badong Chen}, doi = {10.1016/j.neucom.2025.131099}, journal = {Neurocomputing}, month = {11}, pages = {131099}, shortjournal = {Neurocomputing}, title = {Diverse route-driven deep reinforcement learning for vehicle routing problems}, volume = {653}, year = {2025}, }
• (2025). IE-collaborative attention for spatial feature refinement and boundary aware in real-time semantic segmentation. NEUCOM, 653, 131096. (www)

  @article{ , abstract = {Real-time semantic segmentation plays a crucial role in practical applications such as autonomous driving, robotic navigation, and video surveillance. However, existing methods often suffer from insufficient accuracy in high-resolution images due to loss of detail and blurred boundaries, and they struggle to fully integrate global semantic information with local contextual cues. To address these challenges, this paper proposes a novel IE-CRB network that leverages an Internal-External Collaborative Cross-Attention Block (IE-CCAB), a Boundary-Aware Deformable Module (BADS), and a Differentiated Channel-Spatial Attention Module (DCSA) to enable bidirectional inter-layer feature interaction, precise edge detail extraction, and dynamic suppression of background noise. Specifically, the IE-CCAB module captures both global semantics and local details simultaneously via a collaborative cross-attention mechanism between high-level and low-level features, effectively compensating for the shortcomings of traditional methods. The BADS module employs deformable convolutions to focus on target boundary regions, adaptively extracting fine-grained edge features to enhance the model’s boundary awareness; the DCSA module concurrently establishes channel and spatial attention branches and adopts differentiated weighted fusion to emphasize target region features while effectively mitigating background interference. Experimental results on the Cityscapes dataset show that IE-CRB achieves state-of-the-art segmentation accuracy at a comparable model scale while maintaining real-time inference, with IE-CRB-S reaching an mIoU of 79.4 % at 80.0 FPS, and IE-CRB-B attaining 80.6 % at 35.2 FPS. On other benchmarks such as CamVid and ADE20K, IE-CRB consistently maintains a favorable trade-off between mIoU (above 81 % and 44.8 %, respectively) and real-time performance. These results clearly demonstrate the strong generalization and efficiency of our approach.}, archive = {J_NEUCOM}, author = {Ning Li and Xudong Zhang and Bo Li and Baohua Yuan and Gaochao Yang}, doi = {10.1016/j.neucom.2025.131096}, journal = {Neurocomputing}, month = {11}, pages = {131096}, shortjournal = {Neurocomputing}, title = {IE-collaborative attention for spatial feature refinement and boundary aware in real-time semantic segmentation}, volume = {653}, year = {2025}, }
• (2025). Fine-grained grading network based on sparse transformer and spectral attention for multiparametric MR image segmentation. NEUCOM, 653, 131094. (www)

  @article{ , abstract = {Multiparametric magnetic resonance imaging (mpMRI) is crucial in excluding serious diseases by providing clear scans that can quickly locate diseased tissue. Subjective assessment of lesion aggressiveness can vary among clinicians. We propose a network for mpMRI segmentation capable of effectively segmenting fine-grained lesion groups in prostate cancer (PCa), with potential applicability to other magnetic resonance imaging (MRI) segmentation tasks such as brain strokes. Specifically, a bi-oriented bi-level routing transformer (BBRT) with a flexible sparse attention mechanism is designed to enhance the network’s target perception. This design optimizes the Transformer’s long-range modeling capabilities. An adaptive multispectral channel attention (AMSA) module with an automatically selected frequency components mechanism is designed to enhance target contour recognition accuracy. It improves the network’s sensitivity to detect crucial texture information. Experiments are conducted using three MRI datasets. Our method achieved a dice similarity coefficient (DSC) of 76.09 % for grading PCa on PCAMM, demonstrating improvements of 2.77 %, 3.79 %, 3.64 %, and 0.29 % over the suboptimal method in clinically significant Gleason score (GS) groups (GS 3+4, GS 4+3, GS=8, and G S ≥ 9 ), respectively. The prostate region segmentation obtained DSC of 94.03 % and 84.52 % for the central gland (CG) and peripheral zone (PZ), respectively, on PROSTATEx. The stroke segmentation achieved a DSC of 86.33 % on ISLES2022. The experimental results demonstrate that the proposed model excels in generalization and outperforms other state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Yatong Liu and Wei Wang and Yu Zhu and Hangyu Li and Zeyan Zeng and Yuhao Zhang}, doi = {10.1016/j.neucom.2025.131094}, journal = {Neurocomputing}, month = {11}, pages = {131094}, shortjournal = {Neurocomputing}, title = {Fine-grained grading network based on sparse transformer and spectral attention for multiparametric MR image segmentation}, volume = {653}, year = {2025}, }
• (2025). CSPANet: Cross-route statistical partition attention network for style transfer. NEUCOM, 653, 131091. (www)

  @article{ , abstract = {Attention-based style transfer methods have made remarkable advancements in generating high-quality stylized images. However, they often struggle with an essential duality: maintaining precise content structure while reproducing both fine-grained style details and holistic visual concepts from reference artworks. To address these issues, we propose a Cross-Route Statistical Partition Attention Network (CSPANet) for style transfer. Specifically, our proposed CSPANet is a novel stylization framework that can learn cross-route style information for constructing realistic style details of generated stylized images through intuitive feature processing. Next, we design a statistical partition attention migrator (SPAM). It not only can enhance the representation of style semantics correlation by learning global dependency and local awareness, but also depict the captured visual-concept information into the content structure in a coarse-to-fine manner. Furthermore, we construct a detail-aware skip connection (DSC) between the encoder and decoder. It can dynamically inject shallow information into the decoder, improving content detail structure retention and overall stylized quality. Extensive experiments have verified the effectiveness of our proposed CSPANet, surpassing previous state-of-the-art style transfer methods. Code is available: https://github.com/DUT-ZYu/CSPANet-main .}, archive = {J_NEUCOM}, author = {Yu Zhang and Yongsheng Dong and Xiangyu Pei}, doi = {10.1016/j.neucom.2025.131091}, journal = {Neurocomputing}, month = {11}, pages = {131091}, shortjournal = {Neurocomputing}, title = {CSPANet: Cross-route statistical partition attention network for style transfer}, volume = {653}, year = {2025}, }
• (2025). Efficient and dynamic layer-wise structured N:M pruning of deep neural networks. NEUCOM, 653, 131090. (www)

  @article{ , abstract = {Sparsity in deep neural networks (DNNs) has become a key strategy for reducing model complexity and improving execution efficiency, particularly in resource-constrained environments. Pruning techniques, which remove redundant information through structured and unstructured patterns, play a critical role in achieving sparsity. In this paper, we propose USD, a novel hybrid pruning method that seamlessly transitions from unstructured pruning to dynamic layer-wise N : M structured sparsity, where each group of M consecutive weights contains exactly N non-zero elements, effectively combining the strengths of both approaches. Our framework dynamically adjusts the N : M sparsity pattern across layers to optimize both model accuracy and computational efficiency. Extensive experiments, on various CNN and Transformer-based architectures using popular datasets such as CIFAR and ImageNet, demonstrate that our hybrid approach outperforms state-of-the-art pruning techniques, achieving superior accuracy and efficiency. The code is publicly available at https://github.com/usdsparsity/USD .}, archive = {J_NEUCOM}, author = {Dehbia Ahmed Zaid and Badis Djamaa and Mohamed Akrem Benatia}, doi = {10.1016/j.neucom.2025.131090}, journal = {Neurocomputing}, month = {11}, pages = {131090}, shortjournal = {Neurocomputing}, title = {Efficient and dynamic layer-wise structured N:M pruning of deep neural networks}, volume = {653}, year = {2025}, }
• (2025). Degradation-resilient transient control of aero-engines via neural dynamic programming: A model-free framework for rate-constrained safety. NEUCOM, 653, 131079. (www)

  @article{ , abstract = {This paper proposes a model-free neural dynamic programming framework to address the coupled challenges of input safety and degradation adaptation in aero-engine transient control. First, by designing dynamic control laws and leveraging the Bellman optimality principle, the optimal control problem for acceleration transients under input safety constraints is reformulated as a Hamilton-Jacobi-Bellman (HJB) equation. Building on this, we develop an input-constrained neural dynamic programming framework through actor-critic architecture, in which the high-dimensional HJB equation is decomposed into learnable nonlinear optimization hierarchies. Within this framework, a safety-embedded actor network dynamically regulates the control policy to enforce predefined safety constraints, thereby preventing excessive and rapid fuel flow that could cause engine stall or surge during transients. Extensive simulations on the JT9D virtual prototype demonstrate the effectiveness of the framework under both nominal and degradation conditions. Compared to intelligent PID algorithms, the proposed method reduces settling time by 25 − 38 % and computational time by 80 %, while maintaining high-quality transient response performance. By unifying data-driven learning with constrained optimization, this work advances safe control paradigms for complex aerospace systems and offers a practical alternative to model-dependent strategies.}, archive = {J_NEUCOM}, author = {Shuoshuo Liu and Tao Sun and Peng Li and Xudong Zhao}, doi = {10.1016/j.neucom.2025.131079}, journal = {Neurocomputing}, month = {11}, pages = {131079}, shortjournal = {Neurocomputing}, title = {Degradation-resilient transient control of aero-engines via neural dynamic programming: A model-free framework for rate-constrained safety}, volume = {653}, year = {2025}, }
• (2025). High-performance inference graph convolutional networks for skeleton-based action recognition. NEUCOM, 653, 131078. (www)

  @article{ , abstract = {Recently, the significant achievements have been made in skeleton-based human action recognition with the emergence of graph convolutional networks (GCNs). However, the state-of-the-art (SOTA) models used for this task focus on constructing more complex higher-order connections between joint nodes to describe skeleton information, which leads to complex inference processes and high computational costs. To address the slow inference speed caused by overly complex model structures, we introduce re-parameterization and over-parameterization techniques to GCNs and propose two novel high-performance inference GCNs, namely HPI-GCN-RP and HPI-GCN-OP. After the completion of model training, model parameters are fixed. HPI-GCN-RP adopts re-parameterization technique to transform high-performance training model into fast inference model through linear transformations, which achieves a higher inference speed with competitive model performance. HPI-GCN-OP further utilizes over-parameterization technique to achieve higher performance improvement by introducing additional inference parameters, albeit with slightly decreased inference speed. The experimental results on the two skeleton-based action recognition datasets demonstrate the effectiveness of our approach. Our HPI-GCN-OP achieves performance comparable to the SOTA models, with inference speeds five times faster. Specifically, our HPI-GCN-OP achieves an accuracy of 93 % on the cross-subject split of the NTU-RGB+D 60 dataset, and 90.1 % on the cross-subject benchmark of the NTU-RGB+D 120 dataset. Code is available at github.com/lizaowo/HPI-GCN .}, archive = {J_NEUCOM}, author = {Junyi Wang and Ziao Li and Bangli Liu and Haibin Cai and Mohamad Saada and Qinggang Meng}, doi = {10.1016/j.neucom.2025.131078}, journal = {Neurocomputing}, month = {11}, pages = {131078}, shortjournal = {Neurocomputing}, title = {High-performance inference graph convolutional networks for skeleton-based action recognition}, volume = {653}, year = {2025}, }
• (2025). Active metric learning with hybrid information for ordinal classification. NEUCOM, 653, 131076. (www)

  @article{ , abstract = {In the context of ordinal classification, combining absolute information (examples with class labels) and relative information (couples of examples with preference orders) has recently been shown to be an effective way to build performant models. Usually, sufficient amounts of labelled data are needed to develop good ordinal classification models. In order to reduce the cost and time of labelling, a common alternative is to exploit active learning strategies. In this paper, firstly, we propose a semi-supervised distance metric learning method for ordinal classification with absolute information, relative information and unlabelled data. Subsequently, we propose several active metric learning strategies to select informative and representative examples, couples and triplets. Finally, we test these strategies on some classical ordinal classification datasets. The experimental results show the effectiveness of these strategies.}, archive = {J_NEUCOM}, author = {Mengzi Tang and Bernard De Baets}, doi = {10.1016/j.neucom.2025.131076}, journal = {Neurocomputing}, month = {11}, pages = {131076}, shortjournal = {Neurocomputing}, title = {Active metric learning with hybrid information for ordinal classification}, volume = {653}, year = {2025}, }
• (2025). Memory augmented coherent probabilistic forecasts for hierarchically related time series. NEUCOM, 653, 131075. (www)

  @article{ , abstract = {Coherent probabilistic forecasting, which involves predicting multivariate time series with hierarchical aggregation, is vital for accurate decision-making in industrial applications. Recent advancements have emphasized end-to-end approaches that simultaneously learn from all-time series in the hierarchy while integrating the reconciliation step into a single trainable model. However, prior studies often neglect the distinct characteristics of time series at different hierarchical levels. For instance, time series at the lower hierarchy levels are typically sparse and lack the high-level patterns evident in aggregated levels. Notably, in a hierarchical time series, the aggregated series (parent) is the sum of its lower-level series (child), thereby retaining intrinsic time series properties. Motivated by these observations, we propose hierarchical time series forecasting through Hierarchy-Aware Memory Network (HAMN). HAMN employs a hierarchy-aware memory module that stores representations of aggregated series at each hierarchical level in an external memory module and retrieves these representations to augment features of the sparse lower-level series. Evaluation across four public datasets demonstrates improvements of 6.4 %-20.9 % over state-of-the-art baselines in terms of the CRPS metric.}, archive = {J_NEUCOM}, author = {Junyong Lee and Yunseon Byun and Byoungmo Koo and Jun-Geol Baek}, doi = {10.1016/j.neucom.2025.131075}, journal = {Neurocomputing}, month = {11}, pages = {131075}, shortjournal = {Neurocomputing}, title = {Memory augmented coherent probabilistic forecasts for hierarchically related time series}, volume = {653}, year = {2025}, }
• (2025). Improving depression diagnosis using a brain module-based weighted hypergraph convolutional network framework. NEUCOM, 653, 131074. (www)

  @article{ , abstract = {Traditional functional connectivity network (FCN) is limited to capturing pairwise correlations between electrodes or brain regions. In contrast, hyper-connectivity network (HCN), which can capture complex higher-order relationships among brain regions, has gained increasing attention in brain disease diagnosis. While existing HCN-based models have shown effectiveness in disease recognition, several challenges remain. Reliable HCN construction is hindered by methodological constraints and electroencephalogram (EEG) noise. Moreover, these models often underutilize disease-related abnormal neural mechanisms, resulting in limited interpretability and suboptimal recognition performance. To address these issues, we propose a brain module-based adaptive weighted hypergraph convolutional network (BM-WHGCN) framework to improve depression diagnosis. Firstly, node-level feature representations of EEG are comprehensively extracted from temporal, frequency, and randomness domains. Secondly, a local synchronicity- constrained sparse method is proposed to construct HCN that integrates both global and local correlations. Subsequently, densely connected node subsets are identified as brain modules through modularity optimization, and long-range connections are attenuated to develop a learning framework with enhanced neuropathological interpretability. Finally, an extended hypergraph convolutional network aggregates node features and hyperedges to generate a deep representation of the spatiotemporal information. Extensive experiments conducted on a public and our own EEG dataset demonstrate that BM-WHGCN outperforms other state-of-the-art methods for depression recognition. Additionally, our framework demonstrates strong interpretability by effectively identifying hyperedges and electrode nodes closely associated with depression. In conclusion, the proposed method offers significant potential for the early detection and intervention of depression and can be readily adapted for the diagnosis of other brain disorders.}, archive = {J_NEUCOM}, author = {Na Li and Ziyi Wang and Zhenwen Zhang and Jun Huang and Jing Zhu and Xiaowei Li and Bin Hu}, doi = {10.1016/j.neucom.2025.131074}, journal = {Neurocomputing}, month = {11}, pages = {131074}, shortjournal = {Neurocomputing}, title = {Improving depression diagnosis using a brain module-based weighted hypergraph convolutional network framework}, volume = {653}, year = {2025}, }
• (2025). Modeling hierarchical structure for conversational aspect-based quadruple sentiment analysis. NEUCOM, 653, 131055. (www)

  @article{ , abstract = {Conversational Aspect-Based Quadruple Sentiment Analysis (DiaASQ) aims to extract structured sentiment quadruples—comprising target, aspect, opinion, and sentiment polarity—from multi-turn dialogues. This task presents significant challenges, as the elements of a quadruple often span across different utterances and involve multiple speakers, necessitating both fine-grained syntactic comprehension and global discourse-level reasoning. Existing methods typically process utterances in isolation or rely on shallow context modeling, which limits their ability to capture long-range dependencies and structural coherence in conversation flows. To address these challenges, we propose HierASQ , a hierarchical structure-aware framework designed to jointly model intra-utterance syntax and inter-utterance discourse structure. Specifically, we leverage syntactic dependency trees and dialogue thread structures to construct hybrid graphs, which are then encoded by a Hierarchical Graph Attention Network (HGAT) to capture both local and global interactions via tailored message passing. To further enhance representation integration, we introduce a dynamic hierarchical aggregation (DHA) mechanism that progressively fuses multi-scale contextual signals from utterance windows. Finally, we develop a high-order decoding strategy that enables global reasoning over token pairs to extract structurally consistent sentiment quadruples. Extensive experiments on two benchmark DiaASQ datasets—Chinese and English—demonstrate that our approach significantly outperforms strong baselines, particularly on challenging cross-utterance and overlapping scenarios. Qualitative analyses further validate the effectiveness of our model in capturing nuanced discourse patterns and sentiment dynamics in realistic dialogue settings. The future work includes expanding HierASQ in several directions: LLM integration, hybrid systems, multilingual support, and pedagogical use cases.}, archive = {J_NEUCOM}, author = {Ling Zhao and Da Shen and Ying Liu}, doi = {10.1016/j.neucom.2025.131055}, journal = {Neurocomputing}, month = {11}, pages = {131055}, shortjournal = {Neurocomputing}, title = {Modeling hierarchical structure for conversational aspect-based quadruple sentiment analysis}, volume = {653}, year = {2025}, }
• (2025). MSSCR: Multi-scale semantic collaborative reasoning model for explainable multimodal rumor detection. NEUCOM, 653, 131042. (www)

  @article{ , abstract = {Multimodal rumors, which leverage visual “evidence” to enhance their credibility, are more deceptive than textual rumors on online social media. This makes multimodal rumor detection a critical area of research in the field of rumor analysis. While existing methods primarily rely on the coarse-grained alignment between images and text, the increasing sophistication of image generation and manipulation technologies has rendered such approaches inadequate for precise rumor detection. To address this challenge, we propose a Multi-Scale Semantic Collaborative Reasoning model (MSSCR) for explainable multimodal rumor detection. Our model explores complementary relationships from macro, meso, and micro perspectives. At the macro level, a co-attention mechanism is employed to align and fuse cross-modal global semantics, capturing holistic relationships among the post, its comments, and the associated image. At the meso level, we construct graph structures to model local interactions between the post and its comments, as well as between the image and the comments. The Graph Attention Network (GAT) is utilized to extract structured local semantic features. At the micro level, a Large Language Model (LLM) is used to extract fine-grained semantics, including keyword and sentiment consistency, to identify subtle semantic discrepancies. Finally, a dynamic weighting mechanism integrates multi-scale features, and an additional fine-grained semantic consistency score derived from micro-level semantics further refines the final rumor prediction. Experimental results on two real-world datasets, PHEME and Weibo, demonstrate that our model achieves improvements of 1.08 % and 0.82 % over the best baseline method, respectively.}, archive = {J_NEUCOM}, author = {Xue Yuan and Anhao Shu and Yue Wu and Jiyu Liu and Min Wang and Jiahong Liu}, doi = {10.1016/j.neucom.2025.131042}, journal = {Neurocomputing}, month = {11}, pages = {131042}, shortjournal = {Neurocomputing}, title = {MSSCR: Multi-scale semantic collaborative reasoning model for explainable multimodal rumor detection}, volume = {653}, year = {2025}, }
• (2025). Self-ensemble for test time adaptation. NEUCOM, 653, 131038. (www)

  @article{ , abstract = {Fully test-time adaptation (TTA) aims to address the cross-domain performance degradation problem by adapting a source pre-trained model on test data. Ensemble methods have been shown to be very effective for out-of-distribution (OOD) uncertainty estimation, and are promising for TTA. However, naive ensembling of multiple models would significantly increase the overall model size and computational load. This work explores more efficient ensemble approaches for TTA that can improve the cross-domain performance but without significant increase of the overall model size and computational load. To this end, our contribution is twofold. First, we show that, compared with using a single model, an ensemble of multiple smaller models with fewer or comparable overall parameters can also markedly improve TTA performance. Second, we propose a novel self-ensemble method that ensembling is performed on multiple intermediate outputs of a single model rather than on multiple independent models. The proposed method only incurs slight increase in parameters, and can be readily incorporated into existing TTA methods in a plug-in manner. Experimental results on benchmark datasets demonstrate that the proposed self-ensemble method can significantly improve the performance of existing state-of-the-art TTA methods. Furthermore, when combining self-ensemble with the ensembling of smaller models, even greater improvement can be achieved. Code is available at: https://github.com/BryantLeedy/SelfEnsembleTTA.}, archive = {J_NEUCOM}, author = {Dongyang Li and Liujia Ma and Mingyue Qin and Peilin Liu and Fei Wen}, doi = {10.1016/j.neucom.2025.131038}, journal = {Neurocomputing}, month = {11}, pages = {131038}, shortjournal = {Neurocomputing}, title = {Self-ensemble for test time adaptation}, volume = {653}, year = {2025}, }
• (2025). DSM-CLIP: A framework designed for hard negative sampling in generalizable person re-identification. NEUCOM, 653, 130920. (www)

  @article{ , abstract = {Domain Generalization Person Re-identification (DG-ReID) aims to retrieve the same person across different domains and unseen scenarios. Existing methods employ single-modal models for feature extraction, which struggle to capture rich cross-domain semantic information. The vision-language model CLIP, pre-trained on large-scale coarse-grained image-text pairs, facilitates DG-ReID models in capturing diverse high-level semantics, thereby improving performance. However, coarse-grained image-text pairs may blur background and detailed features of images, hindering the model’s ability to generalize effectively in DG-ReID tasks. To address these limitations, this paper proposes a novel Dynamic Semantic Memory framework (DSM-CLIP) to tackle the domain shift problem in DG-ReID. First, a Dual-modal Hard Negative Sampling (DHS) method based on CLIP is introduced, leveraging complementary dual-modal information to generate hard negative samples with highly similar foregrounds and backgrounds in batches, which can make the model focus on fine-grained differences in images while suppressing identity-irrelevant background features. To further enhance the model’s discriminative capability for hard negative samples, a dynamic semantic memory strategy is adopted to refine textual representations, thereby strengthening text embeddings by capturing subtle differences between images of the same identity and updating textual centroids iteratively. Experimental results demonstrate that the proposed method achieves state-of-the-art recognition performance across multiple protocols, effectively mitigates domain shift issues, and enhances the model’s generalization capability.}, archive = {J_NEUCOM}, author = {Yingchun Guo and Tianyang Xu and Yi Liu and Liwen Zhang}, doi = {10.1016/j.neucom.2025.130920}, journal = {Neurocomputing}, month = {11}, pages = {130920}, shortjournal = {Neurocomputing}, title = {DSM-CLIP: A framework designed for hard negative sampling in generalizable person re-identification}, volume = {653}, year = {2025}, }
• (2025). SEMKR: Joint learning of semantic and topological representations for knowledge graph completion. NEUCOM, 653, 130909. (www)

  @article{ , abstract = {Knowledge Graph Completion (KGC) methods concentrate on predicting unknown components in the (head entity) ⟶ [ r e l a t i o n ] (tail entity) triplet. However, large-scale Knowledge Graphs (KGs), especially those in legal domains, often face the problem of imbalanced data distribution, which brings significant challenges to achieve accurate prediction. For instance, high-frequency criminal charges generally produce well-structured and densely connected sub-graphs, whereas low-frequency/confusing criminal charges are sparsely connected. Current KGC methods learn individual features, such as relation types or sub-graph structures, while overlooking joint embedding associations. To address these challenges, we propose a text-guided graph reasoning mechanism that integrates semantic and topological representations synergistically. Our model, SEMKR , SEM antic and topological representations for K nowledge g R aph completion, incorporate two key modules: (1) Entity Description Representation Learning enhances semantic features by pre-train and fine-tune BERT; (2) Relational Topological Representation Learning refines context representations by learning edge and path information in two stage pre-training and fine-tuning. Experimental results demonstrate that SEMKR outperforms state-of-the-art baselines in relation or link prediction tasks, achieving significant improvements, including over 7.8% H@1 on the NELL995 dataset. All implementations will be available at https://github.com/SUSTech-TP/SEMKR.git .}, archive = {J_NEUCOM}, author = {Pengjie Liu and Wang Zhang and Yulong Ding and Jie Jiang and Shuang-Hua Yang}, doi = {10.1016/j.neucom.2025.130909}, journal = {Neurocomputing}, month = {11}, pages = {130909}, shortjournal = {Neurocomputing}, title = {SEMKR: Joint learning of semantic and topological representations for knowledge graph completion}, volume = {653}, year = {2025}, }
• (2025). DFAN++: Enhanced triple-branch network for generalized zero-shot image classification. NEUCOM, 653, 130880. (www)

  @article{ , abstract = {Generalized zero-shot image classification aims to transfer knowledge from seen data to unseen categories and correctly distinguish both seen and unseen data. Most direct attribute prediction (DAP) methods focus on projecting visual features into attribute space and aligning them with class prototypes. However, domain shifts arise during the dataset acquisition due to camera angles or occlusion. Using the class prototype directly cannot accurately represent each sample and will mislead the alignment of visual features and attributes. Moreover, the regularly used visual feature extraction backbone, such as ResNet, is trained on ImageNet, resulting in cross-dataset bias for the GZSL task. Therefore, this paper proposes an Enhanced Triple-branch Network, referred to as DFAN++, an enhanced version of Dual Feature Augmentation Network (DFAN), which integrates Evolving Attribute Feature Learning (EAFL), a Cross-dataset Denoising Module (CDM), and Class Prototype Purification (CPP). EAFL progressively refine attribute features by correcting the semantic prototypes and profoundly explores the relationship between individual and group attribute features. CDM uses a mask to mitigate cross-dataset bias between ImageNet and GZSL benchmarks. CPP learns domain shifts of attributes, guaranteeing the predicted attributes align with specific samples. Furthermore, DFAN++ is conducted on four datasets, demonstrating superior performance compared to some other methods in Generalized Zero-Shot Learning (GZSL). Specifically, on the CUB dataset, the proposed DFAN++ shows an advantage of 1.3% in the GZSL compared to the second-best method. DFAN++ has demonstrated notable advancements compared to DFAN, achieving respective improvements of 2.8% to 3.7% across four benchmark datasets. The codes are available at https://github.com/GAInuist/DFANpp .}, archive = {J_NEUCOM}, author = {Yifan Zhu and Lei Xiang and Yuan Zhou and Haoran Duan and Yang Long}, doi = {10.1016/j.neucom.2025.130880}, journal = {Neurocomputing}, month = {11}, pages = {130880}, shortjournal = {Neurocomputing}, title = {DFAN++: Enhanced triple-branch network for generalized zero-shot image classification}, volume = {653}, year = {2025}, }
• (2025). Finite-time stability analysis of fractional-order neural networks with inertial terms and fuzzy logics. NEUCOM, 652, 131167. (www)

  @article{ , abstract = {Finite-time stability constitutes a fundamental requirement in fractional-order neural network analysis, determining whether a system achieves exponential convergence to equilibrium within prescribed time. Nevertheless, the mathematical foundations of finite-time stability criteria face significant challenges due to the non-commutativity of fractional operators particularly the failure of integer-order chain rules in fractional calculus. This theoretical limitation prevents the direct extension of classical finite-time stability differential inequalities to fractional-order. To address this challenge, we establish two novel finite-time differential inequalities to facilitate stability analysis of fractional-order neural networks. We prove that the system will converge to the equilibrium point in finite time when the fractional derivative of the Lyapunov function satisfies specific nonlinear constraints. Furthermore, we develop two innovative methods for estimating convergence time under these conditions. Additionally, we analyze finite-time stability in fractional-order fuzzy inertial neural networks by applying the proposed discriminant methods and designing appropriate controllers. The finite-time stability criteria of fractional-order fuzzy inertial neural networks are obtained. Finally, numerical simulations validate both the efficacy of our methodology and the precision of convergence time estimates.}, archive = {J_NEUCOM}, author = {Wenxiang Fang and Zhaohai Ma and Fuxiang Li and Hairong Lian}, doi = {10.1016/j.neucom.2025.131167}, journal = {Neurocomputing}, month = {11}, pages = {131167}, shortjournal = {Neurocomputing}, title = {Finite-time stability analysis of fractional-order neural networks with inertial terms and fuzzy logics}, volume = {652}, year = {2025}, }
• (2025). The role of nonlinearity, dimensionality, memory, and input on a mechanical oscillator reservoir computer. NEUCOM, 652, 131158. (www)

  @article{ , abstract = {Physical Reservoir Computing (PRC) leverages the inherent nonlinear dynamics of physical systems for efficient information processing. However, there is a lack of understanding of how nonlinearity in the dynamics, high dimensionality, memory, and different input conditions contribute to information processing capability and thereby influences the PRC performance. This study investigates these key properties within a mechanical spring-mass-damper (SMD) network to understand their influences on computational power for a state estimation task. Spectral analysis and Singular Value Decomposition (SVD) are applied to the system’s readouts to examine dynamic complexity and identify dominant dynamic information. Our results indicate that optimal PRC performance is achieved within specific ranges of nonlinearity, dimensionality, memory, and input conditions. A successful mechanical reservoir computer requires: 1) sufficient nonlinearity in the dynamics to separate input into diverse frequency contents and linearly independent information; 2) adequate number of readout nodes to present this information through the high dimensionality; 3) balanced damping to preserve fading memory; and 4) external input amplitude and frequency carefully tuned to complement system’s inherent nonlinearity and resonance effects. This study highlights the critical interplay between these properties for effective information processing in mechanical reservoirs, enhancing the understanding of the mechanisms that contribute to PRC performance. These insights into the dynamical principles aid in evaluating and designing optimal mechanical systems as effective physical reservoir computers for any given input condition.}, archive = {J_NEUCOM}, author = {Shan He and Patrick F. Musgrave}, doi = {10.1016/j.neucom.2025.131158}, journal = {Neurocomputing}, month = {11}, pages = {131158}, shortjournal = {Neurocomputing}, title = {The role of nonlinearity, dimensionality, memory, and input on a mechanical oscillator reservoir computer}, volume = {652}, year = {2025}, }
• (2025). A survey of deep learning for face presentation attack detection. NEUCOM, 652, 131136. (www)

  @article{ , abstract = {Face anti-spoofing detection (FASD) has become a crucial technology due to the alarming advancements in presentation attacks (PAs). As more novel PAs with realistic generative capabilities emerge, improved biometric security solutions are needed to address these evolving threats. In early and foundational work, FASD techniques focused mainly on handcrafted features that were unreliable due to their limited representation capacity. In the recent decade, with the advancements in deep learning and its capabilities for image processing tasks and the emergence of large datasets, improvements in the performance of detecting PAs have been achieved. However, as research progresses rapidly, there is an absence of a comprehensive analysis of detection methods to understand the strengths and weaknesses of different types of approaches. Although various approaches utilise sensors in addition to RGB cameras, in this paper, we focus specifically on RGB camera-based methods and provide a comprehensive review of deep learning-based FASDs, including generalised models developed to date. In addition, the datasets are presented, along with the evaluation protocols and metrics.}, archive = {J_NEUCOM}, author = {Mohammadreza Sheikhfathollahi and Simon Parkinson and Richard Hill and Saad Khan}, doi = {10.1016/j.neucom.2025.131136}, journal = {Neurocomputing}, month = {11}, pages = {131136}, shortjournal = {Neurocomputing}, title = {A survey of deep learning for face presentation attack detection}, volume = {652}, year = {2025}, }
• (2025). PolaCount: Text-specified zero-shot object counting with pyramid polarity-aware cross-attention. NEUCOM, 652, 131135. (www)

  @article{ , abstract = {Accurate open-world object counting demands fine-grained cross-modal alignment and scale-aware reasoning. We introduce PolaCount, a zero-shot counting framework that leverages a Pyramid Polarity-Aware Cross-Attention Module (PPCAM) to bidirectionally fuse textual and visual features at multiple resolutions. PPCAM divides queries and keys into positive and negative subspaces and applies dual cross-attention, generating sharply focused density maps. Coupled with high-quality DINOv2 visual representations and trained end-to-end using density supervision alone, PolaCount achieves state-of-the-art performance on FSC-147, CARPK, and ShanghaiTech datasets without exemplar boxes or class-specific fine-tuning.}, archive = {J_NEUCOM}, author = {Cheng Qian and Jiwu Cao and Ying Mao and Kai Liu and Peng Zhu and Jun Sang}, doi = {10.1016/j.neucom.2025.131135}, journal = {Neurocomputing}, month = {11}, pages = {131135}, shortjournal = {Neurocomputing}, title = {PolaCount: Text-specified zero-shot object counting with pyramid polarity-aware cross-attention}, volume = {652}, year = {2025}, }
• (2025). M3FPolypSegNet++: Adaptive gaussian frequency decomposition with multi-task learning for polyp segmentation. NEUCOM, 652, 131134. (www)

  @article{ , abstract = {Colorectal cancer remains one of the most prevalent and lethal malignancies globally, primarily due to challenges in timely detection and removal of polyps. Recent CNN- and Transformer-based frameworks have already reported encouraging polyp segmentation accuracy on public benchmarks. Yet no prior approach provides an adaptive, computationally efficient mechanism that jointly leverages frequency cues and global context, leaving robustness under heterogeneous clinical conditions largely unaddressed. To bridge this gap, we design M3FPolypSegNet + + by coupling a learnable Gaussian Adaptive Frequency Atrous Spatial Pyramid Pooling (GAF-ASPP) module with a Transformer backbone, unifying context-rich and frequency-aware features within a lightweight framework suitable for real-time deployment. The GAF-ASPP module adaptively separates feature maps into distinct low- and high-frequency components via a learnable Gaussian filter. Unlike the traditional box-filter with static weight approaches (M3FPolypSegNet), the GAF-ASPP dynamically highlights critical frequency bands according to the input image, effectively reducing artifacts and enhancing segmentation performance. Furthermore, we integrate a Transformer-based encoder, enabling more comprehensive global context modeling and improved representation of long-range dependencies. Extensive experiments on multiple benchmark datasets demonstrate that M3FPolypSegNet + + consistently outperforms state-of-the-art methods across both seen and unseen clinical settings.}, archive = {J_NEUCOM}, author = {Ju-Hyeon Nam and Dong-Hyeon Kang and Sang-Chul Lee}, doi = {10.1016/j.neucom.2025.131134}, journal = {Neurocomputing}, month = {11}, pages = {131134}, shortjournal = {Neurocomputing}, title = {M3FPolypSegNet++: Adaptive gaussian frequency decomposition with multi-task learning for polyp segmentation}, volume = {652}, year = {2025}, }
• (2025). Prescribed performance tracking control of autonomous surface vehicles with reinforcement learning. NEUCOM, 652, 131133. (www)

  @article{ , abstract = {This paper achieves prescribed performance tracking control of an autonomous surface vehicle (ASV) based on a novel reinforcement learning- and sliding mode-based (RLSMB) control algorithm, which is proposed by combining a predefined-time nonsingular controller (PTNC) with a critic-actor neural network-based reinforcement learning (NNBRL) and a novel prescribed performance sliding mode (PPSM). The PPSM based on a predefined-time sliding mode, by integrating a singularity avoidance function, prevents the control input from singularity in predefined-time control, and by integrating a performance constraint function, makes the tracking error stays within an asymmetric time-varying boundary. Compared to the backstepping-based prescribed performance control, the tracking error with the asymmetric time-varying constraint is creatively achieved based on the sliding mode technique, which facilitates the design of the PTNC. Under several sufficient criteria and Lyapunov functions, the theoretical analysis of the RLSMB control algorithm is rigorously conducted, without involving the assumption that the approximation of the actor’s optimal weight is bounded. Finally, the RLSMB control algorithm is applied to the Cyber-Ship II model to verify its efficiency in the numerical simulation.}, archive = {J_NEUCOM}, author = {Jia-Du Zhang and Tao Han and Bo Xiao and Xi-Sheng Zhan}, doi = {10.1016/j.neucom.2025.131133}, journal = {Neurocomputing}, month = {11}, pages = {131133}, shortjournal = {Neurocomputing}, title = {Prescribed performance tracking control of autonomous surface vehicles with reinforcement learning}, volume = {652}, year = {2025}, }
• (2025). Weakly supervised semantic segmentation with multi-task learning and segment anything model. NEUCOM, 652, 131131. (www)

  @article{ , abstract = {Weakly supervised semantic segmentation (WSSS) aims to train segmentation networks using easily available weak annotations, thereby reducing the annotation burden. However, many existing methods encounter two significant bottlenecks: (1) sparse localization of class activation maps (CAMs) and (2) misleading supervision from noisy pseudo annotations. To address these problems, we propose a novel WSSS framework that utilizes multi-task learning and segment anything model (SAM) to activate more complete object regions and robustly train segmentation networks in noisy environments. Specifically, we propose a multi-task CAMs generation network (MCGN), designed to perform the image classification (IC), superpixel segmentation (SS) and semantic affinity prediction (SAP) tasks. The IC task is employed as the primary objective to capture discriminative features, while the latter two tasks function as the auxiliary objectives to capture whole features by learning pixel semantic context. This method enhances the interaction between image-level and pixel-level semantic information, thereby alleviating the sparsity problem of CAMs. Moreover, we propose a knowledge-guided training strategy (KTS), developed to exploit boundary knowledge inferred from SAM to evaluate the confidence of pseudo annotations, which enables the utilization of reliable annotations to train segmentation networks by filtering low-confidence pseudo annotations, thereby enhancing the network’s robustness against noise interference. Extensive quantitative and visual experiments on multiple datasets demonstrate that our method outperforms the existing advanced methods. Our source code is available at: https://github.com/CHENDL-SHEN/MTKL .}, archive = {J_NEUCOM}, author = {Zheng Xue and Dinghao Guo and Dali Chen and Shixin Liu}, doi = {10.1016/j.neucom.2025.131131}, journal = {Neurocomputing}, month = {11}, pages = {131131}, shortjournal = {Neurocomputing}, title = {Weakly supervised semantic segmentation with multi-task learning and segment anything model}, volume = {652}, year = {2025}, }
• (2025). UNICOM: Unified, foreground-aware, and context-realistic deep image composition with diffusion model. NEUCOM, 652, 131129. (www)

  @article{ , abstract = {Image composition aims to combine a given foreground with a background to form a realistic composite image. Prior art decomposes this task into several sub-tasks such as object placement and image harmonization, which require mutual coordination between the foreground and background. Recent diffusion-based image synthesis methods, while combining these components in an end-to-end framework, still require mask annotations to specify target locations, making large-scale deployment impractical. To overcome this limitation, we propose UNIfied deep image COMposition (UNICOM), the first framework to achieve complete sub-task unification without positional masks. By introducing the background prompt, UNICOM can jointly model the foreground posture and the contextual realism in a self-supervised manner, trained on off-the-shelf segmentation datasets without extra manual labeling. Extensive experiments show that UNICOM produces foreground-aware and context-realistic compositions and outperforms relevant baselines in both realism and faithfulness metrics, avoiding exact selection of matched foreground–background pairs.}, archive = {J_NEUCOM}, author = {Lu Yang and Yuanhao Wang and Yicheng Liu and Enze Wang and Ziyang Zhao and Yanqi He and Zexian Song and Hao Lu}, doi = {10.1016/j.neucom.2025.131129}, journal = {Neurocomputing}, month = {11}, pages = {131129}, shortjournal = {Neurocomputing}, title = {UNICOM: Unified, foreground-aware, and context-realistic deep image composition with diffusion model}, volume = {652}, year = {2025}, }
• (2025). Advances on stability analysis for discrete-time markovian jump impulsive stochastic systems with hybrid switching dynamics. NEUCOM, 652, 131128. (www)

  @article{ , abstract = {This paper investigates the p th moment globally asymptotically stable ( p -MGAS) of discrete-time Markovian jump impulsive stochastic systems, which are crucial for modeling complex dynamics influenced by impulsive effect and Markovian jump. We propose a new framework using Lyapunov functions and stochastic analysis to derive stability conditions, incorporating both impulsive and non-impulsive time-varying parameters. The average impulsive interval method is used to handle impulsive effects more flexibly. The results are applied to neural networks with Markovian jumps and impulses via linear matrix inequality (LMI), demonstrating practical relevance. Two numerical examples verify the effectiveness of our methods. This study offers a robust approach to stability analysis for such complex systems.}, archive = {J_NEUCOM}, author = {Wei Shao and Ting Cai and Xiaohui Yan}, doi = {10.1016/j.neucom.2025.131128}, journal = {Neurocomputing}, month = {11}, pages = {131128}, shortjournal = {Neurocomputing}, title = {Advances on stability analysis for discrete-time markovian jump impulsive stochastic systems with hybrid switching dynamics}, volume = {652}, year = {2025}, }
• (2025). RDFCNet: RGB-guided depth feature calibration network for RGB-D salient object detection. NEUCOM, 652, 131127. (www)

  @article{ , abstract = {RGB-D salient object detection aims to fully utilize the complementarity between RGB and depth information to improve detection accuracy, but existing methods still suffer from modal inconsistency, low-quality depth data and information loss in cross-modal feature interaction and fusion. This paper proposes a novel network termed RDFCNet to effectively capture RGB-depth synergy through RGB-guided depth feature calibration and multi-level fusion strategies. Specifically, RDFCNet employs a dual-stream encoder architecture to extract multi-scale hierarchical features, facilitating comprehensive cross-modal feature learning. First, the cross-modal calibration module (CMCM) adaptively refines depth features by aligning them with semantic priors extracted from RGB, mitigating the spatial and semantic gap between modalities. Second, the cross-modal attention module (CMAM) establishes bidirectional, semantic-level dependencies between RGB and depth through a guided attention mechanism, enhancing inter-modal interaction beyond simple concatenation or unidirectional fusion. Third, the dual attention fusion module (DAFM) jointly models channel-wise and spatial salient cues via a effective dual-branch attention structure, reinforcing the integrity and sharpness of salient object representations. Finally, a transformer decoder refines spatial details at the pixel level to achieve precise saliency prediction. Experiments were conducted on multiple RGB-D salient object detection benchmark datasets, and the results demonstrate that RDFCNet outperforms existing state-of-the-art methods across various metrics, exhibiting enhanced capabilities in salient object localization, boundary preservation, and generalization performance.}, archive = {J_NEUCOM}, author = {Weiyi Wei and Yongxin Shi and Yibin Wang and Sixuan Liu}, doi = {10.1016/j.neucom.2025.131127}, journal = {Neurocomputing}, month = {11}, pages = {131127}, shortjournal = {Neurocomputing}, title = {RDFCNet: RGB-guided depth feature calibration network for RGB-D salient object detection}, volume = {652}, year = {2025}, }
• (2025). STSFF-net: A two-stream network for depression detection from facial expressions. NEUCOM, 652, 131126. (www)

  @article{ , abstract = {Depression is a prevalent mental health condition that significantly impairs daily functioning and overall quality of life. Traditional diagnostic methods primarily rely on self-reported questionnaires, which are susceptible to subjective biases. In recent years, facial video-based automatic depression analysis has emerged as a promising alternative, offering enhanced objectivity and greater scalability compared to traditional methods. However, existing approaches typically focus on either facial action units (AUs) or facial image sequences, limiting their ability to comprehensively capture depressive expressions by focusing only on local or global facial information. To address these limitations, we propose a novel spatiotemporal spectral feature fusion network (STSFF-Net). STSFF-Net integrates a one-dimensional residual network for transforming AU time-series data into spectral features, a three-dimensional residual network enhanced with a convolutional block attention module to extract spatiotemporal facial representations, and a late fusion mechanism that effectively combines local and global features. We validated the STSFF-Net model on the Chosun University Hospital Depression Risk Video dataset, which comprises video recordings from 187 clinically diagnosed participants. Our model demonstrated superior performance over existing methods, achieving a mean absolute error of 3.18 and a root mean square error of 4.78. This underscores its effectiveness and potential for real-world depression screening applications.}, archive = {J_NEUCOM}, author = {Daegil Choi and Jisun Hong and Jihun Lee and Zhiqiang Liu and Gengjia Zhang and Jaehyo Jung}, doi = {10.1016/j.neucom.2025.131126}, journal = {Neurocomputing}, month = {11}, pages = {131126}, shortjournal = {Neurocomputing}, title = {STSFF-net: A two-stream network for depression detection from facial expressions}, volume = {652}, year = {2025}, }
• (2025). Triple-residual enhanced convolution- and transformer-based hybrid encoder-decoder network for medical image segmentation. NEUCOM, 652, 131125. (www)

  @article{ , abstract = {In contemporary medical practice, computed tomography and ultrasound imaging play a critical role in disease diagnosis, treatment planning, and monitoring. Image segmentation is a foundational step in medical image analysis, which is essential for extracting key anatomical and pathological information and has been dominated by deep learning. Convolutional Neural Network (CNN) and Vision Transformer (ViT) are two foundational architectures, and both demonstrate promising performance in segmentation. In this paper, Triple-Residual Enhanced Convolution- and Transformer-Based Hybrid Encoder-Decoder Network, named TriResUNet, is proposed. Specifically, a redesigned residual multi-resolution CNN-based block is introduced to enhance multi-level feature extraction ability for Encoders and Decoders, respectively. Secondly, Shift-Window-based ViT with residual connection is explored in the bottleneck aiming to capture long-range relationships. Thirdly, residual connection with computationally efficient CNN-based block is integrated between Encoders and Decoders to mitigate the vanishing gradient problem, facilitating efficient gradient flow and enhancing feature reuse across layers. The proposed network is comprehensively validated on various publicly available datasets against comprehensive baseline methods and various types of similar hybrid networks. The experimental findings suggest the proposed network demonstrates superior performance. The code will be available upon acceptance at https://github.com/ziyangwang007/VIT4UNet .}, archive = {J_NEUCOM}, author = {Minghan Lin and Ziyang Wang}, doi = {10.1016/j.neucom.2025.131125}, journal = {Neurocomputing}, month = {11}, pages = {131125}, shortjournal = {Neurocomputing}, title = {Triple-residual enhanced convolution- and transformer-based hybrid encoder-decoder network for medical image segmentation}, volume = {652}, year = {2025}, }
• (2025). Natural language to SQL conversion using probabilistic spiking return-aligned decision transformer neural networks and subtraction-average-based optimizer. NEUCOM, 652, 131123. (www)

  @article{ , abstract = {The artificial intelligence task of Natural Language to SQL (NL2SQL) conversion empowers non-technical users to query databases using plain language, enhancing data accessibility across sectors such as business intelligence, healthcare, and customer service. However, traditional systems struggle with ambiguity, complex query structures, and schema adaptability, limiting their effectiveness. This study introduces a novel NL2SQL framework that combines advanced preprocessing techniques, a hybrid feature selection approach using Cat Hunting Optimization and War Strategy Optimization, and a Probabilistic Spiking Return-Aligned Decision Transformer Neural Network (PSRANN) for precise SQL generation. Additionally, a Subtraction-Average-Based Optimizer (SABO) is employed to enhance model training efficiency. The proposed framework achieves a remarkable 99.9 % accuracy on the Spider dataset while demonstrating superior performance in precision, recall, F1-score, and query efficiency compared to existing models. The system shows strong adaptability to diverse query structures and database schemas. This work’s novelty lies in its biologically inspired neural architecture and dual-optimization strategy, offering a highly accurate, scalable, and robust solution for NL2SQL tasks. Future work will focus on multilingual support, real-time applications, and deployment in domain-specific environments.}, archive = {J_NEUCOM}, author = {Ashlesha Kolarkar and Kailasam Swathi}, doi = {10.1016/j.neucom.2025.131123}, journal = {Neurocomputing}, month = {11}, pages = {131123}, shortjournal = {Neurocomputing}, title = {Natural language to SQL conversion using probabilistic spiking return-aligned decision transformer neural networks and subtraction-average-based optimizer}, volume = {652}, year = {2025}, }
• (2025). Artificial inductive bias for synthetic tabular data generation in data-scarce scenarios. NEUCOM, 652, 131122. (www)

  @article{ , abstract = {While synthetic tabular data generation using Deep Generative Models (DGMs) offers a compelling solution to data scarcity and privacy concerns, their effectiveness relies on the availability of substantial training data, often lacking in real-world scenarios. To overcome this limitation, we propose a novel methodology that explicitly integrates artificial inductive biases into the generative process to improve data quality in low-data regimes. Our framework leverages transfer learning and meta-learning techniques to construct and inject informative inductive biases into DGMs. We evaluate four approaches (pre-training, model averaging, Model-Agnostic Meta-Learning (MAML), and Domain Randomized Search (DRS)) and analyze their impact on the quality of the generated text. Experimental results show that incorporating inductive bias substantially improves performance, with transfer learning methods outperforming meta-learning, achieving up to 60 % gains in Jensen-Shannon divergence. The methodology is model-agnostic and especially relevant in domains such as healthcare and finance, where high-quality synthetic data are essential, and data availability is often limited.}, archive = {J_NEUCOM}, author = {Patricia A. Apellániz and Ana Jiménez and Borja Arroyo Galende and Juan Parras and Santiago Zazo}, doi = {10.1016/j.neucom.2025.131122}, journal = {Neurocomputing}, month = {11}, pages = {131122}, shortjournal = {Neurocomputing}, title = {Artificial inductive bias for synthetic tabular data generation in data-scarce scenarios}, volume = {652}, year = {2025}, }
• (2025). AcuGPT-agent: An LLM-powered intelligent system for acupuncture-based infertility treatment. NEUCOM, 652, 131116. (www)

  @article{ , abstract = {This paper introduces AcuGPT-Agent, a novel intelligent system powered by a domain-specific large language model (LLM) designed for acupuncture-based infertility treatment. Traditional Chinese Medicine (TCM), particularly acupuncture, presents unique challenges for general-purpose LLM due to its complex theoretical framework and specialized terminology. To bridge this gap, we make three key contributions. First, we develop AcuGPT, the first LLM specifically fine-tuned for the acupuncture domain, utilizing advanced techniques such as LoRA, DoRA, and a Pretrain + Supervised Fine-Tuning (SFT) pipeline. These methods significantly improve the ability of the model to understand and generate medical content related to TCM. Second, we propose AcuGPT-Agent, a modular and extensible system that integrates AcuGPT with standard agent tools, a domain-specific acupuncture knowledge graph for infertility treatment, and a novel LLM-driven Multi-Department Knowledge Base Routing Management Mechanism (MDKBRMM). This mechanism enhances retrieval-augmented generation by effectively managing semantic complexity and routing queries across specialized knowledge bases in medical contexts. Third, we build EvalAcu, a dedicated benchmark dataset for evaluating large language models and intelligent agents within the acupuncture domain. Experimental results demonstrate that AcuGPT exhibits strong domain knowledge and reasoning capabilities, while AcuGPT-Agent further enhances performance in real-world infertility treatment tasks. All code, datasets, and resources are publicly available at: https://github.com/suda7788/AcuGPT-Agent .}, archive = {J_NEUCOM}, author = {Jing Wen and Diandong Liu and Yuxin Xie and Yi Ren and Jiacun Wang and Youbing Xia and Peng Zhu}, doi = {10.1016/j.neucom.2025.131116}, journal = {Neurocomputing}, month = {11}, pages = {131116}, shortjournal = {Neurocomputing}, title = {AcuGPT-agent: An LLM-powered intelligent system for acupuncture-based infertility treatment}, volume = {652}, year = {2025}, }
• (2025). MFEGPT: A PID controller-inspired multimodal feature enhancement for MLLMs. NEUCOM, 652, 131114. (www)

  @article{ , abstract = {To overcome the persistent limitations in spatial detail comprehension and perceptual capacities observed in MLLMs, an innovative multilayer feature enhancement model named MFEGPT is proposed in this paper. Inspired by the dynamic regulation principle of PID controllers, MFEGPT incorporates a visual encoder with collaborative Proportion (P), Integration (I), and Differentiation (D) branches. The proposed architecture consists of three specialized branches: (1) the P branch extracts high-resolution spatial details, (2) the I branch integrates contextual information to mitigate global semantic bias, and (3) the D branch regulates spatial overshooting through contour feature extraction. These components are dynamically integrated through a PID-inspired closed-loop feedback mechanism, significantly improving the joint modeling of both spatial details and semantic consistency. For the projector, we employ a Q-former with linear layers to effectively preserve spatial details. An edge-aware feature mapping module is specifically designed to enhance fine-grained perception. The entire model is trained on a single RTX3090 GPU and requires only 8.9 GB VRAM during inference, making it deployable on standard 10 GB GPU devices. Extensive benchmarking shows that MFEGPT outperforms other lightweight MLLMs while maintaining only 2.8B parameters. Remarkably, it achieves performance competitive with 13B-parameter models, demonstrating an optimal balance between capability and resource efficiency.}, archive = {J_NEUCOM}, author = {Hexiao Li and Zixi Jia and Xiangyue Zhang and Chuanjiang Leng and Heyao Li and Hongbin Gao and Jiqiang Liu and Qinghua Liu}, doi = {10.1016/j.neucom.2025.131114}, journal = {Neurocomputing}, month = {11}, pages = {131114}, shortjournal = {Neurocomputing}, title = {MFEGPT: A PID controller-inspired multimodal feature enhancement for MLLMs}, volume = {652}, year = {2025}, }
• (2025). PONet: Prototype optimization network for few-shot medical image segmentation. NEUCOM, 652, 131113. (www)

  @article{ , abstract = {Although deep learning algorithms have achieved remarkable success in medical image segmentation, their reliance on extensive manually annotated datasets restricts the generalization to scenarios with limited training data. Few-shot segmentation has emerged as a promising solution to this problem by enabling effective segmentation with limited data. However, when applied directly to medical images, there are often two primary challenges: inter-class and intra-class inconsistency. First, inter-class inconsistency refers to the high similarity between different pathological states or tissue structures in pathological images, which negatively impacts the quality of support prototypes. Second, intra-class inconsistency stems from the varying appearances of the same organ across different samples, thereby disrupting the support–query alignment process. To counter these challenges, we propose a prototype optimization network to achieve accurate few-shot medical image segmentation. Specifically, we first employ the simple linear iterative clustering (SLIC) method to generate multiple foreground and background sub-regions within the support image. Subsequently, we introduce a boundary prototype contrastive learning (BPCL) module that utilizes contrastive learning to self-optimize the support prototypes, enhancing the quality of the support prototypes. This reduces the risk of misclassifying the query background as foreground during the support–query alignment process, thereby mitigating inter-class inconsistency. Furthermore, we propose the query-guided prototype optimization (QGPO) module, which introduces a query-guided optimization strategy to generate support prototypes tailored to the current query image. This enhances the connection between support features and the query image, thereby alleviating intra-class inconsistency. Comprehensive experiments are conducted on three medical datasets. Our method achieves state-of-the-art segmentation performance compared to other medical FSS methods. Code is available at: https://github.com/WANGSIQII/FSS.git .}, archive = {J_NEUCOM}, author = {Wang Siqi and Yu Xiaosheng and Chi Jianning and Wu Chengdong and Gao Xiujing}, doi = {10.1016/j.neucom.2025.131113}, journal = {Neurocomputing}, month = {11}, pages = {131113}, shortjournal = {Neurocomputing}, title = {PONet: Prototype optimization network for few-shot medical image segmentation}, volume = {652}, year = {2025}, }
• (2025). Finite-time bipartite synchronization control of coupled inertial neural networks over sign graph. NEUCOM, 652, 131111. (www)

  @article{ , abstract = {In the present paper, finite-time bipartite synchronization problem is investigated for coupled inertial neural networks (INNs) over a sign graph. By employing a nonreduced-order framework, a coupling structure-dependent Lyapunov function is used to explore the finite-time stability of the synchronization error dynamics. Through the design of a discontinuous control strategy incorporating sign functions, an explicit synchronization criteria and a settling time estimation are derived. To enhance practical applicability, adaptive update laws are developed to achieve adaptive synchronization controllers that reduce control energy consumption. The theoretical results are verified through a comprehensive numerical example demonstrating the validity of the proposed control schemes.}, archive = {J_NEUCOM}, author = {Tianhu Yu and Dengqing Cao and Jinde Cao and Liang Hua}, doi = {10.1016/j.neucom.2025.131111}, journal = {Neurocomputing}, month = {11}, pages = {131111}, shortjournal = {Neurocomputing}, title = {Finite-time bipartite synchronization control of coupled inertial neural networks over sign graph}, volume = {652}, year = {2025}, }
• (2025). FeaUn: Feature unlearning in vertical federated learning for IIoT against feature inference attacks. NEUCOM, 652, 131110. (www)

  @article{ , abstract = {Vertical Federated Learning (VFL) has been increasingly popular in Industrial Internet of Things (IIoT) systems due to the independence of feature alignment and privacy preservation capabilities. Although the VFL for IIoT systems offers higher privacy guarantees, there are still security challenges (e.g., feature inference attacks), resulting in the leakage of sensitive features. This challenge could be addressed by retraining a model on a dataset without the sensitive features from scratch or removing all samples from the model using federated unlearning. However, these approaches either introduce high overheads due to retraining or reduce the accuracy of the unlearned model. In this paper, we propose Feature Unlearning (FeaUn) in VFL for IIoT systems, which trades off between the overheads and accuracy. FeaUn differs from a traditional VFL framework only by adding an unlearning round in which the clients and server evaluate and remove the effects of sensitive features. Specifically, a perturbed dataset is constructed using a feature perturbation strategy , where the sensitive features are perturbed with noise. Then, the effects are approximated as the parameter difference between models trained on the original and perturbed dataset. Finally, the unlearned model is obtained through a first-order update, where the effect is removed from the model in one epoch. Experiments on three standard datasets demonstrate the efficacy and fidelity of FeaUn, with an obvious speed-up of 25 × compared with baselines.}, archive = {J_NEUCOM}, author = {Zhaobo Lu and Xu Zhang and Tao Li and Yilei Wang and Guangshun Li and Zhiquan Liu}, doi = {10.1016/j.neucom.2025.131110}, journal = {Neurocomputing}, month = {11}, pages = {131110}, shortjournal = {Neurocomputing}, title = {FeaUn: Feature unlearning in vertical federated learning for IIoT against feature inference attacks}, volume = {652}, year = {2025}, }
• (2025). Koopman framework with self-supervised spectral alignment for multi-domain time-series modeling and prediction. NEUCOM, 652, 131109. (www)

  @article{ , abstract = {In multi-domain time-series modeling, differences in underlying system dynamics across domains present significant challenges for effective knowledge transfer. Traditional domain adaptation methods often focus on aligning feature distributions, while overlooking the temporal evolution laws that govern the data. In this paper, we propose a shared-embedding Koopman framework for learning aligned dynamical models across multiple domains. Our method embeds observations from each domain into a common latent space, where domain-specific linear Koopman operators capture the temporal dynamics. Specifically, nonlinear systems are transformed into linear Koopman representations through learned neural network embeddings, enabling tractable analysis of complex dynamics. To achieve cross-domain consistency, we introduce a dynamics alignment mechanism that enforces similarity between the latent transition operators via operator alignment losses and spectral regularization. We employ Procrustes analysis for eigenvector alignment, which minimizes the Frobenius norm between aligned spectral bases, ensuring that domains share similar fundamental dynamical modes. Furthermore, we incorporate a self-supervised consistency loss to enhance the robustness of latent representations against sequence augmentations. The augmentation strategies include temporal jittering, scaling, and masking, which preserve the essential dynamical properties while introducing controlled variations. This joint modeling approach allows the framework to achieve both intra-domain accurate prediction and inter-domain transferable forecasting. We evaluate our method on three physiological time-series datasets: PAMAP2 Physical Activity Monitoring, UCI Human Activity Recognition (HAR), and EEG Motor Movement/Imagery datasets from multiple subjects, demonstrating superior transfer prediction performance compared to conventional non-aligned models and state-of-the-art domain adaptation baselines including transformer-based methods. Our approach achieves an average improvement of 49.4 % in Mean Absolute Error across all transfer tasks, with particularly strong performance (81.1 % improvement) on low-dimensional physiological signals. The results confirm that our approach effectively aligns latent dynamics across domains and provides a principled solution for time-series domain adaptation tasks. We discuss limitations regarding scalability to very high-dimensional systems and propose future directions for adaptive alignment strategies.}, archive = {J_NEUCOM}, author = {Bailing Zhang}, doi = {10.1016/j.neucom.2025.131109}, journal = {Neurocomputing}, month = {11}, pages = {131109}, shortjournal = {Neurocomputing}, title = {Koopman framework with self-supervised spectral alignment for multi-domain time-series modeling and prediction}, volume = {652}, year = {2025}, }
• (2025). AFMIP-net: Adaptive feature modulation and implicit prompt network for retinal vessel segmentation. NEUCOM, 652, 131106. (www)

  @article{ , abstract = {Accurate segmentation of retinal vessel images is critical for diagnosing ophthalmic diseases. However, retinal vessel segmentation is challenging due to the scarcity of training samples, retinal vessel morphology polymorphism, and the low contrast with the background. Traditional unsupervised methods are sensitive to parameter selection and image preprocessing, making it difficult to adapt to the complex and variable morphology of blood vessels. Although existing deep learning models (such as U-Net and its variants) can extract multi-scale features, they fail to effectively calibrate the positional information of blood vessels, resulting in insufficient suppression of background noise. Most methods suffer from the loss of key details during feature fusion, leading to poor segmentation results for capillaries. Moreover, these models pay insufficient attention to blood vessel segmentation under the interference of lesions, limiting their generalization ability. A novel Adaptive Feature Modulation and Implicit Prompt Network (AFMIP-Net) based on the encoder-decoder structure is proposed to address these issues. To improve the segmentation accuracy with complex background and different scales of blood vessels, the modulating feature maps from the high-level semantic layer to the low-level semantic layer in the encoder are considered. To improve the capillary segmentation, the multi-scale implicit prompt is extracted and applied to each decoder layer by suppressing redundant features with the fusion of the encoding and decoding information. Moreover, data enhancement with the Gaussian filter is adopted to solve the issue of training sample scarcity, which improves the focus on difficult-to-segment complex vessels with the interference of lesions and further enhances generalization. These innovations have provided new technical ideas for the field of retinal vascular segmentation and are of great significance for promoting the development of computer-aided diagnosis in ophthalmology. Experimental results on the mainstream DRIVE, CHASE, STARE, HRF, IOSTAR, and LES-AV datasets show that the F1 score of our model is optimal with all six datasets. Also, its sensitivity and accuracy performance stand out across most datasets. The visual comparison further validates the superiority of our method in identifying retinal vessel topological shapes and micro-vessels. It is concluded that the overall performance of AFMIP-Net is significantly improved in comparison to other state-of-the-art methods. It provides important support for the detection and diagnosis of early eye diseases.}, archive = {J_NEUCOM}, author = {Ying Xie and Jingkai Shang and Ruixiang Deng and Bohao Cao and Shuaiyong Li and Xianlun Tang}, doi = {10.1016/j.neucom.2025.131106}, journal = {Neurocomputing}, month = {11}, pages = {131106}, shortjournal = {Neurocomputing}, title = {AFMIP-net: Adaptive feature modulation and implicit prompt network for retinal vessel segmentation}, volume = {652}, year = {2025}, }
• (2025). A multi relational graph-based approach for stock movement prediction and portfolio optimization. NEUCOM, 652, 131104. (www)

  @article{ , abstract = {The objective of this work is to predict stock price movements and build a profitable portfolio of stocks listed in the NIFTY-50 index of the Indian economy. An accurate prediction of future stock price movements is a complex task due to the nonlinear and volatile nature of stock data. Recent approaches have neglected interdependence between stock pairs or relied on coarse-grained relationships to extract hidden interactions between stocks. This work proposes an end-to-end deep learning framework, Multiple Relation aware Relational Temporal Network (MR2TNet), that considers multiple sources of information to learn the latent interdependence between intra- and inter-sector stock pairs. The financial knowledge graphs are built from three different types of indicators, i.e., sector, technical, and fundamental. The fundamental ratios that provide valuable insights into a company’s financial health have not been extensively utilized in recent approaches. In this work, we incorporate various fundamental ratios along with sector and technical information to provide a more comprehensive understanding of stock relationships. A multi-graph modeling and temporal modeling module is proposed to extract relational and temporal patterns from built knowledge graphs and market data, respectively. We have also proposed a novel loss function that jointly considers prediction and directional movements to efficiently train a model. To demonstrate the application of MR2TNet, a novel EWR-CVaR portfolio optimization approach is proposed to build a profitable portfolio from the predicted results. The phase-wise experiments conducted on the NIFTY-50 dataset show that MR2TNet achieved higher accuracy across the majority of phases, proving its efficiency. Moreover, it has achieved a 16 % higher annual return compared to the other existing baseline models. The results emphasize the superiority of the MR2TNet model in both prediction and portfolio building, proving its efficiency and applicability.}, archive = {J_NEUCOM}, author = {Manali Patel and Shreya Goyal and Krupa Jariwala and Chiranjoy Chattopadhyay}, doi = {10.1016/j.neucom.2025.131104}, journal = {Neurocomputing}, month = {11}, pages = {131104}, shortjournal = {Neurocomputing}, title = {A multi relational graph-based approach for stock movement prediction and portfolio optimization}, volume = {652}, year = {2025}, }
• (2025). Distributed optimal formation control for constrained quadrotor unmanned aerial vehicles via stackelberg-game. NEUCOM, 652, 131102. (www)

  @article{ , abstract = {This paper investigates the distributed optimal formation control problem for quadrotor unmanned aerial vehicles with full-state constraints. First, a time-varying unified universal barrier function is designed to handle the time-varying state constraints. Second, the Stackelberg-game framework is introduced to model the sequential decision-making process between leader and followers, in which the leader first updates its strategy, and the followers subsequently adjust their responses accordingly. Third, the Stackelberg-game problem can be solved by addressing the coupled Hamilton-Jacobi-Bellman equation, and an on-policy reinforcement learning algorithm is designed to obtain the solution. The stability of the system is proved using Lyapunov stability theory. Numerical simulations demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Youneng Li and Guopeng Zhou and Peng Jin and Yutao Zhu and Renzhuo Wan and Ying Han}, doi = {10.1016/j.neucom.2025.131102}, journal = {Neurocomputing}, month = {11}, pages = {131102}, shortjournal = {Neurocomputing}, title = {Distributed optimal formation control for constrained quadrotor unmanned aerial vehicles via stackelberg-game}, volume = {652}, year = {2025}, }
• (2025). Dual-stage learning framework for underwater acoustic target recognition with cross-attention mechanism and audio-guided contrastive learning. NEUCOM, 652, 131101. (www)

  @article{ , abstract = {Underwater acoustic target recognition is crucial for marine exploration and environmental monitoring. However, the redundancy in raw time-domain signals and the challenges in effectively integrating time-frequency representations with audio features hinder current methods. To address these challenges, we introduce a dual-stage learning framework, named audio-guided cross-attention dual-stage network for underwater acoustic target recognition (ACDN-UATR). The approach combines a cross-attention mechanism with audio-guided contrastive learning to improve recognition accuracy. In the first stage, a masked autoencoder (MAE) is used to learn and reconstruct time-frequency features, while a cross-attention mechanism efficiently fuses features from different spectrograms. In the second stage, contrastive learning is employed to align features extracted from time-frequency representations and raw audio signals, enhancing feature consistency and recognition robustness. Experimental results demonstrate that ACDN-UATR effectively integrates both time-frequency and audio-based features, achieving high recognition accuracy on the ShipsEar dataset.}, archive = {J_NEUCOM}, author = {Rongyao Zhao and Feng Liu and Lyufang Zhao and Daihui Li and Jing Xu and Yuanxin Liu and Tongsheng Shen}, doi = {10.1016/j.neucom.2025.131101}, journal = {Neurocomputing}, month = {11}, pages = {131101}, shortjournal = {Neurocomputing}, title = {Dual-stage learning framework for underwater acoustic target recognition with cross-attention mechanism and audio-guided contrastive learning}, volume = {652}, year = {2025}, }
• (2025). RMAF: A replay method based on active forgetting for continual learning. NEUCOM, 652, 131098. (www)

  @article{ , abstract = {The goal of continual learning (CL) is to enable deep neural networks to maintain good performance on old tasks while continuously learning new tasks. The main challenge in the field of CL is catastrophic forgetting (CF). This occurs because of the shared nature of model parameters, where learning a new task can overwrite or cause the loss of knowledge from previous tasks due to parameter updates. To alleviate this problem, the replay method is widely regarded as a simple and effective strategy that restores the performance of the model on the previous task by using a small amount of old task data. However, existing replay methods usually mix replay samples with new task data for training. This strategy makes it difficult to effectively adjust the shared parameters that cause CF, thereby weakening the effect of alleviating forgetting. To address the above problems, this paper proposes a replay method based on active forgetting (RMAF) under continual domain-adaptive pre-training (DAP-training). Unlike traditional methods, RMAF does not mix replay data with current domain data for training, but uses the replay data employed to alleviate CF to train the shared parameters that cause forgetting separately. In addition, to avoid excessive bias of shared parameters towards the current domain, RMAF introduces an active forgetting mechanism to ensure that each domain achieves balanced learning during parameter adjustment. Experimental results show that RMAF outperforms existing methods and achieves the best performance on multiple benchmark datasets.}, archive = {J_NEUCOM}, author = {Huaxiao Qiu and Jianzhou Feng and Lazhi Zhao and Chenghan Gu and Haoran Yu and Yuxuan Zhang and Ziqi Wang}, doi = {10.1016/j.neucom.2025.131098}, journal = {Neurocomputing}, month = {11}, pages = {131098}, shortjournal = {Neurocomputing}, title = {RMAF: A replay method based on active forgetting for continual learning}, volume = {652}, year = {2025}, }
• (2025). Semi-GDE: Generative distribution estimation for semi-supervised medical landmark localization. NEUCOM, 652, 131095. (www)

  @article{ , abstract = {Medical landmark localization is crucial for clinical diagnosis and surgical planning. Semi-supervised medical landmark localization methods alleviate the reliance on large annotated datasets, facilitating clinical adoption. In prior semi-supervised methods, landmarks are commonly assumed to follow a Laplace or a Gaussian distribution, which may not match the true probability distribution well. To address this limitation, we propose a novel semi-supervised landmark localization method with generative distribution estimation, referred to as Semi-GDE. Our method leverages a flow-based generative model to learn the landmark distribution from manual annotations and pseudo-labels that are generated by trained models, thereby improving localization accuracy. Moreover, Semi-GDE adopts a triple-model architecture driven by a pseudo-label refinement module (PRM) to improve the reliability of the estimated pseudo-labels during the training stage. In particular, two models alternate between teacher and student roles, generating pseudo-labels through weak-strong consistency learning. PRM then refines these pseudo-labels based on position inconsistency. Subsequently, the refined pseudo-labels are used to train a third model, which is the model we use to generate output at the inference stage. We additionally introduce residual log-likelihood estimation (RLE) loss, which models the landmark distribution for position regression, and skeleton direction (SDir) loss, which captures the landmark dependency by explicitly constraining the anatomical structure of landmarks. By refining noisy pseudo-labels and incorporating prior information such as landmark distribution and landmark dependency, Semi-GDE enhances localization performance with limited labeled resources. Results of comprehensive experiments conducted on three typical yet challenging datasets of different anatomical structures demonstrated the superior performance of our method over other state-of-the-art methods. Overall, Semi-GDE provides a practical and effective solution for semi-supervised medical landmark localization, reducing the workload of doctors.}, archive = {J_NEUCOM}, author = {Kangqing Ye and Xiaoyang Zou and Wenyuan Sun and Guoyan Zheng}, doi = {10.1016/j.neucom.2025.131095}, journal = {Neurocomputing}, month = {11}, pages = {131095}, shortjournal = {Neurocomputing}, title = {Semi-GDE: Generative distribution estimation for semi-supervised medical landmark localization}, volume = {652}, year = {2025}, }
• (2025). Advancing few-shot rolling element bearing diagnostics with fine-grained similarity learning across variable working environments. NEUCOM, 652, 131093. (www)

  @article{ , abstract = {Bearing fault diagnosis models trained on single-source data often exhibit limited generalization capability due to significant cross-domain distribution shifts and data scarcity under varying operating conditions. To address this challenge, we propose the Fine-Grained Similarity Network (FGSN), a novel few-shot learning framework for rolling element bearing diagnostics. The FGSN architecture first converts vibration signals into 2D tensors, then employs a multi-branch feature representation module with enhanced Squeeze-and-Excitation blocks to generate discriminative feature maps. A hierarchical similarity learning module subsequently computes multi-scale similarity scores, enhancing sensitivity to subtle fault characteristics. Using the episode-based training paradigm (5-way 5-shot), we evaluate the framework on three benchmark datasets: CWRU (10 classes) for base training/testing, PU (13 classes) and SEU (5 classes) for cross-domain validation. Experimental results demonstrate the FGSN’s superior performance, achieving F1-scores of 0.9950 (CWRU), 0.9635 (PU), and 0.9167 (SEU) - representing performance gains of 4.33 % to 11.35 % compared to existing few-shot methods under challenging low-SNR conditions (−4 dB). The results show that the proposed method can generalize fault knowledge from a single-source database to new, unseen cross-domain classes in variable environments.}, archive = {J_NEUCOM}, author = {Tianlei Gao and Xiaoyun Xie and Nuo Wei and Minglei Shu and Yushui Geng}, doi = {10.1016/j.neucom.2025.131093}, journal = {Neurocomputing}, month = {11}, pages = {131093}, shortjournal = {Neurocomputing}, title = {Advancing few-shot rolling element bearing diagnostics with fine-grained similarity learning across variable working environments}, volume = {652}, year = {2025}, }
• (2025). Bearing fault diagnosis under heavy noise: A multi-scale dilated convolution and dense temporal convolutional network. NEUCOM, 652, 131092. (www)

  @article{ , abstract = {In recent years, deep learning techniques have made significant advancements in intelligent fault diagnosis. However, industrial noise contamination frequently obscures critical fault signatures, compromising diagnostic reliability and hindering the practical deployment of data-driven solutions. To address these challenges, we propose the Multi-Scale Dilated Convolution and Dense Temporal Convolutional Network (MSDC-DenseTCN) model. The architecture design is guided by experimental analyses of block connectivity and depth configuration, ensuring optimal feature representation in noisy environments. The model first employs the MSDC module, which consists of three dilated convolution branches with varying dilation rates to filter noise and extract fault features at multiple scales effectively. Subsequently, the DenseTCN module integrates the strengths of DenseNet and Temporal Convolutional Networks (TCN) to improve feature propagation and maintain an expansive receptive field, effectively mitigating the impact of noise on diagnostic performance. The proposed model’s performance is validated using bearing datasets from Paderborn University (PU) and Case Western Reserve University (CWRU) under varying noise conditions. The model achieves the highest accuracy and F1 score across diverse operating conditions, outperforming five state-of-the-art fault diagnosis methods.}, archive = {J_NEUCOM}, author = {Xiangfeng Si and Hao Yan and Yaming Hu and Jian Duan and Tielin Shi}, doi = {10.1016/j.neucom.2025.131092}, journal = {Neurocomputing}, month = {11}, pages = {131092}, shortjournal = {Neurocomputing}, title = {Bearing fault diagnosis under heavy noise: A multi-scale dilated convolution and dense temporal convolutional network}, volume = {652}, year = {2025}, }
• (2025). Wavelet-driven multi-frequency signal unlocking network for image deraining. NEUCOM, 652, 131089. (www)

  @article{ , abstract = {Rain streaks and mist caused by rainfall can significantly degrade image quality, directly hindering downstream tasks. In recent years, wavelet transform-based rain removal methods have advanced due to their ability to capture frequency characteristics, thereby enhancing image structure preservation while suppressing high-frequency noise. However, existing approaches often overlook the importance of independently processing rain components, disrupting the intrinsic relationship between rain and the background. To address this limitation, we propose a novel framework called the wavelet-driven multi-frequency signal unlocking (WD-MFSU) network. This network employs a wavelet decomposition-driven multi-element learning (WDMEL) mechanism that integrates frequency signals into an attention mechanism and differential learning through a collaborative processing mode, which includes global, structural, and detail-oriented branches. This approach alleviates artifacts and background distortion during rain removal. Additionally, we introduce the rain-background augmented dynamic interaction (RBADC) module to enhance cross-component feature interaction and capture potential dependencies between rain and the background. By utilizing sequential compound attention and dynamic convolution, this module recalibrates channels for rain and background, facilitating dynamic feature adjustments. Comprehensive evaluations demonstrate that our method excels across multiple rainy datasets, outperforming recently proposed approaches while maintaining lower complexity, thus showcasing its effectiveness and generalizability in real-world rainy scenes and downstream tasks. The source code will be available at https://github.com/ChenHuyoo .}, archive = {J_NEUCOM}, author = {Jiping Liu and Hongyu Chen and Shaohan Cao and Yong Wang and Jun Zhu and Dejun Feng and Yakun Xie}, doi = {10.1016/j.neucom.2025.131089}, journal = {Neurocomputing}, month = {11}, pages = {131089}, shortjournal = {Neurocomputing}, title = {Wavelet-driven multi-frequency signal unlocking network for image deraining}, volume = {652}, year = {2025}, }
• (2025). Efficient and transferable reversible adversarial attacks utilizing YUV color space. NEUCOM, 652, 131088. (www)

  @article{ , abstract = {Adversarial attacks, which involve adding subtle perturbations to images, pose a significant threat to the secure deployment of deep neural networks. However, when integrated with reversible data hiding (RDH) technology, generated adversarial examples (AEs) can both prevent malicious identification and enable error-free recovery of the original image. This technique is known as error-free reversible adversarial attack. Despite its potential, existing error-free reversible adversarial attack methods primarily focus on feasibility, attack success rate, and image quality, neglecting cross-model transferability and ineffective perturbations, such as embedding-overwritten and generation-redundant perturbations. These issues result in relatively slow operational speeds and limit their applicability to unknown models. To address these challenges, a novel error-free reversible adversarial attack method based on the YUV color space is proposed. By separating the luminance and chrominance channels, this space allows for more efficient image processing. Our method adopts a dual-strategy design: Y-channel attacks (e.g., YFGSM, YI-FGSM, YPGD) are used to eliminate generation-redundant perturbations, while the embedding of perturbation information into the UV channels avoids overwriting, thereby enhancing both transferability and computational efficiency. Furthermore, an ensemble-based attack strategy is introduced to further improve cross-model performance. Experimental results demonstrate that our method not only enables error-free recovery of the original image but also maintains high visual quality, achieves high operational speed, and exhibits strong transferability across multiple models.}, archive = {J_NEUCOM}, author = {Yucheng Fan and Zhaoxia Yin and Jiawei Chen and Wanli Lyu}, doi = {10.1016/j.neucom.2025.131088}, journal = {Neurocomputing}, month = {11}, pages = {131088}, shortjournal = {Neurocomputing}, title = {Efficient and transferable reversible adversarial attacks utilizing YUV color space}, volume = {652}, year = {2025}, }
• (2025). Text-adaptive spatio-temporal interaction network for video question answering. NEUCOM, 652, 131087. (www)

  @article{ , abstract = {Video question answering is an intricate task where the intent of a question involves spatial or temporal visual information. In recent works, cross-modal spatio-temporal interaction is either fixed by a designed framework or achieved dynamically through additional annotations. These methods cannot adaptively determine the degree of interaction in space or time. To address this, we propose the Text-adaptive Spatio-temporal Interaction Network (TSIN), which adaptively interacts based on the spatio-temporal intention of the question. Specifically, video and text initially perform intra-modal self-attention perception to enhance feature representation, and then undergo cross-modal spatial and temporal interaction. The text-adaptive fusion module adaptively integrates the interacted multimodal features based on the question. Finally, the answer is generated by the fusion feature. We evaluate our method on the MSVD-QA, MSRVTT-QA, and TGIF-FrameQA datasets, and the results show competitive performance. Additional experiments demonstrate the effectiveness of each module.}, archive = {J_NEUCOM}, author = {Linlin Zong and Jiahui Wan and Xinyu Zhai and Qianli Zhao and Xinyue Liu and Xianchao Zhang and Bo Xu}, doi = {10.1016/j.neucom.2025.131087}, journal = {Neurocomputing}, month = {11}, pages = {131087}, shortjournal = {Neurocomputing}, title = {Text-adaptive spatio-temporal interaction network for video question answering}, volume = {652}, year = {2025}, }
• (2025). Semantic substructure guided multiple objective molecular generation with discrete diffusion probabilistic model. NEUCOM, 652, 131086. (www)

  @article{ , abstract = {Deep learning methods have paved the highway for drug discovery in past few years, it is still a big challenge that designed molecules satisfy multiple objective constraints. Traditional generative adversarial networks suffer severe mode collapse problems and Transformer based large language models require relative large amount of data to train or fine-tune. Denoising diffusion probabilistic models have achieved huge success in image generation fields, yet it remains an open challenge to explore diffusion based language models in drug discovery. Herein, we bridge the molecule prior distribution with the desired bioactive molecular distribution and propose SGDiff , S ubstructure G uided Diff usion model, which can iteratively inject the sequence-based conditional information into the denoising process to guide the generation process without additional reinforcement learning fine-tuning process. We conduct sufficient discussion about the Transformer based language models and diffusion based language models. SGDiff can achieve the highest conditional success rate in multiple objective generation tasks and can maintain high diversity from conditional guidance. Besides, SGDiff has a great interpretation during sampling. Several explorations have been conducted to detail understanding the mechanism of the guidance paradigm, which can be better used for further multiple objective constraints molecular generation.}, archive = {J_NEUCOM}, author = {Shugao Chen and Bosheng Song and Ruizhe Chen and Lin Liu and Guifei Zhou and Sisi Yuan}, doi = {10.1016/j.neucom.2025.131086}, journal = {Neurocomputing}, month = {11}, pages = {131086}, shortjournal = {Neurocomputing}, title = {Semantic substructure guided multiple objective molecular generation with discrete diffusion probabilistic model}, volume = {652}, year = {2025}, }
• (2025). Unsupervised bidirectional generative smoothing framework with frequency decomposition and attention enhancement. NEUCOM, 652, 131085. (www)

  @article{ , abstract = {Edge-preserving smoothing is a fundamental computer vision task. Filter-based and optimization-based methods require iterative parameter adjustments to accommodate new samples. Learning-based methods offer greater generalizability; however, they predominantly employ supervised learning approaches, which are constrained by the requirement for paired data. To address the aforementioned issues, this paper proposes a method that combines frequency decomposition and attention enhancement in an unsupervised bidirectional generative smoothing framework. It adopts cycle-consistent generative adversarial network, maintaining automatic parameter optimization while being applicable to unpaired data. To calibrate the generation direction, a structure combining frequency decomposition and feature fusion is proposed to enhance feature representation. By integrating the proposed feature reconstruction block, the network’s feature extraction and reconstruction capabilities are enhanced. Additionally, a top-k sparse attention module is proposed to adaptively retain the most critical attention scores, achieving better feature aggregation and improving generation performance. This paper successfully introduces an enhanced unsupervised image smoothing approach that addresses the limitations of existing methods. Experimental results demonstrate that the proposed method achieves superior performance compared to state-of-the-art smoothing techniques.}, archive = {J_NEUCOM}, author = {Jiafu Zeng and Yepeng Liu and Huiyu Li and Fan Zhang}, doi = {10.1016/j.neucom.2025.131085}, journal = {Neurocomputing}, month = {11}, pages = {131085}, shortjournal = {Neurocomputing}, title = {Unsupervised bidirectional generative smoothing framework with frequency decomposition and attention enhancement}, volume = {652}, year = {2025}, }
• (2025). ISPFormer: Learning RAW-to-sRGB mappings with wavelet-based self-attention. NEUCOM, 652, 131084. (www)

  @article{ , abstract = {RAW-to-sRGB mapping, or the simulation of the traditional camera image signal processor (ISP), aims to generate DSLR-quality sRGB images from RAW data captured by smartphone sensors. Despite achieving comparable results to sophisticated handcrafted camera ISP solutions, existing CNN-based methods still suffer from detail disparity and color distortion due to their inherent locality restrictions. In this paper, we present ISPFormer, a novel Transformer-based framework utilizing self-attention to tackle the learnable ISP problem. Specifically, we propose a Wavelet-based Transformer Block (WTB) with two loss functions to correct color and enhance high-frequency details, where WTB integrates wavelet transformation into window-based self-attention to perform self-attention in sub-bands across frequency domains, enabling larger sliding window modeling without additional computational overhead. Based on the proposed WTB, we construct the ISPFormer as a multi-stage network, where a multi-scale window adjustment strategy is further proposed to flexibly assign varying window sizes for each stage, reconstructing visually satisfactory results in a coarse-to-fine manner. Extensive experiments demonstrate that our ISPFormer achieves competitive quantitative and qualitative results. Code is available at https://github.com/RenYangSCU/ISPFormer .}, archive = {J_NEUCOM}, author = {Yang Ren and Xinhan Niu and Hai Jiang and Zhen Liu and Ting Jiang and Guanghui Liu and Shuaicheng Liu}, doi = {10.1016/j.neucom.2025.131084}, journal = {Neurocomputing}, month = {11}, pages = {131084}, shortjournal = {Neurocomputing}, title = {ISPFormer: Learning RAW-to-sRGB mappings with wavelet-based self-attention}, volume = {652}, year = {2025}, }
• (2025). PTDLRec: A recommendation model integrating personality traits and deep learning. NEUCOM, 652, 131083. (www)

  @article{ , abstract = {Recommendation systems have been widely applied in e-commerce, social media, and online content platforms, but traditional recommendation methods struggle to fully capture users’ true preferences. In recent years, the automatic feature extraction capabilities of deep learning have significantly improved recommendation performance, yet they still lack consideration of users’ long-term psychological traits. Therefore, this paper proposes a recommendation model that integrates personality traits and deep learning, empowering the recommendation system with psychological knowledge. Specifically, the model first employs matrix factorization to extract latent features of users and items; then, contrastive learning is used to optimize the embedding representations of user behavior and personality traits. Next, feature interaction is designed to deeply explore the complex relationships between users, items, and personality traits. Finally, a hierarchical attention mechanism is introduced to dynamically assign different feature weights, enabling a more flexible capture of user interests. The experiments conducted on publicly available datasets demonstrate that the proposed model outperforms other advanced recommendation models across multiple evaluation metrics, effectively improving personalized recommendation accuracy.}, archive = {J_NEUCOM}, author = {Xile Wang and Bing Li and Jiangtao Dong and Zijun Lin and Xiaojian Xing}, doi = {10.1016/j.neucom.2025.131083}, journal = {Neurocomputing}, month = {11}, pages = {131083}, shortjournal = {Neurocomputing}, title = {PTDLRec: A recommendation model integrating personality traits and deep learning}, volume = {652}, year = {2025}, }
• (2025). Attribute reduction via an improved granular ball model and attribute significance. NEUCOM, 652, 131082. (www)

  @article{ , abstract = {In the context of big data analysis, attribute reduction plays a crucial role in improving computational efficiency and data processing quality. The rough set method is a classical approach for attribute reduction, where granular ball neighborhood rough set (GBNRS) has shown significant advantages over conventional neighborhood rough set (NRS). However, GBNRS faces efficiency issues with high computational complexity when dealing with high-dimensional data. To address this challenge, this paper proposes improvements to the granular ball generation algorithm. Specifically, a classification significance threshold β is introduced during the granular ball splitting process to adaptively adjust the M value in the M-means algorithm. Furthermore, a customized attribute importance is used to evaluate the contribution of the attribute reduction set to classification, and an efficient model combining granular ball neighborhood rough set and attribute significance is proposed. By combining the positive region’s definition with the average purity of granular balls, granular ball neighborhood rough set with attribute significance (GBNAS) model algorithm is implemented. Experimental results demonstrate that, compared to the classical granular ball attribute reduction algorithms, the improved algorithm significantly enhances learning accuracy and efficiency in handling high-dimensional datasets, and extends the evaluation system for granular ball quality, showing distinct advantages.}, archive = {J_NEUCOM}, author = {Xin Wang and Bing Huang and Tianxing Wang}, doi = {10.1016/j.neucom.2025.131082}, journal = {Neurocomputing}, month = {11}, pages = {131082}, shortjournal = {Neurocomputing}, title = {Attribute reduction via an improved granular ball model and attribute significance}, volume = {652}, year = {2025}, }
• (2025). Cascaded embedded-FPN: A cross-modality multi-scale feature fusion network for varied-sized objects semantic segmentation. NEUCOM, 652, 131081. (www)

  @article{ , abstract = {Contemporary Red-Green-Blue-Thermal (RGB-T) semantic segmentation methods aim to enhance perception robustness by integrating RGB and thermal modalities. Nevertheless, RGB-T typically lacks effective multi-scale feature modeling, which limits its ability to simultaneously segment targets of varying sizes, such as distant pedestrians and nearby vehicles, thereby reducing its effectiveness in complex real-world scenarios, particularly in safety-critical domains like autonomous driving. Spurred by this limitation, this paper proposes a novel feature extraction backbone for RGB-T semantic segmentation tasks called Cascaded Embedded-Feature Pyramid Networks. The developed method embeds multi-scale representation capability into the backbone, allowing for simultaneous feature extraction and fusion within a unified framework. Furthermore, to mitigate the spatial discontinuity introduced by consecutive dilated convolutions in multi-scale feature extraction, we design a deep feature extraction block named Multi-layer FilterNet Block (MFB), where cascaded dilated convolutions with varying dilation rates and shifted anchor points are applied. This design enlarges the receptive field while preserving spatial continuity in the extracted features. On two open-source RGB-T datasets, MFNet and PST900, the developed model achieves pixel accuracies of 76.1 % and 86.9 %, respectively, outperforming the current state-of-the-art (SOTA) methods. Notably, incorporating multi-scale feature fusion within the backbone achieves enhanced performance while neglecting the memory burden on the model.}, archive = {J_NEUCOM}, author = {Yu Chen and Xiang Li and Chao Luan and Congmin Li and Jianguang Ma and Jianqi Zhang and Delian Liu and Xin Wu and Linfang Wei}, doi = {10.1016/j.neucom.2025.131081}, journal = {Neurocomputing}, month = {11}, pages = {131081}, shortjournal = {Neurocomputing}, title = {Cascaded embedded-FPN: A cross-modality multi-scale feature fusion network for varied-sized objects semantic segmentation}, volume = {652}, year = {2025}, }
• (2025). Gaussian-enhanced reinforcement learning for scalable evasion strategies in multi-agent pursuit-evasion games. NEUCOM, 652, 131080. (www)

  @article{ , abstract = {This paper introduces a Gaussian-enhanced multi-agent reinforcement learning framework for developing scalable evasion strategies in dynamic pursuit scenarios. The proposed methodology addresses two critical challenges in unknown environments: sparse reward structures and local optima convergence, while enhancing escape feasibility through probabilistic decision-making. By integrating Gaussian process regression with Q-function approximation, the framework enables efficient online parameter adaptation and demonstrates improved sample efficiency in high-dimensional state spaces. Comprehensive simulations and physical experiments across terrestrial and aerial robotic platforms validate the framework’s effectiveness and robustness in complex evasion tasks. The architecture’s modular design permits generalization to multi-agent pursuit-evasion scenarios with variable participant numbers, establishing a versatile foundation for strategic interactions in large-scale autonomous systems.}, archive = {J_NEUCOM}, author = {Ye Zhang and Yutong Zhu and Jingyu Wang}, doi = {10.1016/j.neucom.2025.131080}, journal = {Neurocomputing}, month = {11}, pages = {131080}, shortjournal = {Neurocomputing}, title = {Gaussian-enhanced reinforcement learning for scalable evasion strategies in multi-agent pursuit-evasion games}, volume = {652}, year = {2025}, }
• (2025). FFSL-aux: An accurate and efficient framework for hybrid federated few-shot learning. NEUCOM, 652, 131077. (www)

  @article{ , abstract = {Federated Learning (FL) decentralizes the model training process by distributing it across multiple clients, diverging from the traditional approach of aggregating data on a central server. This approach enables the server to harness the collective computational resources of all clients to facilitate distributed model training. Traditional federated learning research typically assumes that each client possesses sufficient local data for model training. However, in practical scenarios, client data is often Non-Independent and Identically Distributed (Non-IID) and often suffers from limited sample availability, posing significant challenges to model convergence and generalization. Therefore, it is essential to develop few-shot models that can effectively generalize to clients with limited data within federated learning scenarios. In this paper, this challenge is defined as Federated Few-Shot Learning. To address these challenges, we propose a novel hybrid federated few-shot learning framework. In this framework, participants contribute a minuscule fraction of non-sensitive data to a shared dataset, enabling the construction of an auxiliary model. This auxiliary model, along with decoupled server and client models and customized training strategies, collaboratively addresses the adverse effects of global data heterogeneity and local data scarcity. We conduct comprehensive experimental evaluations on four widely-used benchmark datasets covering both image and text modalities to validate the effectiveness and generalizability of the proposed framework. The results demonstrate that our framework consistently outperforms competitive methods across the evaluated datasets.}, archive = {J_NEUCOM}, author = {Wen Chen and Kuoliang Li and Zhicheng Yu and Ziye Zhang}, doi = {10.1016/j.neucom.2025.131077}, journal = {Neurocomputing}, month = {11}, pages = {131077}, shortjournal = {Neurocomputing}, title = {FFSL-aux: An accurate and efficient framework for hybrid federated few-shot learning}, volume = {652}, year = {2025}, }
• (2025). A modified finite-time prescribed performance adaptive nonsingular terminal super-twisting controller for robot joint module. NEUCOM, 652, 131073. (www)

  @article{ , abstract = {This research proposes a modified finite-time prescribed performance adaptive non-singular terminal super-twisting controller with neural network approximation to achieve high-precision control of the robot joint module. First, a modified finite-time performance function is introduced to ensure the prescribed tracking accuracy within a predetermined convergence time, while avoiding system singularities and enhancing robustness. Second, a non-singular terminal sliding surface is designed, and an adaptive super-twisting approaching law is proposed. This approach alleviates the requirement for precise knowledge of the disturbance boundaries and their derivatives and improves the system's closed-loop performance. Subsequently, an approximator utilizing neural network is developed to estimate system uncertainties, and an adaptive mechanism is put forward to adjust the weight of neural network in real time. Moreover, detailed analysis is presented to demonstrate that the system's state variables are steered towards a vicinity of the origin within a predetermined finite time span with the application of the proposed controller. Simulation and experimental results demonstrate that the proposed controller is superior to the other two controllers in terms of response speed, tracking accuracy and disturbance rejection capability.}, archive = {J_NEUCOM}, author = {Xiao Lin and Yankui Song and Junyang Li and Yu Xia and Hak-Keung Lam}, doi = {10.1016/j.neucom.2025.131073}, journal = {Neurocomputing}, month = {11}, pages = {131073}, shortjournal = {Neurocomputing}, title = {A modified finite-time prescribed performance adaptive nonsingular terminal super-twisting controller for robot joint module}, volume = {652}, year = {2025}, }
• (2025). Synchronization and pinning control of T-S fuzzy complex dynamical networks by using membership-function-dependent method. NEUCOM, 652, 131070. (www)

  @article{ , abstract = {The aim of this paper is to investigate the synchronization and pinning control problems of Takagi-Sugeno (T-S) fuzzy complex dynamical networks (CDNs). The information carried by the membership functions is utilized to obtain the synchronization conditions. Two different approaches are adopted, namely the linear matrix inequalities (LMI) method and the Kronecker order technique (KOT). Using the LMI method, we establish a set of membership-function-dependent synchronization conditions for T-S fuzzy CDNs based on a piecewise membership function approximation technique. With KOT, we construct effective Lyapunov functions by utilizing a weighted combination of normalized left eigenvectors for multiple outer coupling matrices. In pinning control analysis, we adopt the piecewise Lyapunov function approach and design local pinning controllers. In the end, we illustrate the benefits of our approach by presenting several numerical examples.}, archive = {J_NEUCOM}, author = {Xinmin Wei and Xiwei Liu}, doi = {10.1016/j.neucom.2025.131070}, journal = {Neurocomputing}, month = {11}, pages = {131070}, shortjournal = {Neurocomputing}, title = {Synchronization and pinning control of T-S fuzzy complex dynamical networks by using membership-function-dependent method}, volume = {652}, year = {2025}, }
• (2025). Multi-level feature enhancement and dual attention mechanisms for improved osteoporosis diagnosis. NEUCOM, 652, 131069. (www)

  @article{ , abstract = {Osteoporosis is a common metabolic bone disease characterized by decreased bone density and an increased risk of fractures, which can significantly diminish a patient’s quality of life. Early diagnosis is essential for effective management and treatment of this condition. In this study, we introduce a novel deep learning model aimed at improving osteoporosis detection by integrating dual attention mechanisms with multi-level feature fusion. Utilizing EfficientNetV2 as the backbone, our model incorporates deep semantic extraction to capture intricate spatial contexts and employs feature pyramid networks for multi-scale feature extraction. The integration of dual attention mechanisms further enhanced the ability of the model to focus on critical knee X-ray image regions, leading to improved classification performance. Comprehensive ablation studies revealed that the combination of deep semantic extraction and dual attention outperformed the existing methods in terms of Recall, Precision, F1 score, and Accuracy. Additionally, a comparative analysis with state-of-the-art models highlights the ability of our approach to reduce the false positive rate while enhancing diagnostic reliability significantly. The proposed model represents a notable advancement in medical image classification and provides a more accurate and reliable tool for osteoporosis detection.}, archive = {J_NEUCOM}, author = {Routhu Srinivasa Rao and Ishaq Muhammad and Bumshik Lee}, doi = {10.1016/j.neucom.2025.131069}, journal = {Neurocomputing}, month = {11}, pages = {131069}, shortjournal = {Neurocomputing}, title = {Multi-level feature enhancement and dual attention mechanisms for improved osteoporosis diagnosis}, volume = {652}, year = {2025}, }
• (2025). A noniterative linear-variational-inequality based primal-dual neural network for repetitive motion planning of robots. NEUCOM, 652, 131068. (www)

  @article{ , abstract = {In industrial applications, robots are usually employed to perform cyclical tasks. Quadratic programming (QP) based schemes are an effective solution for repetitive motion planning of robots. Since the robot joint variables change over time, the QP is time-variant. Most of the existing time-variant QP (TVQP) solvers are iterative algorithms, suffering from possible failures in meeting specified solution accuracy or real-time computing. Meanwhile, there exist a few noniterative solvers that require continuous and differentiable elements. Besides, bound constraints if involved must be equivalently converted into inequality constraints so that the noniterative solvers can be applied, increasing dimensions of the problem. To lift the above limitations, this paper proposes a novel noniterative linear-variational-inequality based primal-dual neural network (LVI-PDNN) that involves continuous but nondifferentiable projection operator, as well as its Dini derivatives, which is termed Dini-LVI-PDNN for convenience. While converging to the theoretical solution as time progresses, the Dini-LVI-PDNN solves each sampled QP problem by performing only one iteration of computations at each time step. The design procedure and convergence analysis of the Dini-LVI-PDNN solver are first mathematically detailed. Then, comparative numerical studies are conducted to demonstrate the effectiveness and superiority of the Dini-LVI-PDNN solver with regard to solution accuracy and real-time problem solving. Finally, the Dini-LVI-PDNN solver is successfully applied to robotic repetitive motion planning.}, archive = {J_NEUCOM}, author = {Weibing Li and Ruiqi Rao and Yuxuan Chen and Yongping Pan}, doi = {10.1016/j.neucom.2025.131068}, journal = {Neurocomputing}, month = {11}, pages = {131068}, shortjournal = {Neurocomputing}, title = {A noniterative linear-variational-inequality based primal-dual neural network for repetitive motion planning of robots}, volume = {652}, year = {2025}, }
• (2025). Temporal knowledge graph multi-hop path reasoning method based on reinforcement learning. NEUCOM, 652, 131067. (www)

  @article{ , abstract = {Temporal Knowledge Graphs (TKGs) extend traditional knowledge graphs by incorporating temporal information, enabling reasoning over time-dependent facts. However, many real-world knowledge graphs are incomplete, requiring effective reasoning methods to infer missing information and improve their overall quality. Multi-hop reasoning is a promising method for this task, but existing methods often fail to fully utilize temporal information and relationship modeling, resulting in less accurate inference. Additionally, many approaches lack interpretability, making it difficult to explicitly trace the reasoning pathways. To address these challenges, this paper proposes reinforcement learning-based multi-hop path reasoning for TKGs (RLPR), a novel model designed to enhance inference accuracy and interpretability by integrating temporal and relational information. RLPR introduces a timestamp decomposition strategy to better capture temporal dependencies while reducing model complexity. Unlike conventional entity-focused attention mechanisms, RLPR applies attention at the relationship level, allowing for more precise modeling of relational interactions. Furthermore, RLPR employs a reinforcement learning-based strategy network that explicitly constructs multi-hop reasoning pathways, improving both the interpretability and adaptability of the inference process. Compared to SOAT methods, the RLPR model achieved the best results on open-source datasets. For the link prediction results, the metrics improved by 15.07 %, 13.56 %, 4.26 % on the ICEWS14 datasets. When the path length setting for the RLPR model during reasoning is set to 2, the RLPR model improved various metrics on the ICEWS14–2 datasets by 13.04 %, 60.61 %, 38.67 %, and 27.08 %.}, archive = {J_NEUCOM}, author = {Yingping Su and Jianjun Cao and Dechang Pi}, doi = {10.1016/j.neucom.2025.131067}, journal = {Neurocomputing}, month = {11}, pages = {131067}, shortjournal = {Neurocomputing}, title = {Temporal knowledge graph multi-hop path reasoning method based on reinforcement learning}, volume = {652}, year = {2025}, }
• (2025). MOLGMP: A markov approach for molecular graph generation with GNNs. NEUCOM, 652, 131066. (www)

  @article{ , abstract = {Molecule generation has experienced multiple breakthroughs in recent years. While traditional techniques are very reliable they have low space exploration potential; therefore, machine learning techniques are needed to delve deep into the vast space of potential compounds, helping to find new ways of designing candidate molecules. This paper presents a sequential Markovian model based on graph neural networks for molecular generation. The model employs a Breadth–First Search (BFS) ordering strategy and a modular architecture to enhance independence between functions, with a focus on maintaining strict independence at each step of the Markovian process. Experimental results on ZINC (MOSES), ZINC (250 K) and Polymers datasets demonstrate the model’s ability to perform both unconditional and conditional molecule generation while preserving dataset properties. Notably, the step–wise approach achieves state–of–the–art results in terms of uniqueness and validity without using valency masks.}, archive = {J_NEUCOM}, author = {Benoıt Goupil and Antonin Joly and Niccoló Pancino and Pietro Bongini and Franco Scarselli and Monica Bianchini}, doi = {10.1016/j.neucom.2025.131066}, journal = {Neurocomputing}, month = {11}, pages = {131066}, shortjournal = {Neurocomputing}, title = {MOLGMP: A markov approach for molecular graph generation with GNNs}, volume = {652}, year = {2025}, }
• (2025). Safety-critical cooperative path following of uncertain nonlinear systems via unifying control lyapunov and control barrier functions. NEUCOM, 652, 131065. (www)

  @article{ , abstract = {This paper addresses the cooperative path following control of second-order uncertain nonlinear multi-agent systems subject to input constraints and safety constraints. Firstly, a predictor is developed based on neural networks to accurately approximate system nonlinearities and external disturbances. Then, control Lyapunov functions (CLFs) combined with control barrier functions (CBFs) are utilized to formulate the cooperative path following control objectives and safety constraints. A CLF-CBF-quadratic programming (QP) approach using neurodynamic optimization is developed to generate optimal control signals within the constraints to ensure both stability and safety. Compared to existing CBF-based safety-critical approaches designed for high-order systems where multiple time derivatives are required, a CBF backstepping method is proposed herein to consider the top-level of system without using multiple derivatives. The control objective of cooperative path following is realized and the input-to-state safety is guaranteed by using the proposed method. Simulation results of a fleet of autonomous surface vehicles are presented to demonstrate the effectiveness of the proposed CLF-CBF-QP method in cooperative path following.}, archive = {J_NEUCOM}, author = {Siming Cong and Nan Gu and Dan Wang and Yongming Li and Zhouhua Peng}, doi = {10.1016/j.neucom.2025.131065}, journal = {Neurocomputing}, month = {11}, pages = {131065}, shortjournal = {Neurocomputing}, title = {Safety-critical cooperative path following of uncertain nonlinear systems via unifying control lyapunov and control barrier functions}, volume = {652}, year = {2025}, }
• (2025). NSA-AE: An inadequately represented immune spaces NSA augmented via autoencoders. NEUCOM, 652, 131063. (www)

  @article{ , abstract = {Artificial immune systems (AIS) stand as a pivotal subset of artificial intelligence technologies, with the Negative Selection Algorithm (NSA) emerging as the foremost method for crafting immune detectors. Conventional NSAs rely excessively on prior knowledge, particularly for characterizing data and immune spaces (data sample spaces), hindering effective detector generation without predefined assumptions. Additionally, static strategies using constant detector radii as immune domains cause false negatives for non-self and false positives for self. These issues stem from traditional methods’ failure to adequately represent immune space distributions autonomously. Addressing these shortcomings, this paper introduces a novel paradigm, the Inadequately Represented Immune Space NSA augmented by AutoEncoders (AEs), abbreviated NSA-AE. This model initially employs AE to accurately estimate the distribution of incomplete antigens (data samples), which constitute the immune space. It then uses the modeled inadequately represented immune space to autonomously generate a broad and comprehensive assortment of immune candidate detectors. Following this, it incorporates a novel detector filtering mechanism, whereby redundancy is minimized via dynamic immune domain assessment among candidate and mature detectors. Comprehensive evaluations on a wide range of datasets confirm the effectiveness of NSA-AE, especially in the high-dimensional UCI and NSL-KDD datasets, where the anomaly detection rates are improved by 10 %-15 % and 9.16 % respectively, and the number of detectors is reduced by about 50 %.}, archive = {J_NEUCOM}, author = {Jiangchuan Chen and Junjiang He and Wenshan Li and Wenbo Fang and Xiaolong Lan and Wengang Ma and Tao Li and Dawei Zhao}, doi = {10.1016/j.neucom.2025.131063}, journal = {Neurocomputing}, month = {11}, pages = {131063}, shortjournal = {Neurocomputing}, title = {NSA-AE: An inadequately represented immune spaces NSA augmented via autoencoders}, volume = {652}, year = {2025}, }
• (2025). A novel stacking ensemble variant based on machine learning for short-term wind speed forecasting. NEUCOM, 652, 131062. (www)

  @article{ , abstract = {Wind power generation experienced a substantial increase in the last years, which makes wind power forecast tools an essential task to increase the use of renewable sources. This paper proposes a novel Stacking ensemble variant named Data Stacking based on Machine Learning algorithms for wind speed forecasting. This model combines the original input data with the predictions from the base learners to produce the final prediction. The method is validated using a benchmark dataset and a wind speed database from Brazil with several meteorological variables. Input data is preprocessed, and feature selection is applied. Results are evaluated under different error metrics with different forecasting horizons and compared with the performance of several Machine Learning algorithms and ensembles as Stacking and Multi-level Stacking. Results show the proposed Data Stacking presents superior forecasting performance with lower errors in all databases used. For wind speed database, Data Stacking resulted in mean absolute error (MAE) of 0.4855 and normalized root mean square error (nRMSE) of 0.2389. In this case, the error reduction obtained with Data Stacking in relation to Stacking varies from 0.7 % to 4.4 %, depending on the base learner algorithms chosen. An important characteristic observed is that Data Stacking can obtain satisfactory results even using base learners with bad forecasting prediction.}, archive = {J_NEUCOM}, author = {Sebastião B. Fonseca and Roberto C. L. de Oliveira and Carolina M. Affonso}, doi = {10.1016/j.neucom.2025.131062}, journal = {Neurocomputing}, month = {11}, pages = {131062}, shortjournal = {Neurocomputing}, title = {A novel stacking ensemble variant based on machine learning for short-term wind speed forecasting}, volume = {652}, year = {2025}, }
• (2025). Efficiency advantage actor-critic reinforcement learning control for an unmanned surface vehicle with unknown uncertainties. NEUCOM, 652, 131061. (www)

  @article{ , abstract = {For an unmanned surface vehicle (USV) with unknown uncertainties, this paper proposes an efficiency advantage actor-critic reinforcement learning control (EA2CRL) scheme to attain superior efficiency and optimality in tracking control. First, by utilizing control input and backstepping-based tracking error dynamics, a value function is designed to facilitate the derivation of an optimal control solution for the USV with unknown uncertainties. Then, a neural network (NN)-based advantage actor-critic reinforcement learning framework is developed to obtain the optimal solution through rapid learning. To address the challenge posed by unknown uncertainties, a critic NN without the dependency on system dynamics is deployed to approximate the value function, while the advantage function is integrated into an actor NN to reduce variance during the learning process and improve the tracking control efficiency of the USV. Eventually, a theoretical analysis establishes that adherence to optimal value function within this scheme ensures tracking performance. Simulations demonstrate the effectiveness of this scheme for the USV with unknown uncertainties.}, archive = {J_NEUCOM}, author = {Qiang Wang and Chun Liu and Yizhen Meng and Xiaoqiang Ren and Xiaofan Wang}, doi = {10.1016/j.neucom.2025.131061}, journal = {Neurocomputing}, month = {11}, pages = {131061}, shortjournal = {Neurocomputing}, title = {Efficiency advantage actor-critic reinforcement learning control for an unmanned surface vehicle with unknown uncertainties}, volume = {652}, year = {2025}, }
• (2025). Deep learning-based chromosome segmentation and extraction: A comprehensive review of methodologies, challenges, and future directions. NEUCOM, 652, 131060. (www)

  @article{ , abstract = {Chromosome karyotyping is fundamental to cytogenetics, facilitating the diagnosis of genetic disorders and malignancies through detailed structural analysis of chromosomes. A major technical challenge is the precise segmentation and extraction of complete, non-overlapping chromosomes, especially in cases involving dense chromosome clusters or significant morphological variation. Although deep learning has achieved notable success in general image processing, its application to chromosomal analysis has only recently gained momentum, and comprehensive evaluations remain scarce. This review systematically examines recent advances in deep learning-based chromosome segmentation and extraction, summarizing prevailing methodologies and key limitations. It traces the evolution from early convolutional neural networks to encoder-decoder architectures and generative models, highlighting advances in spatial detail recovery, robustness against overlapping structures, and domain adaptation. Furthermore, the paper categorizes chromosomal segmentation into semantic, instance, and hybrid paradigms, elucidates methodological trends such as the incorporation of biological priors and the adoption of multi-task learning, and discusses practical and cognitive challenges that hinder clinical implementation. By providing a comprehensive overview and outlining future directions—including explainable AI and synthetic data augmentation—this work aims to accelerate the development of intelligent, fully automated chromosome karyotyping systems.}, archive = {J_NEUCOM}, author = {Ge Song and Lianzheng Su and Xinmiao Wang and Zhonghao Huang and Shian Wang and Qiuyue Fu and Peng Wang}, doi = {10.1016/j.neucom.2025.131060}, journal = {Neurocomputing}, month = {11}, pages = {131060}, shortjournal = {Neurocomputing}, title = {Deep learning-based chromosome segmentation and extraction: A comprehensive review of methodologies, challenges, and future directions}, volume = {652}, year = {2025}, }
• (2025). Dual-stream collaborative network (DSCN): Multimodal sentiment analysis via modality-invariant and modality-specific representation learning. NEUCOM, 652, 131059. (www)

  @article{ , abstract = {Multimodal Sentiment Analysis (MSA) is a critical technology for human-machine interaction. It performs sentiment analysis by integrating complementary information from multiple modalities such as text, audio, and vision. However, most existing multimodal fusion methods regard the text as the primary modality and treat speech and visual modalities as auxiliary. This approach limits the potential for synergistic interaction among the three modalities. To address this limitation, this paper proposes a Dual-Channel Collaborative Transformer (DCCT), which enables collaborative interaction among the three modalities. DCCT incorporates a Self-modal Binding Channel (SBC) to enhance intra-modal consistency, and a Cross-modal Binding Channel (CBC) to complement the idiosyncratic features of other modalities. Then, a Feed-Forward Network (FFN) integrates the outputs of both channels for effective trimodal feature fusion. Furthermore, to mitigate the label shift problem in multi-task joint learning, we propose a framework that distinguishes between intensity and polarity tasks based on modality decomposition. By separating common and idiosyncratic features, we design dedicated tasks for intensity and polarity, and use a shared-private linear layer in the Trimodal Label Generator (TMLG) to capture the interrelation between the two aspects and facilitate task adaptation. Experiments conducted on the CMU-MOSI and CMU-MOSEI datasets show that the proposed Dual-Stream Collaborative Network (DSCN), which integrates DCCT and TMLG, achieves ACC-2 accuracy of 86.75 % and 86.69 %, respectively. These results outperform the best existing baseline by 0.65 % and 0.83 %.}, archive = {J_NEUCOM}, author = {Shengbing Chen and Dongmei Zhou and Chao Wei and Xingyang Li and Hongyu Yao and Hao Luo}, doi = {10.1016/j.neucom.2025.131059}, journal = {Neurocomputing}, month = {11}, pages = {131059}, shortjournal = {Neurocomputing}, title = {Dual-stream collaborative network (DSCN): Multimodal sentiment analysis via modality-invariant and modality-specific representation learning}, volume = {652}, year = {2025}, }
• (2025). Graph representation based reward shaping approach for addressing reward sparsity in task-oriented dialogue systems. NEUCOM, 652, 131058. (www)

  @article{ , abstract = {Task-Oriented Dialogue (TOD) systems rely on sequential decision-making to assist users in accomplishing specific goals. Due to this sequential nature, Reinforcement Learning (RL) has been widely adopted to train dialogue systems. However, reward sparsity remains a fundamental challenge in RL-based TOD systems, leading to inefficient exploration and suboptimal policy learning. While prior studies have focused on adopting advanced RL techniques to train TOD agents, the design of an effective reward function has received less attention. To address this challenge, this paper proposes a novel Potential-Based Reward Shaping (PBRS) approach that integrates graph representation learning with spectral graph clustering. This approach leverages graph representation learning algorithms to capture both structural and semantic relationships between dialogue states, facilitating more effective reward propagation. Moreover, the integration of Minimum Cut Pooling (MinCutPool) enables adaptive soft clustering, which promotes generalization across large and dynamic state spaces and remains robust to unseen or isolated states caused by frequent task switching or goal resetting. Extensive experiments were conducted to assess the effectiveness of the proposed approach on three benchmark tasks collected via Amazon Mechanical Turk. Evaluation metrics in this paper include success rate, average reward, dialogue turns, inference latency, number of parameters, CPU training time, and maximum CPU memory. Specifically, on the movie ticket booking task, the proposed GAT+MinCut+CSR model achieved 93.00 %, 28.14, 7.14, 96 ms, 80,143, 19982 s, and 7 GB, respectively, demonstrating superior performance and computational efficiency compared to state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Shaghayegh Saffari and Morteza Dorrigiv and Farzin Yaghmaee}, doi = {10.1016/j.neucom.2025.131058}, journal = {Neurocomputing}, month = {11}, pages = {131058}, shortjournal = {Neurocomputing}, title = {Graph representation based reward shaping approach for addressing reward sparsity in task-oriented dialogue systems}, volume = {652}, year = {2025}, }
• (2025). A weakly supervised salient object detection framework based on structured scribble. NEUCOM, 652, 131057. (www)

  @article{ , abstract = {Weakly Supervised Salient Object Detection (WSOD) methods based on scribble annotations provide a solution to the annotation burden from pixel-level annotations. However, due to the sparse nature of scribbles, models built upon them suffer from structure deficiency. In this paper, we propose to learn object structural information from the structured scribbles to address this issue. Structured scribbles incorporate object structural information into the scribble annotations by utilizing the object skeleton feature. On this basis, we propose a novel simple yet effective WSOD learning framework, which consists of a Multiscale Connection Module (MCM), a Detail Mining Module (DMM), and a Structure Guidance Module (SGM). MCM decodes and filters features at different scales to enhance the decoder’s ability to restore structures. DMM suppresses noise and extracts detail features via dual-stream attention. SGM integrates structure and detail features through a fusion, enhancement, and selection strategy to obtain the final prediction map. Moreover, we propose a plug-and-play Proportional Data Augmentation method to alleviate the impact of feature deformation and enhance the model’s generalization and practicality. Extensive experiments demonstrate that our method outperforms state-of-the-art WSOD methods, even reaching the performance level of fully supervised salient object detection methods. The dataset and code will be available at https://github.com/Rouiy/SeNet .}, archive = {J_NEUCOM}, author = {Congjin Gong and Haoyu Dong and Gang Yang and Pengyu Yang}, doi = {10.1016/j.neucom.2025.131057}, journal = {Neurocomputing}, month = {11}, pages = {131057}, shortjournal = {Neurocomputing}, title = {A weakly supervised salient object detection framework based on structured scribble}, volume = {652}, year = {2025}, }
• (2025). RKPOEM: A fine-grained dropout strategy for robust multi-view 3D hand reconstruction. NEUCOM, 652, 131054. (www)

  @article{ , abstract = {With the rapid development of virtual reality (VR), augmented reality (AR), and human-computer interaction, high-precision 3D hand reconstruction has become an important research topic in computer vision. Although existing multi-view reconstruction methods alleviate depth ambiguity by fusing information from multiple views, they often suffer from significant variations in prediction quality across views in complex real-world scenarios. Low-quality predictions from certain views can severely degrade the overall reconstruction performance. To address this issue, this paper proposes a robustness-enhancing strategy based on stochastic keypoint dropout from multi-view predictions, termed RKPOEM. By randomly dropping 2D keypoints estimated from specific views, the method reduces the model’s dependence on particular viewpoints and their keypoints, thereby improving its stability and generalization in the presence of unreliable single-view predictions. This strategy encourages the model to learn more robust spatial representations. Extensive experiments on benchmark datasets such as HO3D-MV, DexYCB-MV, and OakInk-MV, as well as a noise-injected variant of HO3D-MV, demonstrate that RKPOEM significantly reduces both the mean joint position error and the mean vertex position error, leading to substantial improvements in the accuracy and robustness of 3D hand reconstruction. Moreover, the proposed method introduces no additional inference overhead and can be seamlessly integrated into existing multi-view reconstruction frameworks, providing an efficient and cost-effective solution for practical applications.}, archive = {J_NEUCOM}, author = {Shengjiang Zhang and Tianyou Zheng and Zhijie Ren and Silong Sheng and Yang Zhang and Weiwei Fu}, doi = {10.1016/j.neucom.2025.131054}, journal = {Neurocomputing}, month = {11}, pages = {131054}, shortjournal = {Neurocomputing}, title = {RKPOEM: A fine-grained dropout strategy for robust multi-view 3D hand reconstruction}, volume = {652}, year = {2025}, }
• (2025). Multimodal representation learning with hierarchical knowledge decomposition for cancer survival analysis. NEUCOM, 652, 131053. (www)

  @article{ , abstract = {The accurate survival analysis of cancer patients serves as an important decision-making basis for formulating personalized treatment plans. Recent studies illustrate that integrating multimodal information including clinical diagnostic data, genomic features, and whole-slide images (WSIs) can significantly enhance the performance of prognostic assessment models. However, most existing multimodal survival analysis methods excessively focus on extracting shared inter-modal features while failing to effectively mine modality-specific biological information, resulting in inadequate acquisition of comprehensive patient multimodal representations. Furthermore, how to extract prognosis-related tissue microenvironment features from ultra-high-resolution WSIs remains an unresolved open challenge. To address these issues, a multimodal representation learning framework with hierarchical knowledge decomposition (MRL-HKD) is proposed for cancer survival analysis. MRL-HKD transforms multimodal representation learning into a set partitioning problem within multimodal knowledge spaces, and employs a hierarchical multimodal knowledge decomposition module to decouple complex inter-modal relationships. Meanwhile, to address the challenge of high-dimensional pathological image feature extraction, a gated attention mechanism-based multimodal patch attention network is designed. The performance comparison experiments on four cancer datasets demonstrate that MRL-HKD significantly outperforms state-of-the-art methods. Our study demonstrates the potential of gated attention mechanisms and hierarchical knowledge decomposition in multimodal survival analysis, and provides an effective tool for cancer prognosis prediction. The source code will be open-sourced at https://github.com/yangxulin/MRL-HKD .}, archive = {J_NEUCOM}, author = {Xulin Yang and Hang Qiu}, doi = {10.1016/j.neucom.2025.131053}, journal = {Neurocomputing}, month = {11}, pages = {131053}, shortjournal = {Neurocomputing}, title = {Multimodal representation learning with hierarchical knowledge decomposition for cancer survival analysis}, volume = {652}, year = {2025}, }
• (2025). A review of advancements in low-light image enhancement using deep learning. NEUCOM, 652, 131052. (www)

  @article{ , abstract = {In low-light environments, the performance of computer vision algorithms often deteriorates significantly, adversely affecting key vision tasks such as segmentation, detection, and classification. With the rapid advancement of deep learning, its application to low-light image processing has attracted widespread attention and seen significant progress in recent years. However, there remains a lack of comprehensive surveys that systematically examine how recent deep-learning-based low-light image enhancement methods function and evaluate their effectiveness in enhancing downstream vision tasks. To address this gap, this review provides a detailed elaboration on how various recent approaches (from 2020) operate and their enhancement mechanisms, supplemented with clear illustrations. It also investigates the impact of different enhancement techniques on subsequent vision tasks, critically analyzing their strengths and limitations. Our review found that image enhancement improved the performance of downstream vision tasks to varying degrees. Although supervised methods often produced images with high perceptual quality, they typically produced modest improvements in vision tasks. In contrast, zero-shot learning, despite achieving lower scores in image quality metrics, showed consistently boosted performance across various vision tasks. These suggest a disconnect between image quality metrics and those evaluating vision task performance. Additionally, unsupervised domain adaptation techniques demonstrated significant gains in segmentation tasks, highlighting their potential in practical low-light scenarios where labelled data is scarce. Observed limitations of existing studies are analysed, and directions for future research are proposed. This review serves as a useful reference for determining low-light image enhancement techniques and optimizing vision task performance in low-light conditions.}, archive = {J_NEUCOM}, author = {Fangxue Liu and Lei Fan}, doi = {10.1016/j.neucom.2025.131052}, journal = {Neurocomputing}, month = {11}, pages = {131052}, shortjournal = {Neurocomputing}, title = {A review of advancements in low-light image enhancement using deep learning}, volume = {652}, year = {2025}, }
• (2025). Multimodal and multichannel speech separation using location-guided speech feature mapping network. NEUCOM, 652, 131051. (www)

  @article{ , abstract = {In reality, the audio and visual signals of sound sources are closely aligned, working collaboratively to isolate the desired speech signal from overlapping voices of simultaneous talkers. To leverage the complementarity and utilize all available information from both auditory and visual sources in speech separation, we propose a novel robust multimodal and multichannel speech separation method, without requiring known camera parameters. The proposed method exploits the complementarity of audio and visual modalities to estimate the speaker’s location and adopts a location-guided speech feature mapping strategy, wherein the attention mechanism fusion method combines the high-level semantic information of auditory and visual sources, aiding in the separation of target speech with its corresponding directional features. Experimental results suggest that the proposed multimodal and multichannel speech separation system outperforms the baselines, demonstrating improvements of 0.64 dB in SI-SDR and 0.17 in PESQ, respectively. The proposed system consistently outperformed the baselines by achieving a 10.14 % absolute (26.05 % relative) word error rate (WER) reduction.}, archive = {J_NEUCOM}, author = {Yulin Wu and Xiaochen Wang and Dengshi Li and Ruimin Hu}, doi = {10.1016/j.neucom.2025.131051}, journal = {Neurocomputing}, month = {11}, pages = {131051}, shortjournal = {Neurocomputing}, title = {Multimodal and multichannel speech separation using location-guided speech feature mapping network}, volume = {652}, year = {2025}, }
• (2025). Improving vision-language models through intra-modal contrastive learning-based hard sample mining. NEUCOM, 652, 131047. (www)

  @article{ , abstract = {Driving environmental perception is a core component of autonomous driving systems. Recently, emerging vision-language detectors, known for their superior detection accuracy, have gradually replaced traditional detectors and have been increasingly applied in open-world driving scenarios. However, these detectors still face challenges regarding the missed detection of hard positive samples. This study identifies that a primary cause of this problem is the rejection of hard samples due to their low cross-modal consistency. To address this challenge, this work proposes a contrastive learning strategy based on a hard sample prototype memory bank to recall the potential positive samples. Additionally, to enhance the representational capacity of the detection network, an instance-level contrastive learning loss is introduced. This loss aligns the feature representations of the same instance across the deep and shallow network layers, thereby improving the ability of shallow layers to extract features from hard samples. Experimental results demonstrate that the proposed method achieves outstanding detection accuracy and is highly effective in complex urban road scenarios. The code and trained models are available at https://github.com/unbelieboomboom/HSMG_DINO.}, archive = {J_NEUCOM}, author = {Chaojun Lin and Ying Shi and Gang Wang and Shijian Liu}, doi = {10.1016/j.neucom.2025.131047}, journal = {Neurocomputing}, month = {11}, pages = {131047}, shortjournal = {Neurocomputing}, title = {Improving vision-language models through intra-modal contrastive learning-based hard sample mining}, volume = {652}, year = {2025}, }
• (2025). Clip4Vis: Parameter-free fusion for multimodal video recognition. NEUCOM, 652, 131046. (www)

  @article{ , abstract = {Multimodal video recognition has emerged as a central focus due to its ability to effectively integrate information from diverse modalities, such as video and text. However, traditional fusion methods typically rely on trainable parameters, resulting in increased model computational costs. To address these challenges, this paper presents Clip4Vis , a zero-parameter progressive fusion framework that combines video and text features using a shallow-to-deep approach. The shallow and deep fusion steps are implemented through two key modules: (i) Cross-Model Attention , a module that enhances video embeddings with textual information, enabling adaptive focus on keyframes to improve action representation in the video. (ii) Joint Temporal-Textual Aggregation , a module that integrates video embeddings and word embeddings by jointly utilizing temporal and textual information, enabling global information aggregation. Extensive evaluations on five widely used video datasets demonstrate that our method achieves competitive performance in general, zero-shot, and few-shot video recognition. Our best model, using the released CLIP model, achieves a state-of-the-art accuracy of 87.4 % for general recognition on Kinetics-400 and 75.3 % for zero-shot recognition on Kinetics-600. The code will be released later.}, archive = {J_NEUCOM}, author = {Qishi Zheng and Mengnan He and Jiuqin Duan and Gai Luo and Pengcheng Wu and Yimin Han and Qingyue Min and Peng Chen and Ping Zhang}, doi = {10.1016/j.neucom.2025.131046}, journal = {Neurocomputing}, month = {11}, pages = {131046}, shortjournal = {Neurocomputing}, title = {Clip4Vis: Parameter-free fusion for multimodal video recognition}, volume = {652}, year = {2025}, }
• (2025). A functional RNA–based and curvature-aware learning framework for federated optimization. NEUCOM, 652, 131045. (www)

  @article{ , abstract = {Federated Learning (FL) enables decentralized model training without sharing raw data, but faces major challenges under non-IID client distributions, communication inefficiencies, and convergence to suboptimal local minima. To address these issues, we propose FRB-FL—a novel F unctional R NA– B ased and Curvature-Aware Learning framework for Federated Learning. FRB-FL integrates two synergistic components: a biologically inspired decentralized peer-to-peer communication protocol, and a curvature-sensitive optimization strategy modeled on quantum tunneling dynamics. Inspired by SID-1 mediated intercellular RNA signaling, clients exchange models with dynamically selected peers to achieve decentralized consensus and reduce variance without relying on a central server. To further escape sharp local minima, we introduce a stochastic tunneling acceptance rule that probabilistically allows uphill updates based on loss curvature, analogous to escape probabilities in quantum physics. Together, these mechanisms enable FRB-FL to converge faster, generalize better, and remain robust under heterogeneous, asynchronous, and unreliable environments. Experiments on three biomedical datasets (PneumoniaMNIST, EEG motor imagery, and cancer genomics) demonstrate that FRB-FL significantly outperforms strong FL baselines (FedAvg, SCAFFOLD, FedDyn, FedOpt) in both convergence speed and final performance, while also offering resilience to client dropouts, adversarial updates, and privacy leakage.}, archive = {J_NEUCOM}, author = {Koffka Khan and Wayne Goodridge and Winston Elibox}, doi = {10.1016/j.neucom.2025.131045}, journal = {Neurocomputing}, month = {11}, pages = {131045}, shortjournal = {Neurocomputing}, title = {A functional RNA–based and curvature-aware learning framework for federated optimization}, volume = {652}, year = {2025}, }
• (2025). Hyperbolic graph representation learning: Methods, applications and challenges—A survey. NEUCOM, 652, 131044. (www)

  @article{ , abstract = {Graph-structured data is widely used in real-world applications such as social networks, recommender systems, and transportation networks. With the rapid expansion of network scale and the continuous emergence of massive data, graph representation learning has become a research hotspot in the field of graph data mining. Current research on graph representation learning focuses on graph features in Euclidean space. These investigations are simple and efficient but hard to accurately capture the potential power-law distribution and hierarchical structure characteristics of complex networks, which fail to achieve lower distortion when embedding tree-like structured networks. The exponential growth property of hyperbolic geometry is suitable for representing tree-like hierarchical networks and provides unique advantages in embedding heterogeneous or scale-free graphs, making hyperbolic space a novel embedding space for graph representation learning. In this survey, we present a comprehensive overview of the geometric knowledge, methods, tools, and applications on Hyperbolic Graph Representation Learning (HGRL). Firstly, the geometric knowledge involved in HGRL is summarized, with an emphasis on the hyperbolic models. Secondly, we propose a new taxonomy of methods for HGRL and illustrate some examples. Next, we introduce two tools applied in HGRL. Then, the current applications of HGRL in different fields are examined. Finally, directions for further research are outlined, providing insights into future advancements in HGRL.}, archive = {J_NEUCOM}, author = {Ying Chen and Lianming Zhang and Jiusheng Li and Pingping Dong}, doi = {10.1016/j.neucom.2025.131044}, journal = {Neurocomputing}, month = {11}, pages = {131044}, shortjournal = {Neurocomputing}, title = {Hyperbolic graph representation learning: Methods, applications and challenges—A survey}, volume = {652}, year = {2025}, }
• (2025). MambaCOD: Camouflaged object detection with state-space model. NEUCOM, 652, 131043. (www)

  @article{ , abstract = {Camouflaged object detection (COD) is a crucial task in the computer vision field, entailing significant applications in areas such as surveillance, medical imaging, and wildlife monitoring. Despite significant advancements in recent years, state-of-the-art COD algorithms still encounter challenges in handling occlusions, managing high background similarity, detecting small objects, and maintaining performance under low-light conditions. To tackle these challenges, in this study, we propose a novel COD framework named MambaCOD, which capitalizes on the state-space model mechanism. Specifically, we propose the Mamba architecture-assisted cross-scale vision state-space module (CS-VSSM) to enhance the scale and contextual diversity of extracted features in the encoding stage, enabling MambaCOD to more effectively detect camouflaged targets at various scales. Furthermore, the Mamba architecture-aided hierarchical vision state-space module (HVSSM) is designed to capture both global and local semantic dependencies during the decoding phase, thereby improving MambaCOD’s ability to handle complex visual conditions, such as occlusion and high similarity. Extensive experiments on four prominent COD benchmarks, namely CHAMELEON, CAMO, COD10K, and NC4K, demonstrate that MambaCOD effectively addresses these challenges, achieving advanced detection accuracy and surpassing state-of-the-art methods. Overall, this work underscores the significant potential of the Mamba architecture and the state-space model mechanism in advancing COD performance, highlighting promising directions for future research aimed at enhancing the robustness of COD systems.}, archive = {J_NEUCOM}, author = {Zhouyong Liu and Taotao Ji and Chunguo Li and Yongming Huang and Luxi Yang}, doi = {10.1016/j.neucom.2025.131043}, journal = {Neurocomputing}, month = {11}, pages = {131043}, shortjournal = {Neurocomputing}, title = {MambaCOD: Camouflaged object detection with state-space model}, volume = {652}, year = {2025}, }
• (2025). Adversarial detector based on multi-scale texture and natural scene statistics feature fusion in object detection. NEUCOM, 652, 131041. (www)

  @article{ , abstract = {Adversarial attacks can significantly compromise object detectors based on Deep Neural Networks (DNNs), leading to erroneous predictions and posing severe security risks in real-world applications. These attacks typically induce changes in both pixel statistical features and texture characteristics of an image. In response, we propose a novel adversarial sample detection method, the Gabor-NSS Isolation Forest Detector (GN-IFD), which integrates multi-scale texture and Natural Scene Statistics (NSS) features. The proposed approach employs Mean Subtracted Contrast Normalized (MSCN) coefficients to capture the image’s statistical properties, which are crucial for Non-Subsample Sensitivity modeling in detecting shifts in these statistical characteristics. Simultaneously, texture features are extracted using Gabor filters across multiple directions and scales, and subsequently fused with statistical features to create a robust data representation resilient to adversarial perturbations. Additionally, an unsupervised Isolation Forest algorithm is employed to construct an adversarial sample detector, ensuring computational efficiency and suitability for lightweight deployment. The effectiveness and versatility of the proposed method are rigorously assessed through extensive experiments on four distinct object detectors using the PASCAL-VOC dataset. The results demonstrate that GN-IFD exhibits exceptional sensitivity to adversarial samples, achieving a near-perfect recall rate in most detection scenarios, and maintains superior performance across various object detection models, showcasing remarkable transferability.}, archive = {J_NEUCOM}, author = {Heng Zhang and Jiasong Ding and Qin Hang and Jiaxin Guo and Hongtao Liu}, doi = {10.1016/j.neucom.2025.131041}, journal = {Neurocomputing}, month = {11}, pages = {131041}, shortjournal = {Neurocomputing}, title = {Adversarial detector based on multi-scale texture and natural scene statistics feature fusion in object detection}, volume = {652}, year = {2025}, }
• (2025). Behavioral psychology of LLMs: Better task guidance through punishment and reinforcement. NEUCOM, 652, 131040. (www)

  @article{ , abstract = {Recent advances in Large Language Models (LLMs) have sparked interest in their emotional intelligence, but their behavioral psychology—particularly emotion-driven decision-making akin to humans— remains poorly understood. This study introduces Behavioral Consequence Scenarios (BCSs), a structured framework simulating psychologically salient punishment and reinforcement dynamics. We test the emotional perception and empathy abilities of LLMs using psychological scales and examine the emotional variation patterns across different BCSs. The findings suggest that LLMs exhibit rudimentary emotional perception and empathy capabilities, and their emotional responses generally align with human behavior. Building on this, we propose Behavioral Consequence Scenario Prompting (BCSP), a method that injects BCS-driven emotional incentives into task instructions. Experiments on Wizard-of-Wikipedia and CMMLU demonstrate BCSP’s efficacy: in dialogue generation, BCSP achieved at least a 3.53/1.74 improvement in BLEU-2/4, and in the multi-task language understanding test, BCSP gained a 2 % improvement in overall accuracy. Our findings bridge LLMs’ alignment with human-like emotion-behavior dynamics, and encourage deeper investigation into the behavioral psychology and emotional patterns of LLMs.}, archive = {J_NEUCOM}, author = {Sisi Peng and Wenlin Zhang and Shunhang Li and Dan Qu and Han Wei}, doi = {10.1016/j.neucom.2025.131040}, journal = {Neurocomputing}, month = {11}, pages = {131040}, shortjournal = {Neurocomputing}, title = {Behavioral psychology of LLMs: Better task guidance through punishment and reinforcement}, volume = {652}, year = {2025}, }
• (2025). SMAFusion: Multimodal medical image fusion based on spatial registration and local-global multi-scale feature adaptive fusion. NEUCOM, 652, 131039. (www)

  @article{ , abstract = {Aiming at the current image fusion methods relying entirely on paired images after registration, the fusion efficiency is low, and the existing image registration methods overlook the impact of image distribution differences on registration outcomes, as well as the poor capability of the image fusion methods for local-global multi-scale feature extraction from source images during image fusion, we propose a multimodal medical image fusion method based on spatial registration and local-global multi-scale feature adaptive fusion (SMAFusion). In SMAFusion, the spatial registration module is utilized to minimize the impact of image distribution differences on the registration results. Meanwhile, a local-global multi-scale feature encoder is proposed to fully extract local and global information at different scales. And an adaptive fusion strategy is employed to fuse the multi-scale features of different modal images. The experimental results show that this method has an average improvement of 24 %, 121.04 %, 19.57 %, 40.36 %, and 97.62 % compared to the eight comparative methods in five evaluation indicators: mutual information (MI), difference correlation sum (SCD), visual fidelity (VIFF), gradient fusion performance (QAB/F) and structural similarity (SSIM), respectively. The fused image preserves more details and texture information, and exhibits high consistency with the original image in terms of overall structure while enhancing the fusion ability for unregistered images. The significant improvement in MI and SCD indicates that the fused image better integrates complementary information from multiple modalities, helping doctors obtain more diagnostic clues from a single image and reducing the risk of missed or incorrect diagnoses. The improvements in VIFF and SSIM suggest that the fused image performs better in structure preservation and visual clarity, making lesion areas and anatomical structures clearer, thereby enhancing clinical readability and diagnostic efficiency. The enhancement in QAB/F further demonstrates the ability to preserve image edges and texture details, which is directly beneficial for observing fine structures.}, archive = {J_NEUCOM}, author = {Wei Guo and Lifang Wang and Jianchao Zeng and Qiang Han and KaiXin Jin and Xiwen Wang}, doi = {10.1016/j.neucom.2025.131039}, journal = {Neurocomputing}, month = {11}, pages = {131039}, shortjournal = {Neurocomputing}, title = {SMAFusion: Multimodal medical image fusion based on spatial registration and local-global multi-scale feature adaptive fusion}, volume = {652}, year = {2025}, }
• (2025). Advancing mobile robot navigation with DRL and heuristic rewards: A comprehensive review. NEUCOM, 652, 131036. (www)

  @article{ , abstract = {Robotic navigation is a critical component of autonomy, requiring efficient and safe mobility across diverse environments. The advent of Deep Reinforcement Learning (DRL) has spurred significant research into enabling mobile robots to learn effective navigation by optimizing actions based on environmental rewards. DRL has shown promise in addressing challenges such as dynamic environments and cooperative exploration. However, traditional DRL-based navigation faces several limitations, including the need for extensive training data, susceptibility to local traps in complex environments, low transferability to real-world scenarios, slow convergence, and low learning efficiency. Additionally, designing an appropriate reward function to achieve desired behaviors without unintended consequences remains complex; poorly designed rewards can lead to suboptimal or harmful outcomes. Recent studies have explored integrating heuristic search-based rewards into DRL algorithms to mitigate these issues. This study reviews the limitations of traditional DRL navigation and explores recent advancements in integrating heuristic search to design dynamic reward functions that enhance robot learning processes.}, archive = {J_NEUCOM}, author = {Mazbahur Rahman Khan and Azhar Mohd Ibrahim and Suaib Al Mahmud and Farah Asyiqin Samat and Farahiyah Jasni and Muhammad Imran Mardzuki}, doi = {10.1016/j.neucom.2025.131036}, journal = {Neurocomputing}, month = {11}, pages = {131036}, shortjournal = {Neurocomputing}, title = {Advancing mobile robot navigation with DRL and heuristic rewards: A comprehensive review}, volume = {652}, year = {2025}, }
• (2025). Optimizing global parameter synchronization for geo-distributed machine learning in reconfigurable optical wide area networks. NEUCOM, 652, 131025. (www)

  @article{ , abstract = {Geo-distributed machine learning (Geo-DML) usually uses a hierarchical training architecture, local parameter synchronization (LPS) within data center and global parameter synchronization (GPS) between data centers. Compared to fast LAN bandwidth, the heterogeneous and scarce WAN bandwidth becomes one of the main bottlenecks of training performance for Geo-DML. Fortunately, the emerging optical technologies render the modern WAN topology reconfigurable, which has been adopted to improve the performance of some traditional traffic with the help of software-defined networking (SDN). However, the reconfigurable WAN topology is often overlooked by most schemes aimed at accelerating Geo-DML. In this paper, we propose AdaptivePS , an adaptive global parameter synchronization scheduling scheme that leverages the reconfigurable feature of WAN topology and the training characteristics to speed up Geo-DML training. Specifically, mathematical optimization models considering the topology construction and parameter synchronization scheduling are firstly established. Then AdaptivePS solves the mathematical models through relaxing and deterministic rounding scheme, obtaining the deployment of global aggregation nodes, wavelength allocation, path and rate allocation. The simulation results based on real WAN topologies show that compared to RoWAN , RAPIER and Baseline , AdaptivePS can reduce global communication time (GCT) by up to 73.4 % , 86.7 % , 96.2 % , respectively. This demonstrates that AdaptivePS can effectively cope with different network environments, with the help of adaptive selection of global aggregation nodes, reconfigurable topology, and mathematical model based scheduling.}, archive = {J_NEUCOM}, author = {Ling Liu and Wang Xu and Pan Zhou and Xiaoqiong Xu and Xi Chen and Hongfang Yu and Gang Sun}, doi = {10.1016/j.neucom.2025.131025}, journal = {Neurocomputing}, month = {11}, pages = {131025}, shortjournal = {Neurocomputing}, title = {Optimizing global parameter synchronization for geo-distributed machine learning in reconfigurable optical wide area networks}, volume = {652}, year = {2025}, }
• (2025). Semi-supervised label distribution learning via global factorization and local constrain. NEUCOM, 652, 131024. (www)

  @article{ , abstract = {In label distribution learning, properly handling samples with missing label distributions is a particularly challenging task. When dealing with unlabeled samples, leveraging correlation is especially crucial as it reveals the intrinsic patterns of data distribution and effectively reduces the model’s hypothesis space. Currently, semi-supervised label distribution learning follows the same correlation mining methods as those used under complete supervision. However, due to the lack of supervision information for some samples, these methods designed for complete supervision are insufficient in a semi-supervised context. On one hand, the absence of labels for some samples makes it difficult to mine label correlations; on the other hand, label correlations mined solely based on samples are biased, leading to imprecise label correlations due to the missing labels. To address these issues, this paper innovatively proposes two strategies for mining label correlations in semi-supervised label distribution learning: first, exploring the common correlations between known and unknown label distributions; second, using the information of known label distributions to reveal the correlations of unknown label distributions. Specifically, globally, we employ independent component analysis for matrix completion of missing sample labels, and locally, we improve the k -NN framework to utilize the label constraints of known label distributions to restrict the label distribution values of unknown label distributions. Based on these mined correlations, we designed a semi-supervised label distribution learning algorithm. The algorithm outperforms existing methods in 67.27 % of cases, achieving outstanding performance, and demonstrates significant statistical significance in two-sample t -tests.}, archive = {J_NEUCOM}, author = {Peiqiu Yu and Xiuyi Jia}, doi = {10.1016/j.neucom.2025.131024}, journal = {Neurocomputing}, month = {11}, pages = {131024}, shortjournal = {Neurocomputing}, title = {Semi-supervised label distribution learning via global factorization and local constrain}, volume = {652}, year = {2025}, }
• (2025). A weight-sharing based RGB-T image semantic segmentation network with hierarchical feature enhancement and progressive feature fusion. NEUCOM, 652, 131023. (www)

  @article{ , abstract = {RGB-T image segmentation algorithms have been widely adopted in various fields, such as surveillance and autonomous driving. These algorithms typically employ separate encoders to ensure each branch extracts modality-specific features. Nevertheless, this design increases parameters and potential conflicts between multimodal features. An alternative solution is using a weight-sharing encoder, which facilitates consistent encoding across different data types, reducing the number of training parameters and enhancing the encoder’s generalization capability. However, a weight-sharing encoder tends to extract modality-shared features but under-represent modality-specific features, thereby limiting segmentation performance under heterogeneous sensor conditions. To preserve the modality-shared features of multimodal data while simultaneously enhancing modality-specific features, we propose a novel Weight-Sharing based RGB-T image semantic segmentation network (WSRT) with a Hierarchical Feature Enhancement Module (HFEM) and Progressive Fusion Decoder (PFD). HFEM first integrates the modality-shared information from the weight-sharing encoded feature maps to generate enhanced modality-shared feature maps. Subsequently, it utilizes these enhanced feature maps to generate relative modality-specific feature maps of the RGB and thermal modalities. PFD is proposed to progressively integrate multi-scale features from different stages to decode these enhanced features more effectively. Experimental results on multiple RGB-T image semantic segmentation datasets demonstrate that our method achieves top-ranking performance or competitive results. The code is available at: https://github.com/bearxwm/WSRT .}, archive = {J_NEUCOM}, author = {Xue Weimin and Liu Yisha and Zhuang Yan}, doi = {10.1016/j.neucom.2025.131023}, journal = {Neurocomputing}, month = {11}, pages = {131023}, shortjournal = {Neurocomputing}, title = {A weight-sharing based RGB-T image semantic segmentation network with hierarchical feature enhancement and progressive feature fusion}, volume = {652}, year = {2025}, }
• (2025). Safe coordinated optimization of the thickening-dewatering process via reinforcement learning with real-time human guidance. NEUCOM, 652, 131022. (www)

  @article{ , abstract = {Due to its trial-and-error learning mechanism and limited intelligence, current reinforcement learning (RL) faces significant safety risks when applied to complex industrial scenarios. To improve its deployability in high-risk environments, this paper takes the thickening-dewatering process, a key and energy-intensive subprocess in mineral processing, as the research object and proposes a safe RL coordination optimization framework that leverages real-time human guidance mechanisms. The framework consists of two human-in-the-loop models: first, a human supervision model based on soft sensing, which predicts the safety of the agent’s actions at each step and identifies potential risks in advance; second, a human demonstration model based on imitation learning, which automatically generates safe alternative actions in line with human expertise when unsafe actions are detected. Finally, the safe actions, evaluated and filtered by the models, are used for interaction with the environment to ensure the safety of the RL process. Furthermore, this paper derives the upper bound of the discounted failure probability for the algorithm, theoretically validating the safety enhancement provided by the human guidance mechanism. Experimental results demonstrate that, while achieving a 100 % training safety rate, the proposed algorithm reduces energy consumption by 15.62 % compared to existing optimization algorithm, showing significant potential for practical application and broader deployment.}, archive = {J_NEUCOM}, author = {Ranmeng Lin and Runda Jia and Fengyang Jiang and Jun Zheng and Dakuo He and Kang Li and Fuli Wang}, doi = {10.1016/j.neucom.2025.131022}, journal = {Neurocomputing}, month = {11}, pages = {131022}, shortjournal = {Neurocomputing}, title = {Safe coordinated optimization of the thickening-dewatering process via reinforcement learning with real-time human guidance}, volume = {652}, year = {2025}, }
• (2025). Epistemic graph: A plug-and-play module for hybrid representation learning. NEUCOM, 652, 131021. (www)

  @article{ , abstract = {In recent years, deep models have achieved remarkable success in various vision tasks. However, their performance heavily relies on large training datasets. In contrast, humans exhibit hybrid learning, seamlessly integrating structured knowledge for cross-domain recognition or relying on a smaller amount of data samples for few-shot learning. Motivated by this human-like epistemic process, we aim to extend hybrid learning to computer vision tasks by integrating structured knowledge with data samples for more effective representation learning. Nevertheless, this extension faces significant challenges due to the substantial gap between structured knowledge and deep features learned from data samples, encompassing both dimensions and knowledge granularity. In this paper, a novel Epistemic Graph Layer (EGLayer) is introduced to enable hybrid learning, enhancing the exchange of information between deep features and a structured knowledge graph. Our EGLayer is composed of three major parts, including a local graph module, a query aggregation model, and a novel correlation alignment loss function to emulate human epistemic ability. Serving as a plug-and-play module that can replace the standard linear classifier, EGLayer significantly improves the performance of deep models. Extensive experiments demonstrate that EGLayer can greatly enhance representation learning for the tasks of cross-domain recognition and few-shot learning, and the visualization of knowledge graphs can aid in model interpretation.}, archive = {J_NEUCOM}, author = {Jin Yuan and Shikai Chen and Yicheng Jiang and Yang Zhang and Zhongchao Shi and Jianping Fan and Yong Rui}, doi = {10.1016/j.neucom.2025.131021}, journal = {Neurocomputing}, month = {11}, pages = {131021}, shortjournal = {Neurocomputing}, title = {Epistemic graph: A plug-and-play module for hybrid representation learning}, volume = {652}, year = {2025}, }
• (2025). ST-RLNet: Spatio-temporal representation learning for multi-step traffic flow prediction. NEUCOM, 652, 131020. (www)

  @article{ , abstract = {Traffic flow prediction provides valuable traffic information to transportation agencies and individuals in advance. Compared to next-step prediction, multi-step prediction provides users with traffic information for a longer time horizon, allowing users to have a more comprehensive understanding of traffic conditions. So far, various methods have been proposed for multi-step traffic flow prediction. However, most of them become sub-optimal in effectively detecting the spatio-temporal correlations of traffic data. Furthermore, as the number of prediction steps increases, the input data used to predict the flow of the next step tends to deviate further from the ground truth value. This deviation leads to a rapid decrease in prediction accuracy as the number of prediction steps increases. To address these issues, in this paper, we propose a deep spatio-temporal representation learning network named ST-RLNet for multi-step traffic flow prediction. The goal is to effectively generate the traffic data representation by better capturing the complex correlations of the data. In particular, we design a network called 3D-ConvLSTMNet to effectively extract short-term and long-term spatio-temporal data correlations for the next step prediction. To solve the performance degradation problem, we propose a feedback mechanism called PS-Feedback to dynamically reconstruct temporal correlation representations of input traffic flow for each round of next-step prediction. To evaluate the performance of the ST-RLNet, we conduct extensive experiments on two real-world datasets. Experimental results show that the ST-RLNet outperforms the state-of-the-art methods in both next-step and multi-step predictions, and exhibits consistent high performance under different traffic flows.}, archive = {J_NEUCOM}, author = {Lihua He and Si Shi and Dian Zhang and Wuman Luo}, doi = {10.1016/j.neucom.2025.131020}, journal = {Neurocomputing}, month = {11}, pages = {131020}, shortjournal = {Neurocomputing}, title = {ST-RLNet: Spatio-temporal representation learning for multi-step traffic flow prediction}, volume = {652}, year = {2025}, }
• (2025). CLIP-based autonomous visual prompting for unsupervised domain incremental learning. NEUCOM, 652, 131019. (www)

  @article{ , abstract = {Domain Incremental Learning (DIL) focuses on updating the model with knowledge from novel domains (i.e., currently being processed domain) while maintaining the knowledge acquired from previous domains, mitigating the catastrophic forgetting in deep learning. In real-world applications, annotating a large-scale dataset for DIL is expensive and time-consuming. This paper focuses on a more realistic and challenging task: Open-Vocabulary Unsupervised Domain Incremental Learning (OV-UDIL) assumes that all the data in the novel domain are unlabeled during the incremental learning stage, and the vocabulary list is opened across all domains during training. Considering that Vision-Language Models (VLMs) contain abundant semantic knowledge, they can be applied to many downstream tasks in a zero-shot setting. Therefore, we propose a novel CLIP-based Autonomous Visual Prompting (AVP) approach for UDIL. Specifically, we first generate and tune the visual prompt using the unlabeled data from the current domain to uncover and leverage the knowledge embedded in CLIP. Then, we leverage visual prompts from the previous domain to adjust the corresponding prompts in the current domain, thereby preserving each prompt’s particular attribute. Overall, AVP alleviates the adverse effects of domain feature shifts on CLIP’s generalization capability, enhancing recognition performance on the UDIL task. Extensive experiments on four popular datasets demonstrate that our method achieves superior performance compared with state-of-the-art DIL approaches.}, archive = {J_NEUCOM}, author = {Jiaping Yu and Muli Yang and Jiexi Yan and Cheng Deng}, doi = {10.1016/j.neucom.2025.131019}, journal = {Neurocomputing}, month = {11}, pages = {131019}, shortjournal = {Neurocomputing}, title = {CLIP-based autonomous visual prompting for unsupervised domain incremental learning}, volume = {652}, year = {2025}, }
• (2025). Black-box targeted adversarial attacks for deep neural networks. NEUCOM, 652, 131017. (www)

  @article{ , abstract = {Deep neural networks have exhibited exceptional performance across a wide spectrum of applications. However, they remain vulnerable to adversarial samples. To address the issue of poor transferability of adversarial samples in cross-domain targeted attacking scenarios, we propose a novel method named Cross Domain Dual Training (CD-DT) utilizing Generative Adversarial Networks (GANs). Within this framework, the generative adversarial network comprises one generator and two trainable discriminators. To improve the transferability of adversarial samples in cross domain attacks, we incorporate three techniques into the training process of GAN. The differential minimization training technique with three distinct loss functions is employed to amplify the efficacy of adversarial samples against different target models. The Cross-domain Cross-Entropy (CCE) function is applied to replace the Cross-Entropy function used in classical GANs to diminish the sensitivity of the generated perturbations to the underlying data distribution. In addition, we introduce a data distribution alignment method that utilizes three loss functions to ensure the generated adversarial samples keep the similar distribution with the target samples from both global and local perspectives, thereby improving their transferability to the target class. Experimental results demonstrate that our method significantly enhances the transferability of adversarial samples in complex cross-domain and cross-model scenarios compared to existing baseline methods.}, archive = {J_NEUCOM}, author = {Xueqin Zhang and Peilin Geng and Wei Hong and Yixuan Wang and Chunhua Gu}, doi = {10.1016/j.neucom.2025.131017}, journal = {Neurocomputing}, month = {11}, pages = {131017}, shortjournal = {Neurocomputing}, title = {Black-box targeted adversarial attacks for deep neural networks}, volume = {652}, year = {2025}, }
• (2025). ConvMamba: Combining mamba with CNN for hyperspectral image classification. NEUCOM, 652, 131016. (www)

  @article{ , abstract = {Hyperspectral Image (HSI) Classification is crucial in remote sensing and various practical applications. Traditional methods face significant challenges: CNN-based approaches have limited receptive fields, while Transformer-based models, though effective in modeling long-range dependencies, are computationally expensive for high-dimensional hyperspectral data. To address these issues, we propose ConvMamba, a framework that combines the advantages of the State Space Model (SSM) for long-sequence modeling with the capability of CNNs for local spatial feature extraction, thereby enhancing computational efficiency. Specifically, we introduce the Mamba-Conv structure, which leverages Mamba (a linear sequence model) to establish dependencies between the central pixel (the pixel to be classified) and a wide range of neighboring pixels, achieving the integration of spectral information from the central pixel with spatial information from the surrounding pixels. Concurrently, CNN compensates for the linear sequence model’s limitations in modeling local spatial invariance, accelerating model convergence. Additionally, to further enhance spectral-spatial information fusion, we propose a novel fusion structure, CenterFusion. This structure explicitly fuses patch-level features obtained through 1D convolution with the central pixel information extracted by Mamba. This fusion fully utilizes the spectral information contained in the central pixel and the spatial information in the surrounding pixels, further improving classification performance. Extensive experiments on three benchmark hyperspectral datasets demonstrate that ConvMamba achieves state-of-the-art performance while significantly reducing computational costs. Ablation studies further validate the effectiveness of the key components in the proposed framework. The source code of ConvMamba is publicly available at https://github.com/zhysess/ConvMamba .}, archive = {J_NEUCOM}, author = {Haoyuan Zhang and Hui Liu and Zexi Shi and Shanjun Mao and Ning Chen}, doi = {10.1016/j.neucom.2025.131016}, journal = {Neurocomputing}, month = {11}, pages = {131016}, shortjournal = {Neurocomputing}, title = {ConvMamba: Combining mamba with CNN for hyperspectral image classification}, volume = {652}, year = {2025}, }
• (2025). Integrating large foundation models into multimodal named entity recognition with evidential fusion. NEUCOM, 652, 131015. (www)

  @article{ , abstract = {Multimodal Named Entity Recognition (MNER) is a crucial task for information extraction from social media platforms such as Twitter. Most current methods rely on attention weights to extract information from both text and images but are often unreliable and lack interpretability. To address this problem, we propose incorporating uncertainty estimation into the MNER task, producing trustworthy predictions. Our proposed algorithm models the distribution of each modality as a Normal-inverse Gamma distribution, and fuses them into a unified distribution with an evidential fusion mechanism, enabling hierarchical characterization of uncertainties and promotion of prediction accuracy and trustworthiness. Additionally, we explore the potential of pre-trained large foundation models in MNER and propose an efficient fusion approach that leverages their robust feature representations. Experiments on two datasets demonstrate that our proposed method outperforms the baselines and achieves new state-of-the-art performance. Our code is available at https://github.com/ZhongAobo/evi-mner .}, archive = {J_NEUCOM}, author = {Weide Liu and Xiaoyang Zhong and Jingwen Hou and Shaohua Li and Haozhe Huang and Wei Zhou and Yuming Fang}, doi = {10.1016/j.neucom.2025.131015}, journal = {Neurocomputing}, month = {11}, pages = {131015}, shortjournal = {Neurocomputing}, title = {Integrating large foundation models into multimodal named entity recognition with evidential fusion}, volume = {652}, year = {2025}, }
• (2025). Continually learn to map visual concepts to language models in resource-constrained environments. NEUCOM, 652, 131013. (www)

  @article{ , abstract = {Continually learning from non-independent and identically distributed (non-i.i.d.) data poses a significant challenge in deep learning, particularly in resource-constrained environments. Visual models trained via supervised learning often suffer from overfitting, catastrophic forgetting, and biased representations when faced with sequential tasks. In contrast, pre-trained language models demonstrate greater robustness in managing task sequences due to their generalized knowledge representations, albeit at the cost of high computational resources. Leveraging this advantage, we propose a novel learning strategy, Continual Visual Mapping (CVM), which continuously maps visual representations into a fixed knowledge space derived from a language model. By anchoring learning to this fixed space, CVM enables training small, efficient visual models, making it particularly suited for scenarios where adapting large pre-trained visual models is computationally or data-prohibitive. Empirical evaluations across five benchmarks demonstrate that CVM consistently outperforms state-of-the-art continual learning methods, showcasing its potential to enhance generalization and mitigate challenges in resource-constrained continual learning settings.}, archive = {J_NEUCOM}, author = {Clea Rebillard and Julio Hurtado and Andrii Krutsylo and Lucia Passaro and Vincenzo Lomonaco}, doi = {10.1016/j.neucom.2025.131013}, journal = {Neurocomputing}, month = {11}, pages = {131013}, shortjournal = {Neurocomputing}, title = {Continually learn to map visual concepts to language models in resource-constrained environments}, volume = {652}, year = {2025}, }
• (2025). Causal dreamer for partially observable model-based reinforcement learning. NEUCOM, 652, 131012. (www)

  @article{ , abstract = {Model-based reinforcement learning (MBRL) leverages an agent’s experience to construct a dynamics model, thereby improving the data efficiency of policy learning. While MBRL methods have achieved impressive results on fully observable (Markovian) tasks, their effectiveness on partially observable tasks remains limited. Among existing methods, Dreamer—built upon a recurrent variational autoencoder to model latent dynamics and perform trajectory imagination—has shown the most promise for such tasks. However, Dreamer struggles in environments that require long-term memory reasoning due to the large history space, which renders dynamics prediction and policy learning inefficient. To address this issue, we propose Causal Dreamer , which introduces a self-supervised task called history counterfactual inference to identify and discard uninformative observations from the agent’s history. By retaining only the observations that are critical for accurate dynamics prediction, Causal Dreamer improves multi-step transition modeling and enhances policy performance. Experiments on long-horizon partially observable tasks, including Maze, BabyAI, and robotic navigation environments, demonstrate that Causal Dreamer significantly outperforms prior methods in both sample efficiency and final performance.}, archive = {J_NEUCOM}, author = {Haichuan Gao and Tianrun Xu and Tianren Zhang and Yuqing Guo and Chujie Zhao and Jinsheng Ren and Yizhou Jiang and Shangqi Guo and Feng Chen}, doi = {10.1016/j.neucom.2025.131012}, journal = {Neurocomputing}, month = {11}, pages = {131012}, shortjournal = {Neurocomputing}, title = {Causal dreamer for partially observable model-based reinforcement learning}, volume = {652}, year = {2025}, }
• (2025). SportSummarizer: A unified multimodal fusion transformer for context-aware sports video summarization. NEUCOM, 652, 131011. (www)

  @article{ , abstract = {Automated sports video summarization faces critical challenges due to the complexity of dynamic gameplay, event variability, and the intricate rules governing sports like cricket and soccer. Existing methods often struggle to capture key moments accurately, resulting in false positives, redundant content like replays, ineffective multimodal data integration, and challenges in spatio-temporal modeling and semantic event understanding. To overcome these limitations, a novel Unified Multimodal Fusion Transformer is proposed for the summarization of cricket and soccer videos. This approach utilizes advanced feature encoding across multiple modalities: ViViT for video, OpenL3 for audio, and DistilBERT for text, ensuring robust multimodal representations. A multimodal fusion transformer with contextual cross-quadrimodal attention is introduced to address weak multimodal integration, enabling the model to capture complex interactions across visual, audio, and textual data for precise event detection. Further, Hierarchical Temporal Convolutional Networks (Hierarchical TCN) module integrates hierarchical temporal and metadata-enhanced positional to model both short and long-term game sequences effectively. Additionally, replay and redundancy elimination mechanisms are applied to remove repetitive content, generating concise and high-quality video summaries that reflect the game's critical moments. The proposed method achieves state-of-the-art results, with the highest precision (99.2 %) and recall (98.9 %), and a low error rate (4 %). It also demonstrates superior ROC-AUC performance (0.88) and maintains peak accuracy (89.5 %) with strong performance in mIoU (0.82) and highlight diversity (0.93), highlighting its robustness across various event detection metrics.}, archive = {J_NEUCOM}, author = {D. Minola Davids and A. Arul Edwin Raj and C. Seldev Christopher}, doi = {10.1016/j.neucom.2025.131011}, journal = {Neurocomputing}, month = {11}, pages = {131011}, shortjournal = {Neurocomputing}, title = {SportSummarizer: A unified multimodal fusion transformer for context-aware sports video summarization}, volume = {652}, year = {2025}, }
• (2025). Blind source separation of single-channel mixtures via multi-encoder autoencoders. NEUCOM, 652, 131008. (www)

  @article{ , abstract = {The task of blind source separation (BSS) involves separating sources from a mixture without prior knowledge of the sources or the mixing system. Addressing the challenges of single-channel and non-linear mixtures, we propose a novel approach leveraging the feature subspace specialization capability of the multi-encoder autoencoder variant introduced in this paper. During training, our method unmixes source representations into the separate encoder outputs and subsequently remixes them in the decoder for reconstruction of the input. To facilitate source inference, we introduce a novel encoding masking technique, enabling the decoder to estimate individual source signals by selectively masking all but one encoding. Additionally, we propose two new loss functions: a sparse mixing loss to encourage sparse remixing of source encodings, and a zero reconstruction loss to ensure coherent source estimations. We validate our approach using a toy dataset that highlights feature subspace specialization and real-world biosignal data from polysomnography studies, demonstrating effective extraction of respiratory signals from electrocardiogram (ECG) and photoplethysmography (PPG) recordings. Our results show that multi-encoder autoencoders offer a promising solution for single-channel, non-linear BSS.}, archive = {J_NEUCOM}, author = {Matthew B. Webster and Joonnyong Lee}, doi = {10.1016/j.neucom.2025.131008}, journal = {Neurocomputing}, month = {11}, pages = {131008}, shortjournal = {Neurocomputing}, title = {Blind source separation of single-channel mixtures via multi-encoder autoencoders}, volume = {652}, year = {2025}, }
• (2025). Federated graph neural networks in non-IID scenarios—A comprehensive survey. NEUCOM, 652, 131007. (www)

  @article{ , abstract = {Federated Graph Neural Networks (FedGNNs) have emerged as a promising solution to securely train structured graph data in Federated learning (FL) settings. In this paper, we present one of the first works that categorizes the various non-IID (Non-Identically and Independently Distributed) scenarios and challenges occurring in FedGNNs, offering insights into horizontal and vertical non-IID cases. Horizontal non-IID refers to variations in data distributions among clients, while vertical non-IID involves attribute and label disparities within the clients. We briefly discuss works addressing these scenarios and their respective advantages and disadvantages. Additionally, we explore other approaches like centralized and decentralized methods, in mitigating non-IID effects, highlighting their benefits in terms of shared knowledge, privacy preservation, and scalability. Furthermore, we emphasize the importance of evaluating and quantifying non-IIDness in graph data through statistical measures. Our work contributes to the understanding of FedGNNs’ applicability in healthcare, finance, recommender systems, transportation, and other domains. We also identify future research directions, such as taxonomy development, handling complete structural heterogeneity, and exploring adaptive mechanisms, to enhance the robustness and reliability of FedGNNs in real-time scenarios.}, archive = {J_NEUCOM}, author = {Abdullah Abdul Sattar Shaikh and Saeed Samet}, doi = {10.1016/j.neucom.2025.131007}, journal = {Neurocomputing}, month = {11}, pages = {131007}, shortjournal = {Neurocomputing}, title = {Federated graph neural networks in non-IID scenarios—A comprehensive survey}, volume = {652}, year = {2025}, }
• (2025). Distributional consistency-guided prompt learning for vision-language models. NEUCOM, 652, 131003. (www)

  @article{ , abstract = {In the field of Vision-Language (V-L), prompt learning has emerged as a crucial technique for enhancing model adaptability. However, existing methods often focus on using a single set of multi-modal prompts and fail to fully utilize label information, limiting their ability to capture diverse feature representations. To address these challenges, we propose a novel approach named Distributional Consistency-Guided Prompt Learning (DCG), which combines the strengths of a partial coupling projection scheme for generating multiple sets of prompts and a consistency constraint using Maximum Mean Discrepancy (MMD) to enhance feature alignment. The former enables the generation of multiple sets of prompts while preserving modality-specific characteristics. To mitigate the feature misalignment introduced by diverse prompts, we employ Maximum Mean Discrepancy (MMD) to align the distributions of features extracted from multiple sets of prompts with those of the original pre-trained model, thereby enhancing generalization to downstream tasks. We conduct extensive experiments on 11 publicly available vision-language datasets spanning diverse domains and tasks, and the results consistently demonstrate that DCG achieves superior performance in both effectiveness and robustness compared to existing prompt learning approaches.}, archive = {J_NEUCOM}, author = {Zhichao Wu and Bo Sun and Jun He}, doi = {10.1016/j.neucom.2025.131003}, journal = {Neurocomputing}, month = {11}, pages = {131003}, shortjournal = {Neurocomputing}, title = {Distributional consistency-guided prompt learning for vision-language models}, volume = {652}, year = {2025}, }
• (2025). Fixed-time synchronization of octonion-valued neural networks with mixed delays: A non-separation norm approach. NEUCOM, 652, 130995. (www)

  @article{ , abstract = {This work emphasizes the fixed-time synchronization (FTS) of a specific class of octonion-valued neural networks (OVNNs), which incorporate discrete and distributed time delays. This study also establishes several norm properties for the octonion domains and explores FTS and fixed-time projective synchronization (FTPS) in OVNNs having mixed time delays by a suitable choice of the Lyapunov function, controllers and the one norm property. Unlike previous research on the decomposition of neural networks with octonion-valued and quaternion-valued components, this study introduces an enhanced one-norm method based on the non-separation approach. It employs a direct analytical approach to address two synchronization challenges using several norm properties. The computational complexity is reduced to provide less conservative results for OVNNs. This article analyzes various properties of the one norm of octonion domains and introduces effective controllers for achieving FTS and FTPS between the drive and response systems of OVNNs. The present article also establishes results in a compact and more generalized form using one norm criteria, which are easily verifiable to ensure synchronization, even with mixed time delays, which can be achieved within fixed time intervals. The settling time in each case illustrates its effectiveness compared to the existing results in a more straightforward way through a unique analytical process and more versatile activation functions. Finally, the theoretical results are validated through two numerical examples, with the overall results presented and discussed. Additionally, OVNNs are employed to demonstrate their effectiveness in storing and retrieving true-color images. This application showcases the ability of OVNNs to handle high-dimensional data representations, particularly in contexts where color channels and spatial features are strongly interrelated. The results highlight the robustness and efficiency of the proposed OVNNs framework, confirming its potential for advanced multidimensional data processing tasks.}, archive = {J_NEUCOM}, author = {Vaibhav Agrawal and Sunny Singh and Vineet Kumar Singh and Subir Das}, doi = {10.1016/j.neucom.2025.130995}, journal = {Neurocomputing}, month = {11}, pages = {130995}, shortjournal = {Neurocomputing}, title = {Fixed-time synchronization of octonion-valued neural networks with mixed delays: A non-separation norm approach}, volume = {652}, year = {2025}, }
• (2025). Trifocal SLAM: A dynamic SLAM based on three frame views. NEUCOM, 652, 130993. (www)

  @article{ , abstract = {This paper proposes a new and effective approach, named Trifocal SLAM, to address the intensely complex task of SLAM for environments containing different types of dynamic objects. These include objects clearly identified as dynamic (e.g. cars, humans, etc) as well as those which are not originally classified as dynamic yet still in motion (such as chairs, boxes etc). Trifocal SLAM is built upon the basic structure of ORB SLAM2, with additional semantic segmentation thread and a novel motion consistency module. Unlike traditional two-view geometric methods used by many popular approaches, this work introduces a new and improved method based on a trifocal coplanarity constraint, which leverages three consecutive frames for its motion consistency module. This method effectively addresses the major limitation of epipolar geometry and provides better separation of dynamic and static points in the environment. It turns out that this geometric method alone achieves comparatively better results, not only compared to other geometric methods (including epipolar geometry) but also to many deep learning-based methods. To further improve performance, we integrated additional semantic segmentation using YOLOv8-seg. The combination of our multiview geometric method and semantic segmentation enables Trifocal SLAM to outperform many new and established methods, demonstrating up to a 98.38 % improvement in RMSE of absolute trajectory compare to ORB SLAM2. We validated our approach using public datasets, ranging from less complex environments like TUM RGB-D to more complex scenarios such as the BONN dataset. Additionally, we validated and compared our method in real-world scenarios within a custom laboratory environment.}, archive = {J_NEUCOM}, author = {Sujan Dhali and Bhaskar Dasgupta}, doi = {10.1016/j.neucom.2025.130993}, journal = {Neurocomputing}, month = {11}, pages = {130993}, shortjournal = {Neurocomputing}, title = {Trifocal SLAM: A dynamic SLAM based on three frame views}, volume = {652}, year = {2025}, }
• (2025). Adaptive-TOD: An LLM-driven and adaptive agent for diverse interaction modes. NEUCOM, 652, 130991. (www)

  @article{ , abstract = {Previous studies on Task-Oriented Dialogue systems have primarily concentrated on straightforward tasks, utilizing basic strategies such as “Request and Inform" within dialogues. Although expected results have been achieved in simulated scenarios, it becomes evident that this solitary interaction mode is inadequate for complex and dynamic processes in real industrial business. Handling such intricate real-world tasks requires not only consideration of unpredictable user behaviors but also execution feedback from back-end services and workflow constraints. To tackle this challenge, we introduce, for the first time, a novel task termed Diverse Interaction Conversation (DIC) and its corresponding framework in this paper. It offers a comprehensive analysis of varied user behaviors, dynamic execution feedback, and the relationship between them and system interaction policies. Next, we specifically design a task flow schema based on a directed graph structure to illustrate these dynamic and complex task workflows. In response to DIC scenarios, we further develop an adaptive LLM-driven agent (Adaptive-TOD), which includes three modules: an Adaptive Interaction Policy Planning module for generating flexible interactive responses; a Session State Management module for maintaining dialogue status and a Task Flow Engine designed to efficiently parse dynamic task flow schema. Collaborating with industry experts, we collect DIC dataset covering eight industrial tasks to comprehensively evaluate the adaptive system. The experimental findings reveal that, even when solely utilizing an On-Device LLM with instruction-following modes, Adaptive-TOD still exhibits adaptable interaction capabilities and superior task workflow consistency. These results highlight the practical application potential and wide industrial expansion prospects.}, archive = {J_NEUCOM}, author = {Pan Li and Qingwen Yang and Sai Xu and Xuejing Li and Zhi Li and Chen Wang and Yanyi Liu and Tiezheng Guo and Jiawei Tang and Yingyou Wen}, doi = {10.1016/j.neucom.2025.130991}, journal = {Neurocomputing}, month = {11}, pages = {130991}, shortjournal = {Neurocomputing}, title = {Adaptive-TOD: An LLM-driven and adaptive agent for diverse interaction modes}, volume = {652}, year = {2025}, }
• (2025). A DLM watermarking method based on a spatiotemporal chaos with DNA computing. NEUCOM, 652, 130981. (www)

  @article{ , abstract = {Intellectual property (IP) protection for deep learning models (DLM) remains a hotspot, while the main solution is to give each model a universal and useful identity, which is analogous to the identification systems in human society. Recently, black-box watermarking technique has emerged as the primary option for IPP, however, small key space and fraudulent ownership claim attacks are still unresolved. In this paper, we proposed a black-box watermarking method based on a spatiotemporal chaos, Arnold Coupled Logistic Map Lattices (ACLML), with DNA permutation. Firstly, the ACLML can provide favorable chaotic properties to the trigger set and make it unpredictable against machine learning attacks and statistical inference. Secondly, the motion of the ACLML is controlled by particular parameters, which can provide a large key space and assign each model a unique identifier, meeting the commercialization needs of DLM. Thirdly, the trigger samples and chaotic values that build the trigger set are mutually independent, guaranteeing the security of the watermark. Theoretical analysis indicates that our scheme is secure and practical. We also compared it with the previous method, the experimental results demonstrate that our method shows better robustness against fine-tuning attacks and overwriting attacks. Moreover, it also effectively suppresses fraudulent ownership claim attacks.}, archive = {J_NEUCOM}, author = {Dehui Wang and Yingqian Zhang and Qiang Wei and Yumei Xue and Shuang Zhou}, doi = {10.1016/j.neucom.2025.130981}, journal = {Neurocomputing}, month = {11}, pages = {130981}, shortjournal = {Neurocomputing}, title = {A DLM watermarking method based on a spatiotemporal chaos with DNA computing}, volume = {652}, year = {2025}, }
• (2025). Privacy-enhanced data distillation with probability distribution matching. NEUCOM, 652, 130953. (www)

  @article{ , abstract = {Data distillation aims to condense the large-scale original dataset into a small-scale synthetic dataset while preserving as much data utility as possible. As one of the typical implementation mechanisms of data distillation, distribution matching works by aligning the feature distributions of synthetic and original samples, while avoiding the expensive computation and memory costs associated with other matching mechanisms. However, there still exist two primary limitations in distribution matching. On the one hand, distribution matching suffers from inadequate class discrimination, the synthetic samples within the same class may be misclassified as other classes due to the scattered feature distribution. On the other hand, distribution matching raises serious privacy concerns, as the synthetic dataset may inadvertently contain some sensitive information extracted from the original dataset. Taking this cue, we propose here a novel privacy-enhanced distribution matching-based data distillation algorithm. First, we design a probability distribution matching method with intra-class aggregation constraint and inter-class dispersion constraint based on symmetric Kullback-Leibler divergence to strengthen the performance of data distillation. Second, we design a dynamic noise perturbation method based on differential privacy to enhance data privacy guarantees while preserving higher sample quality. Extensive experiments demonstrate that our algorithm can achieve performance improvements of up to 4.5 % on the CIFAR10 dataset and 2.7 % on the SVHN dataset, compared to the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Ke Pan and Yuxin Wen and Yiming Wang and Maoguo Gong and Hui Li and Shanfeng Wang}, doi = {10.1016/j.neucom.2025.130953}, journal = {Neurocomputing}, month = {11}, pages = {130953}, shortjournal = {Neurocomputing}, title = {Privacy-enhanced data distillation with probability distribution matching}, volume = {652}, year = {2025}, }
• (2025). Emotion agent: Unsupervised deep reinforcement learning with distribution-prototype reward for continuous emotional EEG analysis. NEUCOM, 652, 130951. (www)

  @article{ , abstract = {Continuous Electroencephalography (EEG) signals are widely employed in affective brain-computer interface (aBCI) applications. However, only a subset of the continuously acquired EEG data is truly relevant to emotional processing, while the remainder is often noisy or unrelated. Manual annotation of these key emotional segments is impractical due to their dynamic and individualized nature. To address this challenge, we propose a novel unsupervised deep reinforcement learning framework, termed Emotion Agent , which automatically identifies and extracts the most informative emotional segments from continuous EEG signals. Emotion Agent initially utilizes a heuristic algorithm to perform a global search and generate prototype representations of the EEG signals. These prototypes guide the exploration of the signal space and highlight regions of interest. Furthermore, we design a distribution-prototype-based reward function that evaluates the interaction between samples and prototypes to ensure that the selected segments are both representative and relevant to the underlying emotional states. Finally, the framework is trained using Proximal Policy Optimization (PPO) to achieve stable and efficient convergence. Experimental results on three widely used datasets (covering both discrete and dimensional emotion recognition) show an average improvement of 13.46 % when using the proposed Emotion Agent, demonstrating its significant enhancement of accuracy and robustness in downstream aBCI tasks.}, archive = {J_NEUCOM}, author = {Zhihao Zhou and Li Zhang and Qile Liu and Gan Huang and Zhuliang Yu and Zhen Liang}, doi = {10.1016/j.neucom.2025.130951}, journal = {Neurocomputing}, month = {11}, pages = {130951}, shortjournal = {Neurocomputing}, title = {Emotion agent: Unsupervised deep reinforcement learning with distribution-prototype reward for continuous emotional EEG analysis}, volume = {652}, year = {2025}, }
• (2025). Dynamic T-distributed stochastic neighbor graph convolutional networks for multi-modal contrastive fusion. NEUCOM, 652, 130950. (www)

  @article{ , abstract = {As the continuous advancement of data acquisition technologies progresses, multi-modal data have emerged as a prominent focus in various domains. This paper aims to tackle critical challenges in the multi-modal fusion process, specifically in representation learning, modal consistency invariance learning, and model diversity complementarity learning, by employing graph convolutional networks and contrastive learning methods. Current GCN-based methods generally depend on predefined graphs for representation learning, limiting their capacity to capture local and global information effectively. Furthermore, some current models do not adequately compare the representations of consistency and diversity across different modalities during the fusion procedure. To address the identified challenges, we propose a novel T-distributed Stochastic Neighbor Contrastive Graph Convolutional Network (TSNGCN). It consists of the adaptive static graph learning module, the multi-modal representation learning module, and the multi-modal contrastive fusion module. The adaptive static graph learning module constructs graphs without relying on any predefined distance metrics, which creates a pairwise graph adaptively to preserve the local structure of general data. Moreover, a loss function based on T-distributed stochastic neighbor embedding is designed to learn the transformation between the embeddings and the original data, thus facilitating the exploration of more discriminative information within the learned subspace. In addition, the proposed multi-modal contrastive fusion module effectively maximizes the similarity of the same samples across different modalities while ensuring the distinction of dissimilar samples, thereby enhancing the model’s consistency objective. Extensive experiments conducted on several multi-modal benchmark datasets demonstrate the superiority and effectiveness of TSNGCN compared to existing methods.}, archive = {J_NEUCOM}, author = {Bo Xu and Guoxu Li and Jie Wang and Zheng Wang and Jianfu Cao and Rong Wang and Feiping Nie}, doi = {10.1016/j.neucom.2025.130950}, journal = {Neurocomputing}, month = {11}, pages = {130950}, shortjournal = {Neurocomputing}, title = {Dynamic T-distributed stochastic neighbor graph convolutional networks for multi-modal contrastive fusion}, volume = {652}, year = {2025}, }
• (2025). A novel pain sentiment detection system utilizing a PainCapsule model and textual facial patterns. NEUCOM, 652, 130907. (www)

  @article{ , abstract = {Patient sentiment analysis establishes an intricate relationship between pain management and sentiment analysis in delivering high-quality medical care. This work presents an efficient pain sentiment recognition system within a smart healthcare framework designed to assess patients’ pain levels by analyzing their facial expressions. The proposed system is implemented in four distinct phases. First, facial regions are detected using efficient face-detection techniques. In the second phase, the extracted facial regions undergo feature computation using advancements in deep learning techniques, including end-to-end and pre-trained convolutional neural networks ( C N N ) to capture complex and discriminative facial features associated with pain emotions. In the third phase, a novel P a i n C a p s u l e model is introduced, which evaluates pain intensity by analyzing both macro- and microfacial expressions. This phase also employs attention networks, feature tuning, and transfer learning techniques to optimize the system’s performance. Finally, in the fourth phase, score fusion techniques are applied to the deep pain recognition models to enhance accuracy and robustness further. The system’s effectiveness is rigorously evaluated using two benchmark video datasets: the BioVid Heat Pain Dataset and the Multimodal Intensity Pain (MIntPAIN) database. Extensive experiments and comparative analysis with existing state-of-the-art methods reveal that the proposed system achieves an F1-score of 65.51% for BioVid and 58.31% for MIntPAIN datasets, outperforming other pain recognition systems, demonstrating its potential to advance pain sentiment recognition within smart healthcare frameworks.}, archive = {J_NEUCOM}, author = {Anay Ghosh and Saiyed Umer and Bibhas Chandra Dhara and Deepak Kumar Jain and Ranjeet Kumar Rout and Amir Hussain}, doi = {10.1016/j.neucom.2025.130907}, journal = {Neurocomputing}, month = {11}, pages = {130907}, shortjournal = {Neurocomputing}, title = {A novel pain sentiment detection system utilizing a PainCapsule model and textual facial patterns}, volume = {652}, year = {2025}, }
• (2025). M3DP: Optimizing 2D vision tasks with minimal 3D object information. NEUCOM, 652, 130905. (www)

  @article{ , abstract = {The availability of 3D data (such as LiDAR, Radar, and other point-cloud data) enriches the exploration of leveraging 3D priors for specific 2D downstream tasks. However, existing methods require aligned 3D data annotation for 2D data, which can be time-consuming and labor-intensive. To reduce this reliance, we propose Minimal 3D object information Prior (M3DP) for 2D vision feature learning using unaligned 3D data as prior. Specifically, M3DP only requires 3D objects and their corresponding classification labels within the same dataset to learn 3D priors, which greatly saves time and labor costs. Moreover, we introduce multiview rotation augmentation (MRA) and two alignments—K-Best-of-N alignment and instance alignment—to encourage 2D representation learning from 3D representations in the unified 2D–3D space. This approach helps the model learn geometric properties. This strategy fully leverages the multi-view geometric information of 3D objects, enabling precise localization and matching in 2D images. Through instance alignment, our method also efficiently facilitates information transfer across different categories of 3D objects. Our method effectively enhances the performance of 2D tasks by learning the geometric properties of 3D objects. Extensive experiments on three datasets demonstrate the superiority of our approach over prior state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Ziming Wang and Yanjing Li and Linlin Yang and Xinkai Liang and Xianbin Cao and Qi Wang and Baochang Zhang}, doi = {10.1016/j.neucom.2025.130905}, journal = {Neurocomputing}, month = {11}, pages = {130905}, shortjournal = {Neurocomputing}, title = {M3DP: Optimizing 2D vision tasks with minimal 3D object information}, volume = {652}, year = {2025}, }
• (2025). Generalized Cauchy–Schwarz divergence: Efficient estimation and applications in deep learning. NEUCOM, 652, 130904. (www)

  @article{ , abstract = {Divergence measures play a fundamental role in machine learning and deep learning; however, efficient methods for handling multiple distributions (i.e., more than two) remain largely underexplored. This challenge is particularly critical in scenarios where managing multiple distributions simultaneously is both necessary and unavoidable, such as clustering, multi-source domain adaptation, and multi-view learning. A common approach to quantifying overall divergence involves computing the mean pairwise distances between distributions. However, this method suffers from two key limitations. First, it is restricted to pairwise comparisons and fails to capture higher-order interactions or dependencies among three or more distributions. Second, its implementation requires a double-loop traversal over all distribution pairs, leading to significant computational overhead, particularly when dealing with a large number of distributions. In this study, we introduce the generalized Cauchy–Schwarz divergence (GCSD), a novel divergence measure specifically designed for multiple distributions. To facilitate its practical application, we propose a kernel-based closed-form sample estimator, which enables efficient computation in various deep-learning contexts. Furthermore, we validate GCSD through two representative tasks: deep clustering, achieved by maximizing the generalized divergence between clusters, and multi-source domain adaptation, achieved by minimizing the generalized discrepancy among feature distributions. Extensive experimental evaluations highlight the robustness and effectiveness of GCSD in both tasks, underscoring its potential to advance machine learning techniques that require the quantification of multiple distributions.}, archive = {J_NEUCOM}, author = {Mingfei Lu and Shujian Yu and Robert Jenssen and Badong Chen}, doi = {10.1016/j.neucom.2025.130904}, journal = {Neurocomputing}, month = {11}, pages = {130904}, shortjournal = {Neurocomputing}, title = {Generalized Cauchy–Schwarz divergence: Efficient estimation and applications in deep learning}, volume = {652}, year = {2025}, }
• (2025). A new robust semi-supervised clustering method based on adaptive credibility estimation and locality preservation. NEUCOM, 652, 130900. (www)

  @article{ , abstract = {Semi-supervised clustering guides the clustering process by incorporating various prior knowledge, yet many methods overlook the impact of mislabeling. Mislabeling may mislead model learning, affect parameter optimization, and result in overfitting, thereby degrading model performance. To solve the problem, this paper proposes a robust semi-supervised clustering method based on adaptive credibility estimation and locality preservation. Specifically, our method designs an adaptive credibility estimation strategy, which assesses the risk levels of labeled samples by calculating adaptive credibility using local consistency criterion, thereby selecting labeled samples with lower risk to guide the model learning process. Furthermore, by generating a discriminative projection space that maintains the structural integrity of data, our method enables accurate distance metric learning. Comprehensive experiments on a variety of publicly available datasets show that our method achieves the best performance compared with state-of-the-art baseline methods, demonstrating its effectiveness in semi-supervised clustering.}, archive = {J_NEUCOM}, author = {Hengdong Zhu and Rongjiao Liang and Baoshuo Kan and Enliang Yan and Fu Lee Wang and Tianyong Hao}, doi = {10.1016/j.neucom.2025.130900}, journal = {Neurocomputing}, month = {11}, pages = {130900}, shortjournal = {Neurocomputing}, title = {A new robust semi-supervised clustering method based on adaptive credibility estimation and locality preservation}, volume = {652}, year = {2025}, }
• (2025). Sparse and robust alternating direction method of multipliers for large-scale classification learning. NEUCOM, 652, 130893. (www)

  @article{ , abstract = {Support vector machine (SVM) is a highly effective method in terms of classification learning. Nonetheless, when faced with large-scale classification problems, the high computational complexity involved can pose a significant obstacle. To tackle this problem, we establish a new trimmed squared loss SVM model known as TSVM. This model can be designed for achieving both sparsity and robustness at the same time. A novel optimality theory has been developed for the nonsmooth and nonconvex TSVM. Utilizing this new theory, the innovative fast alternating direction method of multipliers with low computational complexity and working set has been proposed to solve TSVM. Numerical tests show the effectiveness of the new method regarding the computational speed, number of support vector and classification accuracy, outperforming eight alternative top solvers. As an illustration, when tackling the real dataset with more than 1 0 7 instances, compared to seven other algorithms, our algorithm exhibited a 34 times enhancement in computation time, alongside achieving a 6.5% enhancement in accuracy and a 25 times decrease in support vector rates.}, archive = {J_NEUCOM}, author = {Huajun Wang and Wenqian Li and Yuanhai Shao and Hongwei Zhang}, doi = {10.1016/j.neucom.2025.130893}, journal = {Neurocomputing}, month = {11}, pages = {130893}, shortjournal = {Neurocomputing}, title = {Sparse and robust alternating direction method of multipliers for large-scale classification learning}, volume = {652}, year = {2025}, }
• (2025). PhraseBT: A phrase-level back-translation data augmentation method for neural machine translation. NEUCOM, 652, 130832. (www)

  @article{ , abstract = {In the field of machine translation, parallel corpora from news, movie, and TV subtitles are often used, but acquiring these corpora is challenging. We began our exploration with Japanese song lyrics and their Chinese translations, collecting over 700K manually translated Japanese–Chinese parallel sentences by web crawling, expanding the parallel corpora for machine translation in this language pair, and providing researchers with novel lyrical data. To enhance the translation quality of this corpus, we proposed a Phrase-level back-translation data augmentation method (PhraseBT) for neural machine translation. Experiments across various language pairs have demonstrated that this method effectively improves machine translation performance, offering a new avenue for data augmentation in this field. Additionally, we combined this method with the Meta-Llama-3-8B-Instruct (Llama3-8B-Instruct) Large Language Model (LLM), utilizing the strengths of both techniques: the grammar correction and general language modeling capabilities of LLMs, and the focused enhancements provided by phrase-level back-translation. This hybrid approach further enhanced translation quality, showcasing cutting-edge practices in machine translation.}, archive = {J_NEUCOM}, author = {Jinyi Zhang and Haowei Li and Dashuai Deng and Yanshu Wang and Tadahiro Matsumoto}, doi = {10.1016/j.neucom.2025.130832}, journal = {Neurocomputing}, month = {11}, pages = {130832}, shortjournal = {Neurocomputing}, title = {PhraseBT: A phrase-level back-translation data augmentation method for neural machine translation}, volume = {652}, year = {2025}, }
• (2025). Semi-supervised few-shot learning using critical-descriptor prototypes. NEUCOM, 652, 130769. (www)

  @article{ , abstract = {Semi-supervised few-shot learning aims to exploit limited number of labeled samples and massive unlabeled samples to improve the model classification capability. One of the mainstream methods is to obtain class means as prototypes to characterize each class. Nonetheless, these mean-of-class prototypes can exhibit bias due to sparse or outlier samples. To mitigate this, we propose Critical-Descriptor Prototypes (CDP), a novel framework that refines prototype learning through dictionary-driven feature sparsification and supervised contrastive optimization. Specifically, CDP introduces two key innovations: firstly, Dictionary-based Descriptor Sparsification: Instead of directly averaging embeddings, CDP decomposes each prototype into 25 discriminative descriptors and reconstructs unlabeled samples via sparse coding. This forces the model to prioritize high-attribution local features by minimizing reconstruction errors. Secondly, Supervised Contrastive Learning: Leveraging pseudo-labels from CDP-refined prototypes, we design a supervised contrastive loss that explicitly clusters samples from the same class while dispersing inter-class features. Extensive experiments on mini-ImageNet and tiered-ImageNet demonstrate state-of-the-art results. Notably, CDP outperforms our prior work (SS-VPE) by 13.94% in 1-shot scenarios, validating its superiority in mitigating prototype bias through critical feature localization.}, archive = {J_NEUCOM}, author = {Yukun Liu and Hongsong Wan and Daming Shi}, doi = {10.1016/j.neucom.2025.130769}, journal = {Neurocomputing}, month = {11}, pages = {130769}, shortjournal = {Neurocomputing}, title = {Semi-supervised few-shot learning using critical-descriptor prototypes}, volume = {652}, year = {2025}, }
• (2025). Assessing feature scorer results on high-dimensional datasets with t-SNE. NEUCOM, 652, 130561. (www)

  @article{ , abstract = {While vast literature on high-dimensional data visualization is available, there are not many works regarding the visualization of feature scorers and their results. Feature scorers are algorithms that assign numerical importance to each feature of multi-dimensional datasets. These importance scores can be used in several applications, such as feature selection, knowledge discovery, and machine learning interpretability. There are several feature scorers to choose from, and often no single metric or ground truth is available to guarantee the quality of their results. In this scenario, visualization can become valuable to support the decision of which method to choose and how good its results are. For this goal, this work presents “weighted t-SNE.” It modifies the relationship between data points in the embedded 2D space to reflect the importance of each dimension of the original datasets as assessed by a feature scorer. This research discusses how to implement weighted t-SNE, proposes the silhouette coefficient as a numerical evaluation of the results, and shows several examples of its use in practice. Synthetic and real-world tabular datasets are used in the experiments together with nine feature scorers, ranging from Mutual Information to neural networks. Each feature scorer produces unique visualizations, and weighted t-SNE can be used to compare and choose the one that better suits a given dataset and task. Weighted t-SNE can also visually show the importance of features learned by machine learning models and help us see how they are organizing the data, increasing their interpretability.}, archive = {J_NEUCOM}, author = {Bruno Iochins Grisci and Mario Inostroza-Ponta and Márcio Dorn}, doi = {10.1016/j.neucom.2025.130561}, journal = {Neurocomputing}, month = {11}, pages = {130561}, shortjournal = {Neurocomputing}, title = {Assessing feature scorer results on high-dimensional datasets with t-SNE}, volume = {652}, year = {2025}, }
• (2025). UAV formation control based on ensemble reinforcement learning. NEUCOM, 651, 131056. (www)

  @article{ , abstract = {Based on the frameworks of Multi-Agent Deep Deterministic Policy Gradient (MADDPG) and Deep Deterministic Policy Gradient (DDPG) algorithms, this paper investigates the UAV formation control problem. To address the convergence difficulties inherent in multi-agent algorithms, curriculum reinforcement learning is applied during the training phase to decompose the task into incremental stages. A progressively hierarchical reward function tailored for each stage is designed, significantly reducing the training complexity of MADDPG. In the inference phase, an ensemble reinforcement learning strategy is adopted to enhance the accuracy of UAV formation control. When the UAVs approach their target positions, the control strategy switches from MADDPG to the DDPG algorithm, thus achieving more efficient and precise control. Through ablation and comparative experiments in a self-developed Software in the Loop (SITL) simulation environment, the effectiveness and stability of the ensemble reinforcement learning algorithm in multi-agent scenarios are validated. Finally, real-world experiments further verify the practical applicability of the proposed algorithm ( https://b23.tv/7ceLpLe ).}, archive = {J_NEUCOM}, author = {Kaifeng Wu and Lei Liu and Chengqing Liang and Lei Li}, doi = {10.1016/j.neucom.2025.131056}, journal = {Neurocomputing}, month = {10}, pages = {131056}, shortjournal = {Neurocomputing}, title = {UAV formation control based on ensemble reinforcement learning}, volume = {651}, year = {2025}, }
• (2025). Joint super-resolution and inverse tone-mapping: A feature decomposition aggregation network and a new benchmark. NEUCOM, 651, 131050. (www)

  @article{ , abstract = {Joint Super-Resolution and Inverse Tone-Mapping (joint SR-ITM) aims to increase the resolution and dynamic range of low-resolution and standard dynamic range images. Recent networks mainly resort to image decomposition techniques with complex multi-branch architectures. However, the fixed decomposition techniques would largely restrict their power on versatile images. To exploit the potential power of decomposition mechanism, in this paper, we generalize it from the image domain to the broader feature domain. To this end, we propose a lightweight Feature Decomposition Aggregation Network (FDAN). In particular, we design a Feature Decomposition Block (FDB) to achieve learnable separation of detail and base feature maps, and develop a Hierarchical Feature Decomposition Group by cascading FDBs for powerful multi-level feature decomposition. Moreover, for better evaluation, we collect a large-scale dataset for joint SR-ITM, i.e. , SRITM-4K, which provides versatile scenarios for robust model training and evaluation. Experimental results on two benchmark datasets demonstrate that our FDAN is efficient and outperforms state-of-the-art methods on joint SR-ITM. The code of our FDAN and the SRITM-4K dataset are available at https://github.com/CS-GangXu/FDAN .}, archive = {J_NEUCOM}, author = {Gang Xu and Ao Shen and Yuchen Yang and Xiantong Zhen and Wei Chen and Jun Xu}, doi = {10.1016/j.neucom.2025.131050}, journal = {Neurocomputing}, month = {10}, pages = {131050}, shortjournal = {Neurocomputing}, title = {Joint super-resolution and inverse tone-mapping: A feature decomposition aggregation network and a new benchmark}, volume = {651}, year = {2025}, }
• (2025). Focus DETR: Focus detection transformer for ship wall-climbing robot real-time object detection. NEUCOM, 651, 131049. (www)

  @article{ , abstract = {With the growing use of wall-climbing robots for ship paint removal in repair yards, real-time detection of the targets of the wall-climbing robots and accurate positioning of them has become a very important task. Due to the changes in the external environment and the diversity of working postures of the wall-climbing robot, maintaining stable detection accuracy and real-time performance is a challenging task. To address this issue, a focus detection transformer (Focus-DETR) architecture is proposed. Firstly, the spatial attention recursive gated convolution(Sa-Gn) module is employed in the final stage of the backbone to extract high-level features, not only gaining high accuracy but also maintaining real-time speed. Secondly, to improve the ability to extract key features of the wall-climbing robot object, a hybrid encoder method is introduced to integrate features from adjacent stages of the neck. What’s more, the focal EIOU loss function in the detection head optimizes the width and height errors of the detection box and adjusts its loss weight. It not only improves the alignment accuracy of the detection center point of the object bounding box of the wall-climbing robot, but also is suitable for lightweight application deployment. Experimental results show that compared with RT-DETR-R50, Focus DETR-S improves the detection mAP by 6.0 % on the wall-climbing robot dataset, with the inference speed very close to that of RT-DETR-R50. It has also achieved the same performance improvement on the UAVDT dataset.}, archive = {J_NEUCOM}, author = {Pan Li and Xiaofang Yuan and Haozhi Xu and Jinlei Wang and Yaonan Wang}, doi = {10.1016/j.neucom.2025.131049}, journal = {Neurocomputing}, month = {10}, pages = {131049}, shortjournal = {Neurocomputing}, title = {Focus DETR: Focus detection transformer for ship wall-climbing robot real-time object detection}, volume = {651}, year = {2025}, }
• (2025). A fault-tolerant model predictive control approach based on deep operator network. NEUCOM, 651, 131048. (www)

  @article{ , abstract = {This paper is concerned with the fault-tolerant control problem for unknown nonlinear systems in the model predictive control (MPC) framework. A modified deep operator network is designed to learn system dynamics from input-output data. However, the input data that contain fault information cannot be acquired directly in the presence of actuator faults. To overcome this difficulty, a mode simulation method is presented via adequate and uniform sampling of virtual fault information in a hyperspace. In this way, the system responses in different faulty modes are simulated to ensure excellent prediction accuracy of the modified network. Moreover, an improved fault estimation method is designed with historical input-output data of the modified network. Then, based on the fault estimates, the design problem of the fault-tolerant MPC controller is converted into a constrained optimization problem, which is further solved using an adaptive gradient descent method. Finally, two simulation experiments are provided to illustrate the validity of the proposed approach.}, archive = {J_NEUCOM}, author = {Yun-He Zhang and Xiao-Jian Li}, doi = {10.1016/j.neucom.2025.131048}, journal = {Neurocomputing}, month = {10}, pages = {131048}, shortjournal = {Neurocomputing}, title = {A fault-tolerant model predictive control approach based on deep operator network}, volume = {651}, year = {2025}, }
• (2025). Enhancing syntactic and semantic features via TextGINConv and Kolmogorov–Arnold networks for aspect-based sentiment analysis. NEUCOM, 651, 131037. (www)

  @article{ , abstract = {Aspect-based sentiment analysis identifies the sentiment polarity of a given aspect in a sentence. Recent advances have demonstrated the effectiveness of combining syntactic dependency structures with graph convolutional networks. However, traditional graph convolutional networks usually have a simpler message passing mechanism that may describe the relationship between nodes only through the adjacency matrix, often ignoring messages passed by edge features and performing sentiment analysis without considering the specific semantic information of different aspects. In this paper, we innovatively propose a syntactic and semantic enhancement network model for aspect-based sentiment analysis (ESSGKA). To be specific, we combine a self-attention mechanism with an aspect-oriented attention mechanism, enabling the simultaneous learning of aspect-related semantics and the overall semantics of the sentence. TextGINConv focuses more on edge features, which are leveraged to facilitate dense message passing. To enhance the fusion of two distinct feature messages, we propose a novel gated fusion framework based on Kolmogorov–Arnold networks. Extensive experiments on four publicly available datasets show that our model is more competitive than state-of-the-art models.}, archive = {J_NEUCOM}, author = {Xiaoru Li and Yuxia Lei}, doi = {10.1016/j.neucom.2025.131037}, journal = {Neurocomputing}, month = {10}, pages = {131037}, shortjournal = {Neurocomputing}, title = {Enhancing syntactic and semantic features via TextGINConv and Kolmogorov–Arnold networks for aspect-based sentiment analysis}, volume = {651}, year = {2025}, }
• (2025). SCR: A completion-then-reasoning framework for multi-hop question answering over incomplete knowledge graph. NEUCOM, 651, 131027. (www)

  @article{ , abstract = {Reinforcement learning has become the widely adopted technique for multi-hop knowledge graph question answering task thanks to its excellent interpretability in reasoning process. However, it is severely affected by the incompleteness of knowledge graphs and the sparse rewards caused by weak supervision. In this paper, we propose a completion-then-reasoning framework, called SCR, to address these two issues. For the incompleteness of knowledge graphs, we first extract a subgraph from the given knowledge graph for a given question, and use the knowledge graph embedding model to predict and complete missing triples, followed by reinforcement learning for answer reasoning on the completed subgraph. To alleviate the sparse rewards in reinforcement learning, we introduce a semantic reward based on the semantic similarity between original question and full relational path, enabling the model to receive partial rewards for partially correct paths instead of a zero reward. Detailed experiments on PQ, PQL, MetaQA, and WebQSP datasets demonstrate that our SCR model effectively improves the performance of multi-hop knowledge graph question answering task. Particularly, under sparse KG setting, SCR model outperforms baselines by a large margin, highlighting the effectiveness of completion-then-reasoning framework in mitigating the incompleteness of knowledge graphs.}, archive = {J_NEUCOM}, author = {Ridong Han and Jia Liu and Haijia Bi and Tao Peng and Lu Liu}, doi = {10.1016/j.neucom.2025.131027}, journal = {Neurocomputing}, month = {10}, pages = {131027}, shortjournal = {Neurocomputing}, title = {SCR: A completion-then-reasoning framework for multi-hop question answering over incomplete knowledge graph}, volume = {651}, year = {2025}, }
• (2025). Safe reinforcement learning for discrete-time nonlinear zero-sum games with unknown state constraints and asymmetric input constraints. NEUCOM, 651, 131026. (www)

  @article{ , abstract = {In this paper, we propose a novel safe reinforcement learning (RL) algorithm for discrete-time nonlinear zero-sum games with unknown state constraints and asymmetric input constraints. To address this constrained optimal problem, we adopt a value iteration framework based on neural networks, incorporating a critic-only structure. Given the unknown safety constraints, we tackle the state constraint issue by introducing a neural network-based control barrier function (CBF) using collected data to augment the reward function. Furthermore, by leveraging the non-monotonic increasing property of the value function, we ensure the system’s safety. Additionally, we construct a non-quadratic function to further augment the reward function, thereby satisfying the asymmetric input constraints. This paper also includes a series of theoretical proofs that rigorously demonstrate the convergence and safety of the proposed algorithm. Finally, experiments conducted under different scenarios and parameter settings, compared with existing algorithms, validate the algorithm’s effectiveness and safety.}, archive = {J_NEUCOM}, author = {Shihan Liu and Zhi Chen and Dongxu Gao}, doi = {10.1016/j.neucom.2025.131026}, journal = {Neurocomputing}, month = {10}, pages = {131026}, shortjournal = {Neurocomputing}, title = {Safe reinforcement learning for discrete-time nonlinear zero-sum games with unknown state constraints and asymmetric input constraints}, volume = {651}, year = {2025}, }
• (2025). MarIns3D: An open-vocabulary 3D instance segmentation model with mask refinement. NEUCOM, 651, 131018. (www)

  @article{ , abstract = {Open-vocabulary 3D instance segmentation has gained significant attention due to its potential role in scene perception. Existing methods typically involve two stages: generating class-agnostic 3D instance masks using segmentation models, followed by semantic classification of these masks. However, poor classification performance often stems from low-quality masks in the first stage. This paper proposes two key components to optimize the mask generation process: a dynamic offset module and a projection consistency loss. By dynamically adjusting sampling point positions, query points can capture key scene features to generate high-quality masks. Then the projection consistency loss compares these 3D instance masks with ground truth in 2D projections to refine them, improving segmentation accuracy. Experimental results on the ScanNetV2 validation set show that MarIns3D outperforms SOLE on zero-shot segmentation, with a 1.8 % and 1.7 % improvement in AP25 and AP50, respectively, and also demonstrates enhanced open-set segmentation capabilities. These results confirm our model’s superior mask quality and segmentation performance. Furthermore, ablation studies verify that the synergy between the dynamic offset module and the projection consistency loss is crucial for these enhancements.}, archive = {J_NEUCOM}, author = {Haiyang Li and Jinhe Su and Dong Zhou and Mengyun Cao}, doi = {10.1016/j.neucom.2025.131018}, journal = {Neurocomputing}, month = {10}, pages = {131018}, shortjournal = {Neurocomputing}, title = {MarIns3D: An open-vocabulary 3D instance segmentation model with mask refinement}, volume = {651}, year = {2025}, }
• (2025). A comprehensive review of deep learning techniques for interaction-aware trajectory prediction in urban autonomous driving. NEUCOM, 651, 131014. (www)

  @article{ , abstract = {Autonomous vehicles can improve urban transport by using multiple components that accurately represent their surroundings and improve decision-making processes. One essential component is trajectory prediction, which estimates the future states of traffic participants and anticipates hazardous scenarios. There are different approaches for trajectory prediction, in which Intention-aware and Interaction-aware approaches represent the state-of-the-art since they involve better representation of the surroundings. This paper reviews the literature on Interaction-Aware Trajectory Prediction for autonomous vehicles. It explores how incorporating maneuver intentions and interactions can improve prediction accuracy, and it examines the techniques and datasets employed in this field.}, archive = {J_NEUCOM}, author = {Iago Pachêco Gomes and Denis Fernando Wolf}, doi = {10.1016/j.neucom.2025.131014}, journal = {Neurocomputing}, month = {10}, pages = {131014}, shortjournal = {Neurocomputing}, title = {A comprehensive review of deep learning techniques for interaction-aware trajectory prediction in urban autonomous driving}, volume = {651}, year = {2025}, }
• (2025). Transduction to induction: Unsupervised representation learning based on rank information. NEUCOM, 651, 131010. (www)

  @article{ , abstract = {The use of deep learning in supervised scenarios has become well-established. However, there is growing interest in exploring unsupervised learning methods. Transductive approaches are promising for learning rich contextual relationships in unsupervised scenarios but face challenges when dealing with large amounts of data. The main motivation of this study is to investigate the feasibility of an inductive model, based on a neural network, learning representations from ranked lists generated by transductive methods in unsupervised scenarios. We propose an unsupervised approach called Inductive Ranking Learning (IRL), which leverages techniques to learn similarities and dissimilarities from pairs derived from ranked lists produced by transductive methods. This technique involves weighting the most relevant and irrelevant elements when calculating the error of likely positive and negative pairs, based on the position of the element in the ranked list relative to its pair. This allows learning without the need for labels. The proposed approach enables the use of transductive techniques to train inductive models, promoting generalization to unseen data, which is particularly important in scenarios where new data is constantly being introduced. Experimental results show promising performance, although the method may face challenges when dealing with ranked lists derived from large datasets. Overall, the proposed approach offers significant potential for both unsupervised learning and the exploration of transductive approaches in inductive models.}, archive = {J_NEUCOM}, author = {Deryk Willyan Biotto and Lucas Pascotti Valem and Daniel Carlos Guimarães Pedronette and Denis Henrique Pinheiro Salvadeo}, doi = {10.1016/j.neucom.2025.131010}, journal = {Neurocomputing}, month = {10}, pages = {131010}, shortjournal = {Neurocomputing}, title = {Transduction to induction: Unsupervised representation learning based on rank information}, volume = {651}, year = {2025}, }
• (2025). Learning deep generative models based on binomial log-likelihood. NEUCOM, 651, 131009. (www)

  @article{ , abstract = {Likelihood-based learning algorithms for deep generative models mostly use the Gaussian log-likelihood. One notable exception is the binomial log-likelihood used in the Wasserstein autoencoder; however, it is not commonly used in practice because it does not generalize well. In this paper, we reconsider the binomial log-likelihood for learning deep generative models and study its theoretical properties. We propose two modifications to the original binomial log-likelihood and derive the convergence rates of the corresponding maximum likelihood estimators. These theoretical results explain why the original binomial log-likelihood performs poorly. In addition, motivated by the modified binomial log-likelihood, we propose a parametric heterogeneous Gaussian log-likelihood, which is novel in learning deep generative models. By analyzing various benchmark image datasets, we show that the proposed parametric heterogeneous Gaussian log-likelihood outperforms the standard homogeneous Gaussian log-likelihood. Additionally, we provide several pieces of evidence to explain why the proposed heterogeneous Gaussian log-likelihood works better than others.}, archive = {J_NEUCOM}, author = {Hwichang Jeong and Insung Kong and Yongdai Kim}, doi = {10.1016/j.neucom.2025.131009}, journal = {Neurocomputing}, month = {10}, pages = {131009}, shortjournal = {Neurocomputing}, title = {Learning deep generative models based on binomial log-likelihood}, volume = {651}, year = {2025}, }
• (2025). Multi spectral visible-thermal IR image translation using improved u-net & conditional diffusion. NEUCOM, 651, 131006. (www)

  @article{ , abstract = {Translating images from visible spectrum to thermal IR (TIR) domain to achieve precise and realistic representations of TIR images is a challenging task. Thermal infrared imaging is of great significance in scenarios where vision is severely impaired especially in difficult lighting conditions such as night, haze, fog or cloudy weather. With these advantages, infrared imaging finds extensive applicability in navigation, surveillance, object detection, product inspection, agriculture as well as remote sensing. In order to build high performance deep models for such wide range of applications, it is necessary to have large amount of TIR data for training. However, there is unavailability of sufficient IR based datasets due to high cost of thermal infrared camera setups. While large number of visible image datasets are available, this scarcity of TIR datasets can be addressed by translating visible images to their TIR counterparts. In this paper, we leverage the widely available visible range data to propose two visible to TIR domain translation approaches, one is modified U-Net based non-generative approach called TIR-UNet and the other is conditional diffusion based generative approach that also uses U-Net as neural backbone for synthesizing TIR images. Both the proposed methods have been evaluated on four benchmark datasets and demonstrate high qualitative as well as quantitative performance in generating perceptually realistic, visually plausible and high quality TIR equivalents of given visible images. Compared to state-of-the-art methods which include U-Net and powerful GAN variants, our methods achieve remarkable performance increase on the metrics of MSE, PSNR and SSIM for both day and night images.}, archive = {J_NEUCOM}, author = {Mahroosh Banday and Brejesh Lall}, doi = {10.1016/j.neucom.2025.131006}, journal = {Neurocomputing}, month = {10}, pages = {131006}, shortjournal = {Neurocomputing}, title = {Multi spectral visible-thermal IR image translation using improved u-net & conditional diffusion}, volume = {651}, year = {2025}, }
• (2025). Adaptive memory fusion for multi-frame optical flow estimation. NEUCOM, 651, 131005. (www)

  @article{ , abstract = {Multi-frame optical flow models have demonstrated superior performance compared to two-frame models. Nevertheless, they suffer from several drawbacks, such as the absence of optical flow for the initial frames, the sharp increase in computational load as the number of input frames increases, and performance fluctuations when ≥ 3 frames are incorporated. To address these issues, we propose AMFFlow, based on the Adaptive Memory Fusion (AMF) module, which effectively utilizes multi-frame information through memory fusion, thereby overcoming the aforementioned limitations. By employing a memory fusion mechanism, our method efficiently utilizes multi-frame information to estimate more accurate optical flow while maintaining a constant computational cost per frame. To our knowledge, AMFFlow is a pioneering model that can effectively exploit multi-frame information using ≥ 5 frames for optical flow estimation. The AMF module is characterized by its low latency and plug-and-play capability. Additionally, we propose the Proportionally Debiased Endpoint Error (PDEPE), a novel metric designed to address evaluation errors caused by dataset bias in multi-frame optical flow models. Extensive experimental results demonstrate that AMFFlow achieves state-of-the-art (SOTA) generalization performance. Compared to the competing MemFlow model, the performance improvements on Sintel Training Clean and Sintel Training Final are approximately 5.9 % and 3.9 %, respectively. The source code and pre-trained models will be made publicly available at https://github.com/keacifer/AMFFlow .}, archive = {J_NEUCOM}, author = {Yibo Cui and Shangsheng Li and Xin Yang and Gang Wang and Yizheng Wang}, doi = {10.1016/j.neucom.2025.131005}, journal = {Neurocomputing}, month = {10}, pages = {131005}, shortjournal = {Neurocomputing}, title = {Adaptive memory fusion for multi-frame optical flow estimation}, volume = {651}, year = {2025}, }
• (2025). Predicting occluded skeletal joints via tracking-based feature extraction. NEUCOM, 651, 131004. (www)

  @article{ , abstract = {A rapidly growing research area in computer vision is the recognition of human poses, which imposes strong occlusion problems in subsequent tracking. Therefore, tracking mechanisms should contain algorithms for handling occlusions such that the skeletal data is continuous between frames. This paper introduces a new skeletal tracking approach where a deep-learning-based model for a Skeleton Feature Extractor is embedded into the tracking algorithm. This differs from the common tracking methods that, in the process of image feature extraction, consider feature extraction at joints and the spatial relation between them, thus making occlusion scenarios highly detectable. We further make a comparison with our model and MotioNet, which is a broadly applied model for 3D motion reconstruction. MotioNet can interpolate the missing joints based on the information both spatial and temporal. It, however, does not work when actual joints are occluded for some frames. Our model predicts the skeletal joints that are missing. Experiments on the JHMDB and Penn_Action dataset were meant to show that the method improves the accuracy of forecasting occluded skeletal joint positions by the same PCK metric.}, archive = {J_NEUCOM}, author = {Khac-Anh Phu and Van-Dung Hoang and Van-Tuong-Lan Le}, doi = {10.1016/j.neucom.2025.131004}, journal = {Neurocomputing}, month = {10}, pages = {131004}, shortjournal = {Neurocomputing}, title = {Predicting occluded skeletal joints via tracking-based feature extraction}, volume = {651}, year = {2025}, }
• (2025). Two-stage real-world image dehazing method using physics-based dehazing network and contrastive learning generative adversarial network. NEUCOM, 651, 131002. (www)

  @article{ , abstract = {Real-world image dehazing remains a challenging task due to the ill-posed nature of haze formation and the significant domain gap between synthetic and real foggy scenes. In this paper, a novel two-stage framework is proposed, integrating a Physics-Based Dehazing Network (PBDNet) with a Contrastive Learning-based Generative Adversarial Network (CLGAN). In the first stage, PBDNet is trained on synthetic hazy-clean pairs using the atmospheric scattering model, extracting interpretable and transferable physical priors. In the second stage, CLGAN leverages these priors to guide unpaired image translation between real hazy and clean images. The integration of contrastive learning further enhances the alignment of fog-invariant representations, improving dehazing stability and generalization. Extensive experiments demonstrate the effectiveness of our approach. On the SOTS-outdoor dataset, our method achieves a PSNR of 34.13 dB and SSIM of 0.9863, surpassing state-of-the-art methods. On the real-world RTTS dataset, it achieves a BRISQUE score of 17.54, indicating superior perceptual quality. Additional evaluations using FADE metrics and object detection tasks confirm the practical value of our method in real-world scenarios. These results validate the effectiveness of combining physics-based priors with contrastive learning for robust real-world dehazing.}, archive = {J_NEUCOM}, author = {Huaqiang Xie and Kangwei Wang and Li Zhu and Jie Xie and Cheng Wu and Jie Sheng and Jin Zhang}, doi = {10.1016/j.neucom.2025.131002}, journal = {Neurocomputing}, month = {10}, pages = {131002}, shortjournal = {Neurocomputing}, title = {Two-stage real-world image dehazing method using physics-based dehazing network and contrastive learning generative adversarial network}, volume = {651}, year = {2025}, }
• (2025). Harmony score-guided inpainting: Iterative refinement for seamless image inpainting. NEUCOM, 651, 131001. (www)

  @article{ , abstract = {Inpainting techniques often demand extensive model fine-tuning or the concatenation of latent vectors, which can be time-intensive and prone to overfitting. Such methods frequently lead to inconsistencies between the inpainted regions and the surrounding background, occasionally producing partially satisfactory results where some areas appear natural while others remain unrealistic. To address these limitations, we demonstrate that existing inpainting methods can sufficiently handle certain scenarios but may struggle with specific problematic patches. We propose an iterative enhancement approach guided by an Inpainting Harmony Score, which evaluates the coherence of the inpainted image. Our method selectively enhances only the poorly reconstructed patches, preserving their masks for subsequent inpainting iterations. The process is repeated, followed by a final blending step to ensure seamless integration between the inpainted region and the background. This approach improves the overall quality and consistency of inpainting results while minimizing the risks of overfitting and inefficiency.}, archive = {J_NEUCOM}, author = {Haixin Wang and Jian Yang and Jinjia Zhou}, doi = {10.1016/j.neucom.2025.131001}, journal = {Neurocomputing}, month = {10}, pages = {131001}, shortjournal = {Neurocomputing}, title = {Harmony score-guided inpainting: Iterative refinement for seamless image inpainting}, volume = {651}, year = {2025}, }
• (2025). PLMQ: Piecewise linear mixed-precision quantization for deep neural networks. NEUCOM, 651, 131000. (www)

  @article{ , abstract = {Deep neural networks have demonstrated significant success across a wide range of applications, yet their high storages and computational requirements pose challenges for real-world deployment. While quantization techniques can mitigate these issues by reducing model size and improving efficiency, existing methods typically require access to the full training dataset and extensive fine-tuning, limiting their practical applicability. Post-training quantization (PTQ) alleviates these constraints but usually comes at the expense of accuracy, especially at low bit-widths. We introduce Piecewise Linear Mixed-Precision Quantization (PLMQ), a novel PTQ method that segments weight values into distinct regions, ensuring fine-grained precision in densely distributed regions while preserving appropriate quantization precision in sparse regions. By leveraging the Fisher Information Matrix for layer-wise sensitivity analysis and employing a grouped parameter sharing scheme, we effectively reduce storage overhead while maintaining accuracy. Experiments conducted on architectures such as ResNet, MobileNet, and Transformer models demonstrate that PLMQ achieves superior performance even at ultra-low bit-widths, rivaling certain partially quantization-aware training methods. This provides an effective solution for deploying models on resource-constrained hardware.}, archive = {J_NEUCOM}, author = {Chao Feng and Guoxu Zhou and Xun Wu and Qibin Zhao}, doi = {10.1016/j.neucom.2025.131000}, journal = {Neurocomputing}, month = {10}, pages = {131000}, shortjournal = {Neurocomputing}, title = {PLMQ: Piecewise linear mixed-precision quantization for deep neural networks}, volume = {651}, year = {2025}, }
• (2025). CTMEG: A continuous-time medical event generation model for clinical prediction of long-term disease progression. NEUCOM, 651, 130999. (www)

  @article{ , abstract = {Long-term health monitoring indicates patient’s disease progression which is critical in improving the quality of patient life and physician’s decision-making. Predictive models based on Electronic Health Records (EHRs) can offer substantial clinical support by alerting subsequent disease-associated adverse events. Effective disease progression modeling involves two subtasks: 1) estimation of disease-associated event occurrence times 2) classification of occurred event types Recent time-aware disease predictive models, mainly based on recurrent neural networks or attention networks, specialize in future disease type prediction by accounting for the temporal irregularities in EHRs. This paper focuses on multi-step continuous-time disease prediction, which is more challenging as predictive models can easily fall into task conflicts between subtasks. We propose a multi-task disentangled Continuous-Time Medical Event Generation (CTMEG) model to simultaneously tackle the two subtasks. Unlike conventional continuous-time models, CTMEG encodes multi-view historical medical events and then simultaneously predicts multi-step disease types and occurrence times. First, a discrete Conditional Intensity Function (CIF) is designed to better estimate the disease occurrence time with limited available data. Second, to reduce task conflicts, a gated network is proposed to disentangle the rough patient representation into task-specific representations. Finally, we utilize a tailored CIF attention module to reduce error accumulation during the prediction process. Extensive experiments on the eICU and BFH databases demonstrate that the proposed CTMEG outperforms twelve competing models in long-term disease progression prediction. Our codes are available on github. 2}, archive = {J_NEUCOM}, author = {Mengxuan Sun and Xuebing Yang and Jiayi Geng and Jinghao Niu and Chutong Wang and Chang Cui and Xiuyuan Chen and Wen Tang and Wensheng Zhang}, doi = {10.1016/j.neucom.2025.130999}, journal = {Neurocomputing}, month = {10}, pages = {130999}, shortjournal = {Neurocomputing}, title = {CTMEG: A continuous-time medical event generation model for clinical prediction of long-term disease progression}, volume = {651}, year = {2025}, }
• (2025). Lightweight memory-driven self-attention for hyperspectral image classification with CNN-transformer cross-feature fusion. NEUCOM, 651, 130998. (www)

  @article{ , abstract = {Recent advancements in deep learning for hyperspectral image (HSI) classification have primarily focused on convolutional neural network (CNN), Transformer, and hybrid CNN-Transformer methods. However, these methods have several notable limitations: (1) Transformers struggle to capture the complex dependencies between target pixels and their neighboring pixels, and (2) the direct fusion strategies employed in current hybrid CNN-Transformer models significantly hinder the ability of CNN and Transformer to fully leverage their respective strengths. Moreover, the excessive complexity of some models leads to a substantial increase in network parameters. To address these challenges, we propose a lightweight memory-driven self-attention hybrid CNN-Transformer network (MDSAnet). MDSAnet comprises four key components: spectral feature joint memory-driven self-attention integration (SJMA), spectral-spatial convolutional multiplicative attention (S 2 CMA), cross-feature channel attention (CFCA), and dual-branch feature integration (DBFI) module. Specifically, SJMA enhances the global feature extraction capability of self-attention by incorporating the memory mechanism of long short-term memory networks, while S 2 CMA effectively captures local spatial-spectral features. Furthermore, CFCA and DBFI efficiently integrate the extracted features, ensuring more effective and robust representation learning. Experimental results on four publicly available HSI datasets demonstrate that the proposed MDSAnet achieves superior classification accuracy and computational efficiency while maintaining a lightweight model architecture, outperforming state-of-the-art methods. The code will be available at https://github.com/JiahongLiang477/MDSAnet .}, archive = {J_NEUCOM}, author = {Jiahong Liang and Yi Gao}, doi = {10.1016/j.neucom.2025.130998}, journal = {Neurocomputing}, month = {10}, pages = {130998}, shortjournal = {Neurocomputing}, title = {Lightweight memory-driven self-attention for hyperspectral image classification with CNN-transformer cross-feature fusion}, volume = {651}, year = {2025}, }
• (2025). Beyond decoders: Learning prompt-aware features for few-shot object counting. NEUCOM, 651, 130997. (www)

  @article{ , abstract = {Few-shot object counting involves estimating the quantity of objects from an arbitrary category in an image, given a few exemplars as visual prompts. This is typically achieved by matching image and exemplar features to establish a class-agnostic similarity map, which is used to regress a density map for the target class. Prevailing approaches primarily focus on improving the matching phase, designing various intricate decoders to perform sophisticated feature correlation. However, these methods still face challenges when initial features lack discriminative power. In this work, we shift our focus from decoder design to learning discriminative prompt-aware image features, enabling more effective similarity matching and density estimation. Specifically, we first establish a straightforward baseline that leverages a transformer-based backbone to enable direct interactions between images and exemplars. To ensure effective feature learning given limited exemplars, we further introduce a class-relevant prompts guided prediction module, which enhances the backbone’s ability to learn discriminative features by incorporating class-relevant visual cues and auxiliary training objectives. This module is designed to be auxiliary and can be discarded at inference, ensuring no additional computational overhead. Extensive experiments on FSC147 and CARPK demonstrate the effectiveness of our method, highlighting the efficacy of learning prompt-aware feature representations for few-shot counting.}, archive = {J_NEUCOM}, author = {Muming Zhao and Guang Li and Piotr Koniusz and Chongyang Zhang and Yongshun Gong}, doi = {10.1016/j.neucom.2025.130997}, journal = {Neurocomputing}, month = {10}, pages = {130997}, shortjournal = {Neurocomputing}, title = {Beyond decoders: Learning prompt-aware features for few-shot object counting}, volume = {651}, year = {2025}, }
• (2025). SMT-DL: A semi-supervised multi-task learning framework based on dictionary learning for robust feature sharing. NEUCOM, 651, 130996. (www)

  @article{ , abstract = {Multi-task learning (MTL) leverages shared representations across related tasks, facilitating the utilization of latent data information and enhancing classification performance. In complex learning scenarios, two major challenges frequently arise: limited labeled data and inefficient cross-task knowledge transfer. To address these issues, we propose a semi-supervised multi-task learning (SMT-DL) method based on dictionary learning. Specifically, we establish a dual dictionary architecture by innovatively combining dictionary learning with a multi-task coordination mechanism: (1) a semi-supervised synthetic dictionary that jointly reconstructs labeled and unlabeled data to capture potential cross-task features, and (2) an analytical dictionary that aligns sparse representations with discriminative decision boundaries. The framework incorporates three technical innovations: a block sparse regularization scheme that enforces feature sharing across tasks, a dual-space reconstruction mechanism that separates task-specific and shared representations, and a cross-task support vector synchronization strategy. In addition, we rigorously demonstrate the convergence of the proposed optimization algorithm. Extensive experimental results validate that the proposed SMT-DL approach outperforms existing methods in terms of robustness and classification performance.}, archive = {J_NEUCOM}, author = {Bo Liu and Boxu Zhou and Yanshan Xiao and Zhitong Wang and Baoqing Li and Shengxin He and Chenlong Ye and Fan Cao}, doi = {10.1016/j.neucom.2025.130996}, journal = {Neurocomputing}, month = {10}, pages = {130996}, shortjournal = {Neurocomputing}, title = {SMT-DL: A semi-supervised multi-task learning framework based on dictionary learning for robust feature sharing}, volume = {651}, year = {2025}, }
• (2025). ESC-DRKD: Enhanced skip connection-based direct reverse knowledge distillation for medical image anomaly detection. NEUCOM, 651, 130994. (www)

  @article{ , abstract = {Image anomaly detection has emerged as a prevalent research area in medical diagnosis, focusing on using models trained exclusively on normal samples to identify and locate anomalous images during testing. While many image anomaly detection (AD) models exhibit remarkable performance on industrial datasets, they frequently struggle to effectively detect anomalies in medical datasets characterized by complex distributions. This is because most methods are prone to overgeneralization, resulting in ineffective restoration of normal image features. To tackle this challenge, we introduce a novel approach called Enhanced Skip Connection-Based Direct Reverse Knowledge Distillation (ESC-DRKD), specifically designed to facilitate anomaly detection and localization in medical images. ESC-DRKD consists of pre-trained teacher encoders, trainable projection layers, and student decoders. Firstly, pseudo abnormal images are generated using the CutPaste method. By extracting multi-scale features from both normal and pseudo abnormal images through the pretrained teacher encoders, projection layers are employed to project the features of pseudo abnormal images onto important information of normal features. These projected features are then added to the outputs of corresponding student decoders at each level through skip connections to enhance the restoration of normal features. Furthermore, the output of the final layer of the teacher encoders serves as the input to the student decoders, thereby preventing the loss of normal information. Extensive experiments on various public medical datasets are tested the effectiveness of our proposed method. The results demonstrate that ESC-DRKD outperforms the state-of-the-art AD models on the five medical datasets, achieving an average improvement of over 3.0 % AUROC for anomaly detection. Code is available at https://github.com/GE-123-cpu/ESC_DRKD}, archive = {J_NEUCOM}, author = {Chenkun Ge and Xiaojun Yu and Hao Zheng and Zeming Fan and Umair Muhammad and Jinna Chen and Perry Ping Shum}, doi = {10.1016/j.neucom.2025.130994}, journal = {Neurocomputing}, month = {10}, pages = {130994}, shortjournal = {Neurocomputing}, title = {ESC-DRKD: Enhanced skip connection-based direct reverse knowledge distillation for medical image anomaly detection}, volume = {651}, year = {2025}, }
• (2025). DynamoSort: Using machine learning approaches for the automatic classification of seizure dynamotypes. NEUCOM, 651, 130992. (www)

  @article{ , abstract = {Epilepsy is characterised by unprovoked and recurring seizures, which can be electrically measured using electroencephalograms (EEG). To better understand the underlying mechanisms of seizures, researchers are exploring their temporal dynamics through the lens of dynamical systems modelling. Seizure initiation and termination patterns of spiking amplitude and frequency can be sorted into “dynamotypes”, which may be able to serve as biomarkers for intervention. However, manual classification of these dynamotypes requires trained raters and is prone to variability. To address this, we developed DynamoSort, a machine-learning algorithm for automatic seizure onset and offset classification. Dynamotype classification of real EEG data lacks a definitive ground truth, with mean inter-rater agreement at 73.4 % for onset and 64.2 % for offset types. Despite this, DynamoSort achieved a mean area under the curve (AUC) of 0.81 for onset and a mean AUC of 0.75 for offset types. Machine-human agreement was not significantly different from human-to-human agreement. To address the lack of ground truth in ratings, DynamoSort assigns probabilistic scores (-20−20), to indicate similarity to each seizure dynamotype based on spiking features, allowing for a characterization of seizure dynamics on a spectrum rather than the traditional qualitative taxonomy. DynamoSort is a straightforward, open-access tool that uses automatic labelling and probabilistic scoring to quantify subtle changes in seizure onset and offset dynamics.}, archive = {J_NEUCOM}, author = {Josh Wooley and Ashley Zachery-Savella and Michelle Le and Sally Y. Scofield and Kishore Jay and Josh Mosse-Robinson and Peter J. West and Karen S. Wilcox and Daria Nesterovich Anderson}, doi = {10.1016/j.neucom.2025.130992}, journal = {Neurocomputing}, month = {10}, pages = {130992}, shortjournal = {Neurocomputing}, title = {DynamoSort: Using machine learning approaches for the automatic classification of seizure dynamotypes}, volume = {651}, year = {2025}, }
• (2025). Multi-level contrastive reasoning over knowledge graphs for fault diagnosis of confounding cases. NEUCOM, 651, 130990. (www)

  @article{ , abstract = {Knowledge graphs provide a new paradigm for managing and understanding vast amounts of textual data and have been widely applied in fault diagnosis to enhance the safety and reliability of systems. However, they still face several challenges. First, existing methods lack deep interactive learning between cases. Second, the differences between some confounding cases are minimal, which significantly increases the diagnostic challenge. To address these issues, this paper introduces a Multi-level Contrastive Reasoning over Knowledge Graphs (MCRKG) approach, which for the first time incorporates a graph contrastive learning framework into knowledge-based fault diagnosis tasks. Specifically, the MCRKG method employs contrastive learning at two levels: node and graph. Its aim is to uncover commonalities and differences among confounding cases from both local and global perspectives. In addition, we propose a graph attention network with adaptive neighborhood awareness to learn the importance of different hop-level node representations. We conducted extensive experiments on real-world fault case data, and the results demonstrate that the MCRKG method outperforms state-of-the-art (SOTA) methods, achieving an accuracy of 97.89 % while also exhibiting robustness and explainability. We provide the code at https://github.com/XiangLiphd/MCRKG.}, archive = {J_NEUCOM}, author = {Xiang Li and Runhai Jiao and Changyu Zhou and Boxu Yan and Kaihang Li and Kangjun Liu}, doi = {10.1016/j.neucom.2025.130990}, journal = {Neurocomputing}, month = {10}, pages = {130990}, shortjournal = {Neurocomputing}, title = {Multi-level contrastive reasoning over knowledge graphs for fault diagnosis of confounding cases}, volume = {651}, year = {2025}, }
• (2025). An integrated multi-attribute graph sequence clustering with fuzzy information granule and knowledge-guidance. NEUCOM, 651, 130989. (www)

  @article{ , abstract = {Attribute graph sequence clustering is a type of vertex clustering that meets the needs of applications, like port scale clustering, which involves vertex attribute information and historical information. However, the existing methods explore such clustering tend to focus on only one of these aspects, missing opportunity for integrated analysis of data characteristics. To address this gap, our study introduces fuzzy information granule (FIG) to include as many vertices as possible while capturing multi-faceted information with precision, thereby enabling the integration of attribute information and historical information simultaneously. Based on these granular knowledges, two algorithms are raised to realize attribute graph sequence clustering, respectively for single-attribute graph sequence (Algorithm 1) and multi-attribute graph sequence (Algorithm 2). Algorithm 1 clusters single-attribute graph sequence employing FIG based fuzzy C-means algorithm, with the objects of FIGs. Leveraging the results of single-attribute clustering, Algorithm 2 extends to multi-attribute graph sequence clustering according to knowledge-guided idea. Ultimately, an accurate clustering result for multi-attribute graph sequence is ensured by considering the clustering result of each individual attribute. Worth noting that two novel proposed clustering algorithms achieve the clustering of attribute graph sequence successfully at application level, and also facilitate that of FIGs at methodological level. After comparing the proposed multi-attribute graph sequence clustering algorithm with six other algorithms on seven datasets, Algorithm 2 wins the highest clustering accuracy (CA) and adjusted rand index (ARI) values, and the smallest FIG-Index value, particularly with CA=0.88, ARI=0.62, and FIG‐Index=0.17 on a synthetic dataset. These results demonstrate that the newly proposed algorithm significantly outperforms other algorithms in clustering accuracy and robustness.}, archive = {J_NEUCOM}, author = {Fang Li and Jingxian Ma and Xinran Cheng}, doi = {10.1016/j.neucom.2025.130989}, journal = {Neurocomputing}, month = {10}, pages = {130989}, shortjournal = {Neurocomputing}, title = {An integrated multi-attribute graph sequence clustering with fuzzy information granule and knowledge-guidance}, volume = {651}, year = {2025}, }
• (2025). A review of object tracking based on deep learning. NEUCOM, 651, 130988. (www)

  @article{ , abstract = {The rapid advancement of deep learning has led to a surge in the development of object-tracking algorithms. Given the diverse objectives, backbone networks, and application methodologies, this study aims to integrate the prevalent tracking approaches comprehensively. We propose a systematic classification scheme based on application scenarios and primary methods, accompanied by a thorough analysis and concise summaries of each category. This approach provides a broader coverage of tracking techniques, facilitating a quicker understanding of the domain for novice researchers. In addition, we present standardized evaluation metrics and widely used datasets, including cross-method performance comparisons of selected algorithms on identical benchmarks to enhance the reader’s contextual understanding. Finally, we offer a critical assessment of current limitations, practical recommendations, and forward-looking perspectives to guide future research directions.}, archive = {J_NEUCOM}, author = {Guochen Zhao and Fanyong Meng and Chengzhuan Yang and Hui Wei and Dawei Zhang and Zhonglong Zheng}, doi = {10.1016/j.neucom.2025.130988}, journal = {Neurocomputing}, month = {10}, pages = {130988}, shortjournal = {Neurocomputing}, title = {A review of object tracking based on deep learning}, volume = {651}, year = {2025}, }
• (2025). Imbalanced open set domain generalization network for sensor fault diagnosis. NEUCOM, 651, 130987. (www)

  @article{ , abstract = {In recent years, the technique of applying domain generalization methods to solve cross-domain fault diagnosis problems has received widespread attention in the industrial community, among which, the open-set domain generalization fault diagnosis method effectively copes with the occurrence of unknown fault states in the target domain. However, issues such as data imbalance where fault data are scarce and normal data are abundant during the long-term operation of industrial sensors, and boundary shifts caused by unknown faults occurring in the target domain, make it difficult for the existing open-set domain generalization techniques to achieve accurate decision-making on sample types. This paper therefore introduces the HSL-ARAN generalization network, which can be generalized to carry out unknown fault diagnosis under imbalanced data conditions. First, a hierarchical style learning network is designed to encourage the generation of samples with relatively rich feature information, to address the issue of class imbalance in the source domain. Then, adversarial training with uncertainty weighting is used to extract reliable domain-invariant representations, and the inter-class relationships are leveraged to determine appropriate class boundaries and rejection thresholds. Finally, a new local clustering method is employed to further enhance the reliability of the class boundaries, which enables the identification of new fault modes. The algorithm is tested on sensor data for a nickel flash furnace system, and the effectiveness and superiority of the HSL-ARAN diagnosis method are verified.}, archive = {J_NEUCOM}, author = {Dongnian Jiang and Zhaiwen Wang and Huichao Cao and Dezhi Xu}, doi = {10.1016/j.neucom.2025.130987}, journal = {Neurocomputing}, month = {10}, pages = {130987}, shortjournal = {Neurocomputing}, title = {Imbalanced open set domain generalization network for sensor fault diagnosis}, volume = {651}, year = {2025}, }
• (2025). Un-CNL: An uncertainty-based continual noisy learning framework. NEUCOM, 651, 130985. (www)

  @article{ , abstract = {The goal of continual learning is to maintain model performance while adapting to new tasks and evolving data environments. This helps address catastrophic forgetting, a common issue in deep learning. However, challenges like human annotation errors and label biases introduce noisy labels into datasets, further intensifying catastrophic forgetting in neural networks. In response to these challenges, the concept of continual noisy learning (CNL) has emerged. While existing methods often rely on sample selection and replay strategies, they tend to focus solely on sample confidence, neglecting representativeness. To improve the reliability and representativeness of replayed samples, we propose a novel method called Un-CNL. This approach uses uncertainty purification techniques based on perturbed samples to separate data streams and select reliable samples for replay. Additionally, we apply CutMix data augmentation to enhance the representativeness of these samples. Subsequently, semi-supervised learning is employed for fine-tuning, combined with contrastive learning to handle the classification challenges posed by noisy data streams. We validated the effectiveness of Un-CNL through experiments on CIFAR-10 and CIFAR-100 datasets, demonstrating its superior performance compared to existing methods.}, archive = {J_NEUCOM}, author = {Guangrui Guo and Jinyong Cheng}, doi = {10.1016/j.neucom.2025.130985}, journal = {Neurocomputing}, month = {10}, pages = {130985}, shortjournal = {Neurocomputing}, title = {Un-CNL: An uncertainty-based continual noisy learning framework}, volume = {651}, year = {2025}, }
• (2025). Fast fixed-/preassigned-time synchronization of clifford-valued neural networks for medical image encryption. NEUCOM, 651, 130984. (www)

  @article{ , abstract = {This paper aims to investigate the fixed-time (FXT) and preassigned-time (PAT) synchronization for Clifford-valued neural networks (CFVNNs) with mixed delays by improving a novel FXT stability theorem and using non-decomposing two-step method. First of all, a novel FXT stability theorem has been derived. Its time estimation formula and settling time are simpler and accurate compared to existing stability theorem. Then, based on this novel FXT stability theorem, the FXT synchronization of the CFVNNs is obtained by designing sample nonlinear controller and Lyapunov function and seeking the settling time. As a special case, the PAT synchronization of CFVNNs is investigated, in which the estimation of settling time is independent of any initial conditions of neural networks and any parameters of the designed controllers. Lastly, numerical examples demonstrate the effectiveness and superiority of the derived theoretical results. The research also extends to the practical domain, evaluating the impact of CFVNNs and the designed controllers on medical image encryption.}, archive = {J_NEUCOM}, author = {Yanlin Zhang and Kit Ian Kou and Yanhui Zhang and Lizhi Liu}, doi = {10.1016/j.neucom.2025.130984}, journal = {Neurocomputing}, month = {10}, pages = {130984}, shortjournal = {Neurocomputing}, title = {Fast fixed-/preassigned-time synchronization of clifford-valued neural networks for medical image encryption}, volume = {651}, year = {2025}, }
• (2025). A review of double compression detection for digital multimedia. NEUCOM, 651, 130983. (www)

  @article{ , abstract = {The rapid advancement of AI-driven multimedia manipulation has created an urgent need for more sophisticated digital forensics solutions. Current detection methods, while effective against specific tampering types, suffer from limited generalizability across diverse manipulation techniques. To address this challenge, researchers have developed Double Compression Detection (DCD) as a universal approach through compression-domain analysis. This review presents the comprehensive analysis of DCD techniques, systematically evaluating cutting-edge techniques for audio, image, and video content forensics. The pros and cons of existing DCD schemes are summarized for the first time from the perspective of generalization and effectiveness in this review. The emerging trends and fundamental limitations of existing researches are critically examined to guide future research directions in DCD.}, archive = {J_NEUCOM}, author = {Tanfeng Sun and Xiao Han and Qiang Xu and Xing Yan and Yueneng Wang}, doi = {10.1016/j.neucom.2025.130983}, journal = {Neurocomputing}, month = {10}, pages = {130983}, shortjournal = {Neurocomputing}, title = {A review of double compression detection for digital multimedia}, volume = {651}, year = {2025}, }
• (2025). Improving multimodal named entity recognition via text-image relevance prediction with large language models. NEUCOM, 651, 130982. (www)

  @article{ , abstract = {Multimodal Named Entity Recognition (MNER) is a critical task in information extraction, which aims to identify named entities in text-image pairs and classify them into specific types such as person, organization and location. While existing studies have achieved moderate success by fusing visual and textual features through cross-modal attention mechanisms, two major challenges remain: (1) image-text mismatch, where the two modalities are not always semantically aligned in real-world scenarios; and (2) insufficient labeled data, which hampers the model’s ability to learn complex cross-modal associations and limits generalization. To overcome these challenges, we propose a novel framework that leverages the semantic comprehension and reasoning capabilities of Large Language Models (LLMs). Specifically, for the mismatch issue, we employ LLMs to generate the text-image relevance score with inference reason to guide the subsequent modules. Then we design T ext-image R elationship P redicting (TRP) module, which determines the final feature fusion weights based on the relevance score provided by LLMs. To mitigate data scarcity, we prompt LLMs to identify the key entities in text and incorporate them into the original input. Additionally, we design T ext-image R elevance F eatures L earning (TRFL) module to construct positive and negative samples based on the relevance score, employing a supervised contrastive learning method to further enhance the model’s ability to extract key features from image-text pairs. Experiments show that our proposed method achieves F1 scores of 75.32 % and 86.65 % on Twitter-2015 and Twitter-2017 datasets, respectively, demonstrating its effectiveness.}, archive = {J_NEUCOM}, author = {Qingyang Zeng and Minghui Yuan and Yueyang Su and Jia Mi and Qianzi Che and Jing Wan}, doi = {10.1016/j.neucom.2025.130982}, journal = {Neurocomputing}, month = {10}, pages = {130982}, shortjournal = {Neurocomputing}, title = {Improving multimodal named entity recognition via text-image relevance prediction with large language models}, volume = {651}, year = {2025}, }
• (2025). SEArch: A self-evolving framework for network architecture optimization. NEUCOM, 651, 130980. (www)

  @article{ , abstract = {This paper studies a fundamental network optimization problem that finds a network architecture with optimal performance (low loss) under given resource budgets (small number of parameters and/or fast inference). Unlike existing network optimization approaches such as network pruning, knowledge distillation (KD), and network architecture search (NAS), in this work we introduce a self-evolving pipeline to perform network optimization. In this framework, a simple network iteratively and adaptively modifies its structure by using the guidance from a teacher network, until it reaches the resource budget. An attention module is introduced to transfer the knowledge from the teacher network to the student network. A splitting edge scheme is designed to help the student model find an optimal macro architecture. The proposed framework combines the advantages of pruning, KD, and NAS, and hence, can efficiently generate networks with flexible structure and desirable performance. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that our framework achieves great performance in this network architecture optimization task.}, archive = {J_NEUCOM}, author = {Yongqing Liang and Dawei Xiang and Xin Li}, doi = {10.1016/j.neucom.2025.130980}, journal = {Neurocomputing}, month = {10}, pages = {130980}, shortjournal = {Neurocomputing}, title = {SEArch: A self-evolving framework for network architecture optimization}, volume = {651}, year = {2025}, }
• (2025). MKE-PLLM: A benchmark for multilingual knowledge editing on pretrained large language model. NEUCOM, 651, 130979. (www)

  @article{ , abstract = {Multilingual large language models have demonstrated remarkable performance across various downstream tasks but are still plagued by factuality errors. Knowledge editing aims to correct these errors by modifying the internal knowledge of pre-trained models. However, current knowledge editing methods primarily focus on monolingual settings, neglecting the complexities and interdependencies within multilingual scenarios. Furthermore, benchmarks specifically designed for multilingual knowledge editing are relatively scarce. Addressing this gap, this paper constructs a novel multilingual knowledge editing benchmark. This benchmark comprehensively evaluates methods for mLLMs based on accuracy, reliability, generalization, and consistency. To ensure the robustness and usability of the benchmark, we conducted detailed analysis and validation. Concurrently, we propose a baseline method that adapts existing monolingual knowledge editing techniques to the multilingual environment. Extensive experimental results demonstrate the effectiveness of our constructed benchmark in evaluating multilingual knowledge editing.}, archive = {J_NEUCOM}, author = {Ran Song and Shengxiang Gao and Xiaofei Gao and Cunli Mao and Zhengtao Yu}, doi = {10.1016/j.neucom.2025.130979}, journal = {Neurocomputing}, month = {10}, pages = {130979}, shortjournal = {Neurocomputing}, title = {MKE-PLLM: A benchmark for multilingual knowledge editing on pretrained large language model}, volume = {651}, year = {2025}, }
• (2025). Leveraging subclass learning for improving uncertainty estimation in deep learning. NEUCOM, 651, 130954. (www)

  @article{ , abstract = {Machine learning is becoming increasingly popular across various applications and has led to state-of-the-art results, but it faces challenges related to its trustworthiness. One aspect of making deep learning models more trustworthy is improving their ability to estimate the uncertainty of whether a sample is from the in-domain (ID) data distribution or not. Especially, neural networks have a tendency to make overly confident extrapolations and struggle to convey their uncertainty, which can limit their trustworthiness. Recent approaches have employed Radial Basis Function (RBF)-based models with great success in improving uncertainty estimation in Deep Learning. However, such models assume a unimodal distribution of the data for each class, which we show is critical for out-of-distribution sample detection, but can be limiting in many real world cases. To overcome these limitations, in this paper, we propose a method for training a deep model utilizing the inherent different modalities that naturally arise in a class in real data, which we call subclasses , leading to improved uncertainty quantification. The proposed method leverages a variance-preserving reconstruction-based representation learning approach that prevents feature collapse and enables robust discovery of subclasses, further improving the effectiveness of the proposed approach. The improvement of the approach is demonstrated using extensive experiments on several datasets.}, archive = {J_NEUCOM}, author = {Dimitrios Spanos and Nikolaos Passalis and Anastasios Tefas}, doi = {10.1016/j.neucom.2025.130954}, journal = {Neurocomputing}, month = {10}, pages = {130954}, shortjournal = {Neurocomputing}, title = {Leveraging subclass learning for improving uncertainty estimation in deep learning}, volume = {651}, year = {2025}, }
• (2025). CMC-GCN: Consistent multi-granularity cascading graph convolution network for multi-behavior recommendation. NEUCOM, 651, 130952. (www)

  @article{ , abstract = {Multi-behavior recommendation (MBR) has garnered growing attention recently due to its ability to mitigate the sparsity issue by inferring user preferences from various auxiliary behaviors to improve predictions for the target behavior. Although existing research on MBR has yielded impressive results, it still faces two major limitations. First, previous methods mainly focus on modeling fine-grained interaction information between users and items under each behavior, which may suffer from sparsity issue. Second, existing models usually concentrate on exploiting dependencies between two consecutive behaviors, leaving intra- and inter-behavior consistency largely unexplored. To this end, we propose a novel approach named Consistent Multi-Granularity Cascading Graph Convolution Network for Multi-Behavior Recommendation (CMC-GCN). To be specific, we first explore both fine- and coarse-grained correlations among users or items of each behavior by simultaneously modeling the behavior-specific interaction graph and its corresponding hypergraph in a cascaded manner. Then, we propose a behavior consistency-guided alignment strategy that preserves consistent representations between the interaction graph and its associated hypergraph for each behavior, while also maintaining representation consistency across different behaviors. Extensive experiments and analyses on four real-world benchmark datasets demonstrate that our proposed approach is consistently superior to previous state-of-the-art methods due to its capability to effectively attenuate the sparsity issue as well as keep both intra- and inter-behavior consistencies. The code is available at https://github.com/marqu22/CMCGCN.git .}, archive = {J_NEUCOM}, author = {Yabo Yin and Xiaofei Zhu and Kunyang Huang and Wenshan Wang and Yihao Zhang and Pengfei Wang and Yixing Fan and Jiafeng Guo}, doi = {10.1016/j.neucom.2025.130952}, journal = {Neurocomputing}, month = {10}, pages = {130952}, shortjournal = {Neurocomputing}, title = {CMC-GCN: Consistent multi-granularity cascading graph convolution network for multi-behavior recommendation}, volume = {651}, year = {2025}, }
• (2025). ExpTamed: An exponential tamed optimizer based on langevin SDEs. NEUCOM, 651, 130949. (www)

  @article{ , abstract = {This study presents a new method to improve optimization by regularizing the gradients in deep learning methods based on a novel taming strategy to regulate the growth of numerical solutions for stochastic differential equations. The method, ExpTamed, enhances stability and reduces the mean-square error across a short time horizon in comparison to existing techniques. The practical effectiveness of ExpTamed is rigorously evaluated on CIFAR-10, Tiny-ImageNet, and Caltech256 across diverse architectures. In direct comparisons with prominent optimizers like Adam, ExpTamed demonstrates significant performance gains. Specifically, it achieved increases in best top-1 test accuracy ranging from 0.86 to 2.76 percentage points on CIFAR-10, and up to 4.46 percentage points on Tiny-ImageNet (without learning rate schedule). On Caltech256, ExpTamed also yielded superior accuracy, precision, and Kappa metrics. These results clearly quantify ExpTamed’s capability to deliver enhanced performance in practical deep learning applications.}, archive = {J_NEUCOM}, author = {Utku Erdoğan and Şahin Işık and Yıldıray Anagün and Gabriel Lord}, doi = {10.1016/j.neucom.2025.130949}, journal = {Neurocomputing}, month = {10}, pages = {130949}, shortjournal = {Neurocomputing}, title = {ExpTamed: An exponential tamed optimizer based on langevin SDEs}, volume = {651}, year = {2025}, }
• (2025). Long-tail item recommendations exploiting graph neural network in collaborating filtering framework. NEUCOM, 651, 130948. (www)

  @article{ , abstract = {Collaborative filtering (CF) refers to a type of recommender system that generates recommendations solely based on users’ ratings on items. The CF approach to recommendations is very popular in academic and industry domains. However, recommendations provided by this CF approach are biased towards popular items. As a result, less popular items form a long tail in product space. Recently, researchers have started developing recommender systems that can mitigate long tail effect in recommendations. These recommender systems include hybrid approaches, heuristic approaches, and neural networks. However, by treating input data as independent features and ignoring the connections between nodes in a network, these methods lose information about the user-item relationship. Because of their capacity to capture intricate user-item interactions, Graph Neural Networks ( G N N s ) have become increasingly popular in recommender systems. In this paper, a two-phase deep neural network approach is proposed to enhance the long-tail item recommendations. A generative neural model such as Generative Adversarial Network (GAN) is utilized to impute long-tail item ratings in the first phase. The second phase uses a GNN to learn the embeddings of users and items based on the imputed user-item rating matrix. These learned embeddings are then used for generating recommendations. Extensive experiments are conducted on MovieLens 1M, MovieLens 100K, Amazon books, and Yelp business recommendation datasets. From the results, it has been observed that the proposed approach outperforms neural graph collaborative filtering (NGCF) and approaches specially developed for improving long-tail recommendations such as Econo-physics inspired approach, and graph augmentation-based approaches. Evaluation metrics such as Precision, Recall, NDCG and Catalog coverage are used.}, archive = {J_NEUCOM}, author = {N.Sangita Achary and Bidyut Kr. Patra and Sambit Bakshi}, doi = {10.1016/j.neucom.2025.130948}, journal = {Neurocomputing}, month = {10}, pages = {130948}, shortjournal = {Neurocomputing}, title = {Long-tail item recommendations exploiting graph neural network in collaborating filtering framework}, volume = {651}, year = {2025}, }
• (2025). A few-shot knowledge reasoning method based on three-way partial order structure and prompt learning. NEUCOM, 651, 130947. (www)

  @article{ , abstract = {As an emerging topic, few-shot knowledge reasoning is of great significance to the advancement of artificial intelligence. KR-POFSA is an approach that employs a three-way partial order structure for knowledge representation, combined with granular computing to achieve few-shot knowledge reasoning. However, it faces a limitation: when using it for knowledge reasoning, if the number of attributes is huge or there are a lot of cross-relationships between the attributes, then it may generate lots of new potential object patterns, most of which are useless. To solve this problem, this paper devises an innovative approach to improve KR-POFSA, which uses LLM to downstream few-shot knowledge reasoning tasks through appropriate prompt design. Specifically, we design a prompt template that guides LLM to output domain knowledge, such as a correlation matrix, and uses thresholds to limit the generation of invalid patterns. Through three experiments—with 8 objects and 9 attributes, 20 objects and 11 attributes, and 23 objects and 12 attributes, respectively—we demonstrate that our method can not only reduce the discovery of invalid attribute granules and object patterns in granular computing by 30 %–50 %, but also may offer practitioners insights into which attributes to prioritize, minimizing empiricism.}, archive = {J_NEUCOM}, author = {Yuxuan Huang and Enliang Yan and Peiming Zhang and Tianyong Hao}, doi = {10.1016/j.neucom.2025.130947}, journal = {Neurocomputing}, month = {10}, pages = {130947}, shortjournal = {Neurocomputing}, title = {A few-shot knowledge reasoning method based on three-way partial order structure and prompt learning}, volume = {651}, year = {2025}, }
• (2025). Meta-learning for physically-constrained neural system identification. NEUCOM, 651, 130945. (www)

  @article{ , abstract = {We present a gradient-based meta-learning framework for rapid adaptation of neural state-space models (NSSMs) for black-box system identification. When applicable, we also incorporate domain-specific physical constraints to improve the accuracy of the NSSM. The major benefit of our approach is that instead of relying solely on data from a single query system, our framework utilizes data from a diverse set of source systems, enabling learning from limited contextualizing data from a query system, as well as with few online training iterations. Through benchmark examples, we demonstrate the potential of our approach, study the effect of fine-tuning subnetworks rather than full fine-tuning, and report real-world case studies to illustrate the practical application and generalizability of the approach to practical problems with physical-constraints. Specifically, we show that the meta-learned models result in improved downstream performance in model-based state estimation in indoor localization and energy systems.}, archive = {J_NEUCOM}, author = {Ankush Chakrabarty and Gordon Wichern and Vedang M. Deshpande and Abraham P. Vinod and Karl Berntorp and Christopher R. Laughman}, doi = {10.1016/j.neucom.2025.130945}, journal = {Neurocomputing}, month = {10}, pages = {130945}, shortjournal = {Neurocomputing}, title = {Meta-learning for physically-constrained neural system identification}, volume = {651}, year = {2025}, }
• (2025). Adaptive neural network tracking control for nonlinear interconnected time-delay systems with prescribed performance. NEUCOM, 651, 130943. (www)

  @article{ , abstract = {This paper investigates the adaptive decentralized tracking control problem for a class of uncertain nonlinear interconnected systems with time-delay. Instead of coupling by output signal, the interconnected terms in the considered systems are allowed to contain all state information of other subsystems. To ensure transient and steady-state tracking behavior meets user-defined specifications, a time-varying performance function and error transformation are introduced for each subsystem, enabling the tracking error to evolve within predefined bounds regardless of initial conditions. An adaptive decentralized controller is then developed by integrating radial basis function (RBF) neural networks with command filters to approximate unknown nonlinearities and overcome the complexity explosion problem in traditional backstepping designs. It is rigorously proven that all closed-loop signals are uniformly bounded, and the tracking error of each subsystem converges to a prescribed residual set within a user-specified time. Simulation results demonstrate the effectiveness and robustness of the proposed control scheme.}, archive = {J_NEUCOM}, author = {Guopin Liu and Zihao Li and Changchun Hua and Yafeng Li and Jiannan Chen}, doi = {10.1016/j.neucom.2025.130943}, journal = {Neurocomputing}, month = {10}, pages = {130943}, shortjournal = {Neurocomputing}, title = {Adaptive neural network tracking control for nonlinear interconnected time-delay systems with prescribed performance}, volume = {651}, year = {2025}, }
• (2025). Large language model with iteratively prompt for temporal knowledge graph completion. NEUCOM, 651, 130941. (www)

  @article{ , abstract = {Temporal Knowledge Graph Completion (TKGC) has emerged as a pivotal undertaking in Artificial Intelligence research, entailing the interpolation of missing data points within temporal contexts. Despite its recent surge in interest, prevailing methodologies primarily rely on graph neural networks, often yielding suboptimal results due to inadequate timestamp extraction and underutilization of implicit relational information. This limitation stems from the inherent inability of traditional approaches to effectively capture complex temporal dependencies and nuanced semantic relationships. To remedy these shortcomings, we introduce a pioneering approach that harnesses the power of Large Language Models (LLMs) enhanced with iterative prompts. Our method systematically transforms sampled quadruples into LLM inputs by employing an auto-prompt mechanism, thereby facilitating the integration of contextual information and relational semantics. Subsequently, we introduce an iterative prompt model to dynamically steer the pre-trained language model through a sequence of knowledge recall steps, facilitating downstream task transformations and enabling the effective propagation of temporal dependencies. Experimental evaluations on three benchmark datasets, alongside comprehensive analyses, underscore our model’s superior performance across three key metrics, affirming its robust competitiveness within the field.}, archive = {J_NEUCOM}, author = {Xi Yang and Jia Zhu}, doi = {10.1016/j.neucom.2025.130941}, journal = {Neurocomputing}, month = {10}, pages = {130941}, shortjournal = {Neurocomputing}, title = {Large language model with iteratively prompt for temporal knowledge graph completion}, volume = {651}, year = {2025}, }
• (2025). Influence maximization in multilayer social networks using transformer-based node embeddings and deep neural networks. NEUCOM, 651, 130939. (www)

  @article{ , abstract = {Studies have shown that individuals frequently interact and influence other people within a society. Understanding and identifying influential individuals within a network is crucial for optimizing information diffusion. This challenge, known as influence maximization, has attracted significant attention, particularly in multilayer social networks where individuals participate across diverse contexts. Traditional approaches rely on heuristic or approximation algorithms, but their scalability and adaptability remain limited. In this paper, we propose a novel deep neural network architecture to predict and maximize influence in multilayer social networks. The framework utilizes a combination of node- and layer-specific feature embeddings, a transformer encoder for contextual feature integration, and multilayer perceptron (MLPs) for influence regression. The input comprises feature vectors representing nodes and layers, which are encoded and aggregated to approximate the influence of each node within its respective layer. A final aggregation step computes the total influence spread across layers, enabling efficient seed set selection of highly influential nodes. Our method yields favorable outcomes, effectively tackling challenges such as hardware resource requirements, scalability, and runtime performance. Empirical evaluations against state-of-the-art algorithms demonstrate the effectiveness of the proposed model in achieving superior influence spread with reduced computational overhead. This approach proposes new paths for influence maximization in large-scale, multilayer social networks.}, archive = {J_NEUCOM}, author = {Xilai Ju and Ali Seyfi and Asgarali Bouyer and Alireza Rouhi and Xiaoyang Liu and Bahman Arasteh}, doi = {10.1016/j.neucom.2025.130939}, journal = {Neurocomputing}, month = {10}, pages = {130939}, shortjournal = {Neurocomputing}, title = {Influence maximization in multilayer social networks using transformer-based node embeddings and deep neural networks}, volume = {651}, year = {2025}, }
• (2025). Latent structure-oriented asymmetric hashing for cross-modal retrieval. NEUCOM, 651, 130938. (www)

  @article{ , abstract = {Cross-modal hashing has attracted considerable attention in cross-modal retrieval due to its excellent computational efficiency and retrieval performance. Most existing methods aim to map multimodal data into a common representation space where either semantic similarity or instance similarity is preserved. However, these methods do not consider the potential clustering structure of instances that characterizes sample separability, resulting in degraded retrieval performance. Furthermore, capturing the consistent instance similarity by effectively fusing similarities of different modalities remains an essential problem to be addressed. To tackle these issues, this paper proposes a novel latent structure-oriented asymmetric cross-modal Hashing method (LSOAH) for cross-modal retrieval. Specifically, LSOAH formulates the common representation learning with orthogonal decomposition, where each modality-specific instance is projected and decomposed into a modality-specific base matrix and a common cluster indicator matrix, and where the indicator matrix is concatenated with the hash code via an asymmetric mechanism. Additionally, we utilize Hadamard product on graphs from different modalities to explore the consistent instance similarity, and embed it in the common representation. Finally, a unified objective function is presented to enable the simultaneous exploration of the cluster structure, instance similarity and semantic similarity, as well as the hash code learning, upon which an alternating optimization algorithm is developed with theoretically proven convergence. Experimental results on three benchmark datasets confirm the superiority of the proposed LSOAH for cross-modal retrieval.}, archive = {J_NEUCOM}, author = {Jiajun Ma}, doi = {10.1016/j.neucom.2025.130938}, journal = {Neurocomputing}, month = {10}, pages = {130938}, shortjournal = {Neurocomputing}, title = {Latent structure-oriented asymmetric hashing for cross-modal retrieval}, volume = {651}, year = {2025}, }
• (2025). MFGDAFormer: Multi-scale frequency-guided dual-branch attention transformer for low-light image enhancement. NEUCOM, 651, 130937. (www)

  @article{ , abstract = {Images captured in low-light conditions, such as at night or in backlit environments, often suffer from uneven illumination, insufficient contrast, and significant noise, which compromise the accuracy and robustness of visual tasks like object detection. Low-light image enhancement techniques are essential for improving visibility, restoring details, and reducing noise, thereby enhancing the performance of various visual tasks. Effective illumination adjustment and detail preservation are key to achieving successful low-light image enhancement. However, existing methods face significant challenges, including inadequate modeling of lighting patterns, insufficient local feature representation, neglect of frequency-domain information, and difficulty balancing noise suppression with detail preservation. This paper introduces the Multi-scale Frequency-Guided Dual-branch Attention Transformer (MFGDAFormer) for low-light image enhancement. The framework incorporates a CNN-based illumination estimator to guide enhancement and lighting adjustment, while the Fourier Sparse Multi-scale Attention Mechanism (FSMAM) facilitates effective frequency-domain analysis, adaptive feature modulation, and multi-scale detail preservation. Furthermore, the U-Net architecture is employed for fine-grained illumination enhancement, and the Dual-Large Kernel Activation Attention Module (DLKA) integrates large-kernel convolutions with adaptive attention mechanisms to optimize feature fusion. Experimental results demonstrate that MFGDAFormer outperforms state-of-the-art models, with PSNR improvements ranging from 0.21 dB to 0.88 dB across multiple low-light datasets.}, archive = {J_NEUCOM}, author = {Faming Gong and Yimeng Zhang and Chengze Du and Xiaofeng Ji}, doi = {10.1016/j.neucom.2025.130937}, journal = {Neurocomputing}, month = {10}, pages = {130937}, shortjournal = {Neurocomputing}, title = {MFGDAFormer: Multi-scale frequency-guided dual-branch attention transformer for low-light image enhancement}, volume = {651}, year = {2025}, }
• (2025). Enhancing collaborative translational metric learning with causal graph network. NEUCOM, 651, 130933. (www)

  @article{ , abstract = {The application of metric learning in recommender systems has gained considerable attention, particularly for enabling supervised training through Euclidean distance. However, conventional collaborative translational metric learning typically represents each user and item as a single point and projects collaborative signals onto a specific translation vector within a uniform space. This approach fails to explicitly capture the high-order semantics and dynamic nature of user-item interactions derived from implicit user behavior. Given the extensive research on graph neural networks (GNNs), a natural extension is to incorporate high-order information into the generation of translation vectors. Yet, our empirical findings reveal that the performance is not as expected. In this paper, we systematically investigate the limitations of existing methods and attribute the issue to semantic confusion. Motivated by this finding, we propose a novel framework that integrates bilateral metric learning with a causal graph network. Specifically, we construct a structural causal model (SCM) to disentangle user behavior into causal features and confounders. To mitigate the influence of these confounders, we propose a conditional intervention module that generates confounder-specific conditions for the target. Moreover, inspired by this conditional intervention mechanism, we develop an additional module conditioned on user behavior to enhance the diversity of the recommended list, recognizing that diversity is also a crucial metric for evaluating recommendation quality. We conducted a series of experiments to validate the effectiveness of our proposed method. The results demonstrate that our approach achieves state-of-the-art performance in terms of both recommendation accuracy and diversity.}, archive = {J_NEUCOM}, author = {Jingjing Wang and Haoran Xie and S.Joe Qin and Xiaohui Tao and Fu Lee Wang and Xiaoliang Xu}, doi = {10.1016/j.neucom.2025.130933}, journal = {Neurocomputing}, month = {10}, pages = {130933}, shortjournal = {Neurocomputing}, title = {Enhancing collaborative translational metric learning with causal graph network}, volume = {651}, year = {2025}, }
• (2025). A new U-shapelet clustering algorithm by capturing morphological trends. NEUCOM, 651, 130932. (www)

  @article{ , abstract = {U-shapelet-based time series clustering methods utilize discriminative subsequences for clustering, ensuring strong interpretability and high accuracy. While many methods focus on improving efficiency by shrinking the U-shapelet candidate set, they often sacrifice clustering accuracy due to inadequate attention to U-shapelet quality. To address this issue, this paper presents Morph, a novel U-shapelet-based clustering algorithm that captures morphological trends in time series. Firstly, the Discrete Feature Representation (DFR) is introduced to effectively capture key information and morphological trends. Secondly, the Extended Euclidean Distance (EED) is developed as a robust similarity measure that enhances the selection of high-quality U-shapelets by integrating seamlessly with DFR. Lastly, random masking is incorporated into Morph to expedite the selection process of the U-shapelet set. Experiments on 14 diverse datasets show that Morph outperforms 15 baselines, demonstrating its effectiveness in achieving high accuracy and interpretability.}, archive = {J_NEUCOM}, author = {Mei Chen and Yu Wang and Mingwei Leng and Boya Liu and Yuanyuxiu You and Yiying Yao and Song Wang}, doi = {10.1016/j.neucom.2025.130932}, journal = {Neurocomputing}, month = {10}, pages = {130932}, shortjournal = {Neurocomputing}, title = {A new U-shapelet clustering algorithm by capturing morphological trends}, volume = {651}, year = {2025}, }
• (2025). Two novel harmonic-resistant zeroing neural networks for time-varying problems in robotic manipulators. NEUCOM, 651, 130930. (www)

  @article{ , abstract = {This paper presents two novel harmonically disturbance-resistant zeroing neural network (ZNN) models: the known frequency harmonic-resistant ZNN (KFHRZNN) and the unknown frequency harmonic-resistant ZNN (UFHRZNN). These models are designed to tackle the pseudoinverse of time-varying matrices and inverse kinematics challenges in robotic manipulators. By precisely accounting for the derivatives of harmonic disturbances, they significantly mitigate these interferences, thereby improving the control efficacy of robots in high-speed, dynamic settings. The study elucidates the design rationale, convergence characteristics, and stability assessments for both KFHRZNN and UFHRZNN. Numerical simulations and physical experiments validate the effectiveness and advantages of these models in resolving time-varying issues within robotic manipulators, highlighting their precision and robustness against harmonic disturbances.}, archive = {J_NEUCOM}, author = {Bing Zhang and Xinglong Chen and Yuhua Zheng and Shuai Li and Duc Truong Pham and Yao Mao}, doi = {10.1016/j.neucom.2025.130930}, journal = {Neurocomputing}, month = {10}, pages = {130930}, shortjournal = {Neurocomputing}, title = {Two novel harmonic-resistant zeroing neural networks for time-varying problems in robotic manipulators}, volume = {651}, year = {2025}, }
• (2025). SNR: One single network for image steganography with robust post-save recovery. NEUCOM, 651, 130929. (www)

  @article{ , abstract = {Image steganography has been proposed to transmit information securely in an open environment. Hiding invisibility, recovery quality and security are three primary challenges in image steganography task. Additionally, most existing image steganography methods treat image concealing and revealing as separate parts, consuming a large computing resources in training. This paper proposes a novel image steganography with robust post-save recovery using one single network, namely SNR. Specifically, we design a Swin Transformer based model with multi-level symmetric structure, which allows image concealing and revealing to share the same set of networks and parameters. A WaveletBlock module and a concealing wavelet loss are proposed to constrain secret information to be hidden in high-frequency subbands of different scales of the cover image to enhance steganographic quality. We find that quantization error is introduced when stego images are saved locally, and we analyze its distribution as well as its impact on recovery images. To improve recovery robustness, we add noise with the same distribution as the error in the revealing process. Experimental results demonstrate that SNR significantly outperforms other state-of-the-art image steganography methods, with more than 3.68 dB and 0.96 dB PSNR improvement in concealing and revealing compared with HiNet. Moreover, it exhibits good generalization ability on multiple datasets. The code is available at https://github.com/Floren1026/SNR .}, archive = {J_NEUCOM}, author = {Chao Yang and Shiyuan Wang and Ying Huang and Mingqiang Guo}, doi = {10.1016/j.neucom.2025.130929}, journal = {Neurocomputing}, month = {10}, pages = {130929}, shortjournal = {Neurocomputing}, title = {SNR: One single network for image steganography with robust post-save recovery}, volume = {651}, year = {2025}, }
• (2025). MFD-CANet: An intelligent identification model for thunderstorm wind gust via multidimensional feature decoupling and spatiotemporal cross-attention. NEUCOM, 651, 130926. (www)

  @article{ , abstract = {Thunderstorm wind gusts are extreme weather phenomena characterized by strong surges and destructive power. Due to their complex evolution and the wide range of influencing factors, traditional methods relying on a single observation, such as automatic weather stations, radar, or satellite, often fail to provide accurate and comprehensive monitoring. To address the issues of insufficient spatial resolution and low accuracy in conventional models, this paper proposes an intelligent identification model for thunderstorm wind gusts, called MFD-CANet. The model integrates multisource observational data, including ground wind speed, radar reflectivity, and satellite brightness temperature, to effectively exploit complementary information between different modalities to accurately capture the spatial organizational structure and evolutionary characteristics of thunderstorm wind gusts. Specifically, the model first introduces a multidimensional nonlinear structural perception module to extract key storm evolution features in the meso- and small-scale convective systems. It then employs a spatiotemporal cross-attention mechanism to effectively integrate temporal dynamics with spatial distribution information, enhancing the ability to identify the occurrence locations of thunderstorm wind gusts. Finally, a loss function based on the Critical Success Index is designed to optimize the balance between accuracy and false alarms. Experimental results demonstrate that MFD-CANet achieves superior performance compared to existing mainstream methods, with average improvements of 6.6 % in Probability of Detection (POD) and 7.8 % in Critical Success Index (CSI), along with a 9.5 % reduction in False Alarm Rate (FAR). This model provides more precise and reliable technical support for real-time monitoring and early warning of severe convective weather.}, archive = {J_NEUCOM}, author = {Bingjian Lu and Zhenyu Lu and Xiaowen Zhang and Quanbo Ge}, doi = {10.1016/j.neucom.2025.130926}, journal = {Neurocomputing}, month = {10}, pages = {130926}, shortjournal = {Neurocomputing}, title = {MFD-CANet: An intelligent identification model for thunderstorm wind gust via multidimensional feature decoupling and spatiotemporal cross-attention}, volume = {651}, year = {2025}, }
• (2025). Improving few-shot object detection via mislabeling mitigation. NEUCOM, 651, 130921. (www)

  @article{ , abstract = {Few-shot object detection (FSOD) aims to extend object detection capabilities with few annotated samples, enabling models to efficiently adapt to novel classes. Recently, FSOD has garnered significant attention. However, existing methods commonly rely on high-quality annotated data. In incomplete annotation scenarios, some foreground objects may be mislabeled as background, leading the model to learn incorrect information, which exacerbates class confusion and impairs novel class detection. Meanwhile, the potential positive information present in unlabeled objects remains underutilized, posing an additional challenge to FSOD, as data scarcity is already a fundamental issue. To this end, we propose a Mislabeling Adaptive Mitigation (MAM) strategy, which selects potential mislabeled samples within the background class based on feature similarity and reallocates them as positive samples to mitigate erroneous learning. Additionally, we dynamically adjust their weights to reasonably control the extent of their positive contribution to the training process. For feature aggregation, we introduce a Dual-Phase Sampling (DPS) strategy and a Weighted Non-Linear Aggregation (WNLA) module, which work synergistically to enhance feature interaction and fusion. Experiments on the PASCAL VOC and MS-COCO benchmark datasets demonstrate that our proposed Mislabeling Mitigation (MLM) method, which integrates MAM, DPS, and WNLA, outperforms state-of-the-art methods across multiple few-shot settings. Notably, it achieves a 1.7 percentage point improvement on VOC (Split 3, 2-shot), along with an average gain of 15 % over the baseline on VOC (Split 1).}, archive = {J_NEUCOM}, author = {Pei Yu and Gaocai Wang and Man Wu and Lili Wen}, doi = {10.1016/j.neucom.2025.130921}, journal = {Neurocomputing}, month = {10}, pages = {130921}, shortjournal = {Neurocomputing}, title = {Improving few-shot object detection via mislabeling mitigation}, volume = {651}, year = {2025}, }
• (2025). DSCN-net: Domain-specific contrastive network for unsupervised low-dose CT denoising. NEUCOM, 651, 130919. (www)

  @article{ , abstract = {Low-dose computed tomography (LDCT) technology reduces patient exposure to X-ray radiation, but results in degraded image quality with significant noise and artifacts. While supervised deep learning methods effectively denoise LDCT images, they rely heavily on numerous paired training samples, which are difficult to obtain in clinical settings. In this study, we present a novel unsupervised network model, the Domain-Specific Contrastive Network (DSCN-Net), for LDCT image denoising, providing a more practical and accessible solution for clinical denoising. Our approach begins by disentangling the input images into content and noise components, thus establishing three distinct data domains: the noisy LDCT image domain, the clean normal-dose CT (NDCT) image domain, and the noise domain. We employ a bidirectional contrastive learning architecture that assigns a dedicated generator to each domain, allowing for tailored, domain-specific data representations. In addition, we introduce two types of contrastive losses to model the relationships between these domains. Specifically, we introduce Intra-Contrastive Loss to align corresponding patches in the embedding space between the noisy and clean image domains for content consistency, and Inter-Contrastive Loss to separate the interactions between the image and noise domains to minimize the intrusion of image content features into the noise components (and vice versa). Experimental results show that DSCN-Net achieves state-of-the-art performance, outperforming other unsupervised methods in noise reduction and fine structure preservation, making it a promising solution for clinical use.}, archive = {J_NEUCOM}, author = {Rui Zhang and Yuanke Zhang and Yanfei Guo and Hanxiang Wang and Bingbing Wei and Fei Ma and Jing Meng and Jianlei Liu and Hongbing Lu and Yang Chen}, doi = {10.1016/j.neucom.2025.130919}, journal = {Neurocomputing}, month = {10}, pages = {130919}, shortjournal = {Neurocomputing}, title = {DSCN-net: Domain-specific contrastive network for unsupervised low-dose CT denoising}, volume = {651}, year = {2025}, }
• (2025). Generalizable agent modeling for agent collaboration-competition adaptation with multi-retrieval and dynamic generation. NEUCOM, 651, 130912. (www)

  @article{ , abstract = {Adapting a single agent to a new multi-agent system brings challenges, necessitating adjustments across various tasks, environments, and interactions with unknown teammates and opponents. Addressing this challenge is highly complex, and researchers have proposed two simplified scenarios, Multi-agent reinforcement learning for zero-shot learning and Ad-Hoc Teamwork. Building on these foundations, we propose a more comprehensive setting, Agent Collaborative-Competitive Adaptation (ACCA), which evaluates an agent to generalize across diverse scenarios, tasks, and interactions with both unfamiliar opponents and teammates. In ACCA, agents adjust to task and environmental changes, collaborate with unseen teammates, and compete against unknown opponents. We introduce a new modeling approach, Multi-Retrieval and Dynamic Generation (MRDG), that effectively models both teammates and opponents using their behavioral trajectories. This method incorporates a positional encoder for varying team sizes and a hypernetwork module to boost agents’ learning and adaptive capabilities. Additionally, a viewpoint alignment module harmonizes the observational perspectives of retrieved teammates and opponents with the learning agent. Extensive tests in benchmark scenarios like SMAC, Overcooked-AI, and Melting Pot show that MRDG significantly improves robust collaboration and competition with unseen teammates and opponents, surpassing established baselines. Our code is available at: https://github.com/vcis-wangchenxu/MRDG.git .}, archive = {J_NEUCOM}, author = {Chenxu Wang and Yonggang Jin and Cheng Hu and Youpeng Zhao and Zipeng Dai and Jian Zhao and Liuyu Xiang and Junge Zhang and Zhaofeng He}, doi = {10.1016/j.neucom.2025.130912}, journal = {Neurocomputing}, month = {10}, pages = {130912}, shortjournal = {Neurocomputing}, title = {Generalizable agent modeling for agent collaboration-competition adaptation with multi-retrieval and dynamic generation}, volume = {651}, year = {2025}, }
• (2025). FairBias: Mitigating bias in medical image diagnosis with mixed noise and class imbalance. NEUCOM, 651, 130910. (www)

  @article{ , abstract = {In medical image diagnosis, class-unbalanced and noisy datasets provide a challenge to deep learning algorithms. Biased models, poor performance for minority classes, and a decline in confidence in AI-based diagnosis are the results of these problems. These problems, especially class imbalance, are not adequately addressed by current approaches like QMix. To address these challenges, we propose FairBias, a novel framework that integrates class-aware sample separation, bias-aware loss functions, dynamic class reweighting, and advanced data augmentation to mitigate bias while maintaining high diagnostic accuracy. FairBias achieves significant fairness improvements by reducing the True Positive Rate (TPR) gap between majority and minority subgroups from 15 % to 5 % on datasets such as CheXpert and Breast MRI/FFDM. Additionally, FairBias narrows the Equal Opportunity (EO) disparity from 12 % to 3 % and improves Demographic Parity (DP) by ensuring that predicted positive rates across subgroups differ by less than 2 %. These fairness enhancements are achieved without compromising diagnostic performance, as evidenced by AUC values of 0.97 on Breast MRI/FFDM, 0.95 on Hep-2, 0.92 on SOKL, and 0.96 on CheXpert. Furthermore, FairBias demonstrates robustness under high noise ratios, achieving a Kappa score of 88.5 % on the Hep-2 dataset with 55 % Mis-Labeled Low-Quality Samples (MLQS) and 20 % Mis-High-Quality Samples (MHQS), outperforming state-of-the-art (SOTA) methods like QMix and vlm-fairness. It obtains AUC values of 0.97 on Breast MRI/FFDM, 0.95 on Hep-2, 0.92 on SOKL, and 0.96 on CheXpert. By addressing both mixed noise and class imbalance, FairBias ensures equitable and accurate disease diagnosis, particularly for minority classes that are often overlooked by existing systems. The framework’s limitations, including computational complexity and generalization to new modalities, are discussed, along with future directions to enhance its scalability and applicability. Overall, FairBias presents a significant step toward fairer and more reliable AI-driven medical image diagnosis, bridging critical gaps in current SOTA techniques.}, archive = {J_NEUCOM}, author = {Saeed Iqbal and Xiaopin Zhong and Muhammad Attique Khan and Zongze Wu and Nouf Abdullah Almujally and Weixiang Liu and Amir Hussain}, doi = {10.1016/j.neucom.2025.130910}, journal = {Neurocomputing}, month = {10}, pages = {130910}, shortjournal = {Neurocomputing}, title = {FairBias: Mitigating bias in medical image diagnosis with mixed noise and class imbalance}, volume = {651}, year = {2025}, }
• (2025). Stability of delayed quaternion-valued neural networks with general probabilistic bounded markovian switching. NEUCOM, 651, 130908. (www)

  @article{ , abstract = {The study delves into the stability problem of quaternion-valued neural networks (QVNNs) with time-varying discrete delays and distributed delays as well as general probabilistic bounded Markovian switching. Firstly, the non-commutative nature of quaternion multiplication complicates theoretical analysis and numerical computation when decomposition methods are employed. To address this, a suitable Lyapunov-Krasovskii functional is constructed and combined with the method of free-weighting matrix and inequality techniques, the QVNNs with Markovian switching are analyzed as a whole, yielding stability criteria in the form of linear matrix inequalities (LMIs). Additionally, as the transition probabilities are general probabilistic bounded, the system becomes more versatile and realistic. Finally, two example with simulations are given to show the validity and applicability of the achieved result.}, archive = {J_NEUCOM}, author = {Miao Shu and Qiankun Song and Yurong Liu}, doi = {10.1016/j.neucom.2025.130908}, journal = {Neurocomputing}, month = {10}, pages = {130908}, shortjournal = {Neurocomputing}, title = {Stability of delayed quaternion-valued neural networks with general probabilistic bounded markovian switching}, volume = {651}, year = {2025}, }
• (2025). GRSDet: Learning to generate local reverse samples for few-shot object detection. NEUCOM, 651, 130903. (www)

  @article{ , abstract = {Few-shot object detection (FSOD) aims to achieve object detection only using a few novel class training data. While transfer learning based methods have demonstrated remarkable performance in FSOD, the confusion between novel classes and similar base classes in knowledge transfer remains an issue to be concerned about. In this paper, we consider the conversion relationship between non-similar samples of the same category (Local Reverse Samples, LRSamples) in the base class can be learned and also transferred to the novel class. So we propose the selection rule and intra-class feature converter (IFC) of LRSamples in our Center Calibration Variance Augmentation (CCVA) module to adaptively adjust the center position of the novel class distribution for FSOD. Moreover, we propose a Feature Density Boundary Optimization (FDBO) module to adaptively adjust the importance of samples depending on their distance from the decision boundary. It emphasizes the importance of the high-density area of the similar class (closer decision boundary area) and reduce the weight of the low-density area of the similar class (farther decision boundary area), optimizing a clearer decision boundary for each category. Our method achieves consistent improvement on the Pascal VOC and MS COCO datasets based on DeFRCN and MFDC baselines.}, archive = {J_NEUCOM}, author = {Hefei Mei and Taijin Zhao and Shiyuan Tang and Heqian Qiu and Lanxiao Wang and Minjian Zhang and Fanman Meng and Hongliang Li}, doi = {10.1016/j.neucom.2025.130903}, journal = {Neurocomputing}, month = {10}, pages = {130903}, shortjournal = {Neurocomputing}, title = {GRSDet: Learning to generate local reverse samples for few-shot object detection}, volume = {651}, year = {2025}, }
• (2025). PyfUS: Python-based open-source software for the analysis of functional ultrasound imaging data. NEUCOM, 651, 130899. (www)

  @article{ , abstract = {Functional ultrasound (fUS) imaging has emerged as an important technology for investigating a multitude of neuro-related research areas in various animal models and patients. Beyond imaging quality and specificity, computational analysis of fUS data sets is critical for accurately and comprehensively characterize brain functions and circuits, under physiological and pathological conditions. To facilitate efficient and reproducible data analysis, we present a python-based open-source software (PyfUS) providing an end-to-end pipeline for registration, signal processing and visualization of fUS datasets. In addition to the conventional analysis - region-based averaging and correlation - we introduce the single-voxel clustering as an alternative analysis method that allows simultaneous spatial and temporal examination of the fUS signals at the finest scale allowed by the fUS. We compare the different strategies for analyzing fUS data and display the results of the analytical pipeline on a dataset comprising awake mice subjected to visual stimulation. In a standard computing environment with 32 GB of memory, a 10-Gb data size of brain-wide fUS images can be loaded in ∼1 h and fully processed with the 3 analysis methods in few minutes. The flexibility of the software allows for the easy extension to other animal models and user-developed modules. We deliver a tool providing a convenient access to state-of-the-art analysis methods and an open platform for the development of new processing strategies.}, archive = {J_NEUCOM}, author = {Théo Lambert and Clément Brunner and Gabriel Montaldo and Alan Urban}, doi = {10.1016/j.neucom.2025.130899}, journal = {Neurocomputing}, month = {10}, pages = {130899}, shortjournal = {Neurocomputing}, title = {PyfUS: Python-based open-source software for the analysis of functional ultrasound imaging data}, volume = {651}, year = {2025}, }
• (2025). GrafoRVFL: A gradient-free optimization framework for boosting random vector functional link network. NEUCOM, 651, 130898. (www)

  @article{ , abstract = {Random Vector Functional Link (RVFL) networks have garnered attention as a rapid and efficient neural network model due to their simplified architecture and reduced training complexity. Nevertheless, the hyperparameter tuning of this network remains a substantial obstacle in the pursuit of enhanced performance across many applications. In this study, we present GrafoRVFL, an open-source framework that employs gradient-free algorithms to optimize RVFL networks’ hyperparameters. GrafoRVFL is a system that is adaptable and helps to enhance the performance of RVFL models. It is constructed on top of Numpy, Mealpy, and Scikit-Learn. We evaluate the proposed framework by comparing 14 hybrid gradient-free trained RVFL models on a variety of regression and classification datasets. The best-performing models achieve classification accuracies of 96%, 92%, and 85% on the breast cancer, waveform, and magic telescope datasets, respectively. For regression, R-scores of 0.70, 0.89, and 0.80 are observed on the diabetes, Boston housing, and California housing datasets. Additionally, we compare three hybrid RVFL models with GridSearchCV and RandomizedSearchCV on the digits dataset. The results show that our hybrid models yield better performance while requiring significantly less computational time. This suggests that our proposed framework can serve as a critical resource for researchers and practitioners who are seeking practical and resilient approaches to real-world issues. The source code of the library is accessible to the public on the GitHub repository: https://github.com/thieu1995/GrafoRVFL .}, archive = {J_NEUCOM}, author = {Nguyen Van Thieu and Nguyen Thanh Hoang and Hossam Faris}, doi = {10.1016/j.neucom.2025.130898}, journal = {Neurocomputing}, month = {10}, pages = {130898}, shortjournal = {Neurocomputing}, title = {GrafoRVFL: A gradient-free optimization framework for boosting random vector functional link network}, volume = {651}, year = {2025}, }
• (2025). Adaptive hypergraph structure regularized semi-supervised non-negative matrix factorization for image clustering. NEUCOM, 651, 130895. (www)

  @article{ , abstract = {Semi-supervised non-negative matrix factorization (SNMF) is a powerful technique used in image clustering. Unlike traditional graphs, hypergraphs can capture higher-order geometric relationships among data samples. However, many existing hypergraph learning methods use a fixed hypergraph structure, which fails to dynamically optimize the hypergraph during the learning process. As a result, these methods may not accurately represent the true relationships between samples. To address this issue, this paper introduces a novel semi-supervised SNMF-based method called Adaptive Hypergraph Structure Regularized Semi-supervised Non-negative Matrix Factorization (AHS-SNMF). This method enhances the capture of high-order similarity between data samples by adaptively adjusting the hypergraph structure and iteratively improving it throughout the learning process. Additionally, we integrate labeled projection matrix with hypergraph regularization to minimize common errors found in traditional clustering methods. This approach strengthens the representation of labeled data in relation to unlabeled data, thereby boosting the model’s robustness and performance. Experimental results from comparative studies on six datasets confirm that the AHS-SNMF method significantly improves performance in semi-supervised learning tasks.}, archive = {J_NEUCOM}, author = {Xiaowan Ren and Youlong Yang}, doi = {10.1016/j.neucom.2025.130895}, journal = {Neurocomputing}, month = {10}, pages = {130895}, shortjournal = {Neurocomputing}, title = {Adaptive hypergraph structure regularized semi-supervised non-negative matrix factorization for image clustering}, volume = {651}, year = {2025}, }
• (2025). A new dataset, model, and benchmark for lightweight and real-time underwater object detection. NEUCOM, 651, 130891. (www)

  @article{ , abstract = {Underwater object detection (UOD) is crucial for monitoring marine ecosystems, underwater robotics, environmental protection, and autonomous underwater vehicles (AUVs). Despite progress, many models struggle under real-world conditions due to poor visibility, dynamic lighting, and domain shifts. Traditional methods like Faster R-CNN are computationally expensive, while YOLO-based models suffer in challenging underwater scenarios. The scarcity of large-scale annotated datasets further limits model generalization. To address these challenges, we introduce UOD-SZTU-2025, a new dataset of 3,133 high-quality underwater images, sourced primarily from video platforms. The dataset is used in EFCWM (Enhanced Feature Correction and Weighting Module) to extract and refine a feature material library for detection targets. We present EFCWM-Mamba-YOLO , a novel lightweight and real-time underwater object detector that integrates enhanced feature correction with state-space modeling to improve detection accuracy and robustness in complex underwater environments. The EFCWM module incorporates domain adaptation for improved robustness. Additionally, a two-stage training strategy first trains on a source domain and fine-tunes with limited target domain samples to enhance generalization. Experiments show our approach surpasses existing lightweight UOD models in accuracy, real-time performance, and robustness. Our dataset, model, and benchmark establish a strong foundation for future UOD research.}, archive = {J_NEUCOM}, author = {Huilin Ge and Pan Sun and Yu Lu}, doi = {10.1016/j.neucom.2025.130891}, journal = {Neurocomputing}, month = {10}, pages = {130891}, shortjournal = {Neurocomputing}, title = {A new dataset, model, and benchmark for lightweight and real-time underwater object detection}, volume = {651}, year = {2025}, }
• (2025). Multi-layer federated learning for bit-width and data heterogeneity in cloud-edge systems. NEUCOM, 651, 130890. (www)

  @article{ , abstract = {In edge intelligence scenarios, quantized models often vary in bit-width due to hardware diversity. Inconsistent model quantization and non-IID local data create dual heterogeneity, which causes significant challenges for federated training across edges. To overcome these challenges, we introduce a novel method (multi-layer Federated Learning with Bit-width Adaptivity) for dual heterogeneity, which facilitates collaborative optimization across differently quantized models. Our method establishes multiple model layers at edges, enabling alternate execution of collaborative and local updates. During collaborative updates, edges synchronize with other edges by selecting suitable models with the same bit-width from their candidate sets. To reduce communication, a Contact Map is designed at the cloud, tracking model selection and guiding efficient collaboration. We theoretically analyze the reduction of communication burden by the Contact Map. Experimental evaluations show that our method outperforms existing approaches in terms of model accuracy in dual heterogeneous settings.}, archive = {J_NEUCOM}, author = {Baoxue Li and Zoujing Yao and Chunhui Zhao}, doi = {10.1016/j.neucom.2025.130890}, journal = {Neurocomputing}, month = {10}, pages = {130890}, shortjournal = {Neurocomputing}, title = {Multi-layer federated learning for bit-width and data heterogeneity in cloud-edge systems}, volume = {651}, year = {2025}, }
• (2025). Latent low-rank tensor wheel decomposition for visual data completion. NEUCOM, 651, 130889. (www)

  @article{ , abstract = {Recently, tensor wheel (TW) decomposition gains increasing attention in the area of low-rank tensor completion (LRTC). Existing tensor factorization-based methods can either capture the global connections among all dimension-pairs of data or the local connections between adjacent modes only via low-rank regularization. In this paper, we propose a novel TW decomposition with latent low-rank factors, where the low-rank regularizations are incorporated in the gradient domain of ring factors to enhance the robustness of TW-ranks. Thus, the global low-rank structure of TW decomposition and local continuity of high-order tensors can be exploited in a unified framework. Additionally, an efficient alternating direction method of multipliers (ADMM) algorithm is developed to solve the optimization. Experimental results on real-world visual data such as color images, multispectral images (MSI), and video sequences have showcased the superiority of the proposed method.}, archive = {J_NEUCOM}, author = {Yihao Luo and Yuning Qiu and Peilin Yang and Hongxia Rao and Zhenhao Huang and Guoxu Zhou}, doi = {10.1016/j.neucom.2025.130889}, journal = {Neurocomputing}, month = {10}, pages = {130889}, shortjournal = {Neurocomputing}, title = {Latent low-rank tensor wheel decomposition for visual data completion}, volume = {651}, year = {2025}, }
• (2025). Modified BiLSTM network for interval prediction based on aerospace load system. NEUCOM, 651, 130887. (www)

  @article{ , abstract = {In structural health monitoring, load identification significantly impacts the safety and reliability of structures, particularly in aerospace, maritime, and other industrial systems. To address the issues of low accuracy and unstable signal identification in airborne load recognition, this paper proposes a novel neural network-based load prediction model called Encoding Meta Attention-Adaptive Snow Ablation Optimizer-BiLSTM (EMA-ABiLSTM). First, a novel meta-attention mechanism is designed to directly train the source data and output weights to the prediction model. The features generated by positional encoding and meta-network are used as inputs to the prediction model. Second, an Adaptive Improved Snow Ablation Optimizer (ASAO) algorithm is proposed to optimize the hyperparameters of the BiLSTM network, forming the Adaptive Snow Ablation Optimizer-BiLSTM (ABiLSTM) model. Finally, both feature input and raw input data are used to train the ABiLSTM model for high-precision load identification. An interval prediction method is introduced to quantify the uncertainty of predictions, thereby enhancing the reliability of the results. Additionally, a novel meta-network feature loss function is designed to improve the model’s identification efficiency. In experiments conducted using real landing gear data from an aircraft, the proposed method is compared with several advanced prediction models. The results demonstrate that the EMA-ABiLSTM model achieves excellent accuracy and reliability in interval prediction and load identification tasks.}, archive = {J_NEUCOM}, author = {Xiaoli Li and Tongming Huo and Lianqing Zhu and Kang Wang}, doi = {10.1016/j.neucom.2025.130887}, journal = {Neurocomputing}, month = {10}, pages = {130887}, shortjournal = {Neurocomputing}, title = {Modified BiLSTM network for interval prediction based on aerospace load system}, volume = {651}, year = {2025}, }
• (2025). Further results on global stability of clifford-valued neural networks subject to time-varying delays. NEUCOM, 651, 130886. (www)

  @article{ , abstract = {This paper investigates the global exponential and asymptotic stability of Clifford-valued neural networks (CLVNNs) with multiple time-varying delays. Due to the non-commutative nature of Clifford algebra, analyzing the stability and other dynamical properties of CLVNNs becomes challenging. To address this issue, we separate the CLVNNs into equivalent real-valued neural networks (RVNNs). This separation simplifies the study of CLVNNs through their RVNN components. By constructing a suitable Lyapunov–Krasovskii functionals (LKFs) and applying inequality techniques, we establish several sufficient conditions that guarantee the existence and uniqueness of the equilibrium point (EP), as well as the global exponential and asymptotic stability of the considered neural networks (NNs). These conditions are expressed as linear matrix inequalities (LMIs), which can be efficiently verified using MATLAB LMI toolbox. To validate the analytical results, we present three numerical examples. Additionally, we propose a novel color image encryption algorithm, and demonstrate its effectiveness through simulation results and detailed performance analysis.}, archive = {J_NEUCOM}, author = {N. Manoj and R. Sriraman and R. Gurusamy and Yilun Shang}, doi = {10.1016/j.neucom.2025.130886}, journal = {Neurocomputing}, month = {10}, pages = {130886}, shortjournal = {Neurocomputing}, title = {Further results on global stability of clifford-valued neural networks subject to time-varying delays}, volume = {651}, year = {2025}, }
• (2025). PHiD: Preserving human identity in pose-guided character animation. NEUCOM, 651, 130885. (www)

  @article{ , abstract = {Generating pose-guided human animation videos is a challenging task, particularly in maintaining consistent facial identity (ID) between the generated video and the reference image. Despite significant advancements in diffusion-based human animation models, existing methods, which mainly rely on basic conditioning mechanisms, often struggle with facial consistency and realism, especially when the face occupies a small region in the reference. This paper proposes a facial-area-aware approach, PHiD , designed to enhance facial ID similarity while ensuring strong structural and temporal coherence. Specifically, we propose a Pose-Driven Face Morphing module that leverages the 3D Morphable Model to synthesize proxy faces based on the reference ID and target pose, generating multi-view features to enhance temporal consistency. Additionally, we introduce a Masked Face Adapter (MFA) that embeds the proxy face and employs masked attention on facial regions to capture and refine localized facial features accurately. To enable effective training of MFA, we design a Facial ID-Preserving Loss that combines feature similarity, reconstruction, and pose consistency terms. Notably, our method demonstrates strong generalization capabilities and can be seamlessly integrated into existing pose-guided image-to-video models. Extensive experiments show that our approach outperforms baseline methods in generating human animation videos with improved facial consistency and similarity.}, archive = {J_NEUCOM}, author = {Wenjie Zheng and Xingze Zou and Lianrui Mu and Jing Wang and Jiaqi Hu and Jiangnan Ye and Jiedong Zhuang and Mudassar Ali and Olumayowa Idowu and Haoji Hu}, doi = {10.1016/j.neucom.2025.130885}, journal = {Neurocomputing}, month = {10}, pages = {130885}, shortjournal = {Neurocomputing}, title = {PHiD: Preserving human identity in pose-guided character animation}, volume = {651}, year = {2025}, }
• (2025). PiMAN: A physics-informed motion prediction network using sEMG signal features for human movement parameters. NEUCOM, 651, 130884. (www)

  @article{ , abstract = {Early and accurate prediction of human movement parameters is critical for assistive robotics to synchronize effectively with a user’s intent. Surface electromyography (sEMG) signals offer a unique advantage by capturing neuromuscular activity prior to visible motion; however, existing model-based and model-free approaches often suffer from limited generalizability, delayed response, or poor biomechanical interpretability. To address these limitations, we propose PiMAN (Physics-informed Motion Anticipation Network), a deep learning framework that combines an attention-based bidirectional gated recurrent unit (BiGRU) architecture with physics constraints derived from the inverse dynamics. The model incorporates subject-specific anthropometric hyperparameters into the inverse dynamics formulation, enabling biomechanically consistent torque estimation across individuals. PiMAN predicts a comprehensive set of joint parameters, including angles, velocities, accelerations, external payloads, and torques, 48–96 ms before visible movement onset, from sEMG windows aligned with electromechanical delay range. This supports real-time control in assistive and neuroprosthetic systems. The model was trained and evaluated on five test subjects under three external load conditions (0 kg, 2 kg, and 4 kg), using both intra- and inter-subject scenarios. It achieved low RMSE ( ≤ 1.3) and high correlation (up to 0.93) across all outputs. Compared to purely data-driven baselines and physics-informed variants lacking attention, PiMAN consistently outperforms in joint torque and load estimation, particularly under higher-load conditions. In addition, PiMAN generalizes to temporally varying load transitions without retraining, and treats external mass as a continuous variable to facilitate seamless integration into inverse dynamics. These findings position PiMAN as a scalable, generalizable, and real-time-ready framework for anticipatory motion prediction in wearable assistive technologies.}, archive = {J_NEUCOM}, author = {Rajnish Kumar and Anand Gupta and Suriya Prakash Muthukrishnan and Lalan Kumar and Sitikantha Roy}, doi = {10.1016/j.neucom.2025.130884}, journal = {Neurocomputing}, month = {10}, pages = {130884}, shortjournal = {Neurocomputing}, title = {PiMAN: A physics-informed motion prediction network using sEMG signal features for human movement parameters}, volume = {651}, year = {2025}, }
• (2025). Hierarchical reasoning enhanced few-shot multimodal sentiment analysis. NEUCOM, 651, 130883. (www)

  @article{ , abstract = {Few-shot Multimodal Sentiment Analysis (FMSA) aims to predict sentiment with minimal labeled data by integrating multiple modalities, such as text and images. While recent FMSA methods have focused on transforming non-linguistic information (e.g., images) into text and leveraging language models to convert them into few-shot filling tasks, they still struggle to capture the latent sentiment information in image–text pairs. These limitations hinder their effectiveness, particularly in real-world applications where labeled data is scarce. To address these limitations, we propose a novel approach, Hierarchical Reasoning Enhanced Few-shot Multimodal Sentiment Analysis (HRE-FMSA), which consists of three main components: the Hierarchical Reasoning Framework (HRF), the Hierarchical Reasoning Representation Fusion Network (H2RF-Net), and label prediction. Concretely, the HRF module excavates latent sentiment information from image–text pairs at three levels: topic/aspect, opinion, and sentiment. Then, H2RF-Net integrates latent sentiment information with the original image–text pairs to generate a prompt, which is fed into a pre-trained Language Model to obtain the final sentiment type. In the experiment, we conducted comprehensive evaluations on three sentence-level datasets and two aspect-level datasets, demonstrating the effectiveness and applicability of HRE-FMSA.}, archive = {J_NEUCOM}, author = {Jiali You and Haoran Li and Jiawen Deng and Wei Li and Yuanyuan He and Fuji Ren}, doi = {10.1016/j.neucom.2025.130883}, journal = {Neurocomputing}, month = {10}, pages = {130883}, shortjournal = {Neurocomputing}, title = {Hierarchical reasoning enhanced few-shot multimodal sentiment analysis}, volume = {651}, year = {2025}, }
• (2025). Exploring interaction concepts for human–object-interaction detection via global- and local-scale enhancing. NEUCOM, 651, 130882. (www)

  @article{ , abstract = {Understanding the interactions between human–object (HO) pairs is the key to the human–object interaction (HOI) detection task. Visual understanding research has been significantly impacted by recent advances in linguistic-visual contrastive learning. For HOI detection studies, the alignment of linguistic and visual features is usually required to be performed when linguistic knowledge is used for enhancement. This usually results in the demands of extra training data or extended training time. In this study, an effective approach for utilizing multimodal knowledge to enhance HOI learning from global and instance scales is proposed. Model performance on Rare HOI categories can be prominently improved by using projection guided by linguistic knowledge at a global scale and merging multimodal features at an instance scale. State-of-the-art performance on the HICO-Det benchmark is achieved by the proposed model, and the effectiveness of the proposed global- and local-scale multimodal learning approach is validated.}, archive = {J_NEUCOM}, author = {Tianlun Luo and Qiao Yuan and Boxuan Zhu and Steven Guan and Rui Yang and Jeremy S. Smith and Eng Gee Lim}, doi = {10.1016/j.neucom.2025.130882}, journal = {Neurocomputing}, month = {10}, pages = {130882}, shortjournal = {Neurocomputing}, title = {Exploring interaction concepts for human–object-interaction detection via global- and local-scale enhancing}, volume = {651}, year = {2025}, }
• (2025). Reducing hubness to improve inductive few-shot learning. NEUCOM, 651, 130879. (www)

  @article{ , abstract = {Few-Shot Learning (FSL) provides a promising paradigm that completes a task using a few labeled samples, benefiting object recognition, outlier detection, and various tasks. However, distance-based FSL is vulnerable to the hubness problem, i.e., a few points (hubs from one class) occur frequently among the nearest neighbors of other points (from other classes) in the high-dimensional space. Therefore, the hubs commonly mislead classifiers resulting in a considerable performance degradation. The major body of current solutions is not designed to approach zero hubness, or limited to the transductive paradigm. As a countermeasure, this work aims to approach zero hubness for inductive FSL explicitly. We propose a new optimization objective that restricts the embeddings to approach zero hubness. A holistic Hubness rEDucing (HED) method is proposed to embed representations on the hypersphere and maintain linear separability, resulting in the decrease of hubness and preservation of class structure. A calibration mechanism is further devised to mitigate the negative impact of FSL’s data limitation, which calibrates the potentially biased distribution of embeddings. Extensive evaluation results demonstrate that our proposed method reduces hubness to improve the inductive FSL, and meanwhile, is compatible with a wide range of backbones.}, archive = {J_NEUCOM}, author = {Wenyi Tang and Haocheng Pei and Xin Wang and Zaobo He and Lei Yu and Xinsong Yang}, doi = {10.1016/j.neucom.2025.130879}, journal = {Neurocomputing}, month = {10}, pages = {130879}, shortjournal = {Neurocomputing}, title = {Reducing hubness to improve inductive few-shot learning}, volume = {651}, year = {2025}, }
• (2025). UniCuboid: Cuboid-based dense shape supervision for monocular 3D object detection. NEUCOM, 651, 130873. (www)

  @article{ , abstract = {Monocular 3D object detection aims to estimate the object’s 3D cuboid from a single RGB image. It is an ill-posed problem due to the inherent lack of depth information. Recent methods predict 3D shape points w.r.t. 2D grids and solve 3D pose with a PnP solver. However, it is difficult to obtain ground truth supervision for 3D shape in driving scenarios. Existing solutions typically require extra LiDAR data or CAD models to obtain 3D shape points, limiting their applicability. Besides, the 3D parameters estimation processes are highly coupled within existing differentiable PnP solvers, hindering the convergence to accurate solutions. In this paper, we propose UniCuboid, a cuboid-based dense shape supervision method for monocular 3D object detection. It contains a new cuboid-based dense shape supervision (CDSS) for shape learning and a PnP solver with a closed-form solution (CF-PnP) for pose estimation. CDSS includes a new cuboid-based dense 3D shape representation and shape supervision to guide 3D shape learning. The new 3D shape representation consists of 3D points positioned on the cuboid surface rather than the object’s surface. The learned 3D shape and weight maps are then sent to our PnP solver for 3D pose. It decouples 3D parameters estimation and offers a closed-form solution for dense 2D-3D point correspondence optimization. The proposed cuboid-based dense shape supervision and weighted PnP solver work together, constraining the 2D grids and enabling the adaptive tuning of the combination of 3D shapes and weight maps. Experiments on KITTI and nuScenes datasets demonstrate the promising performance of UniCuboid compared with previous works.}, archive = {J_NEUCOM}, author = {Xiaoning Zhang and Yuanqi Su and Haoyue Shi and Haoang Lu and Yuehu Liu and Le Wang}, doi = {10.1016/j.neucom.2025.130873}, journal = {Neurocomputing}, month = {10}, pages = {130873}, shortjournal = {Neurocomputing}, title = {UniCuboid: Cuboid-based dense shape supervision for monocular 3D object detection}, volume = {651}, year = {2025}, }
• (2025). Trade-off analysis between finite-time synchronization and energy consumption for fractional-order two-layer neural networks. NEUCOM, 651, 130872. (www)

  @article{ , abstract = {This paper investigates the finite-time synchronization (FTS) of fractional-order two-layer neural networks with time delays and explores the energy consumption of their controllers, as well as the trade-off between the synchronization time cost and the controller energy consumption. The study makes the following key contributions: First, a sufficient criterion is derived to guarantee the FTS of fractional-order delayed complex networks by using a lemma based on fractional-order differential inequalities. This approach circumvents a potential methodological flaw present in some prior studies. Additionally, while existing research on FTS has primarily focused on fractional-order complex networks without time delays, this work extends the analysis to systems with time delays, thereby broadening the applicability of the results. Second, a switching controller is proposed to accelerate the synchronization of the error system. Third, a standardized evaluation function is introduced to analyze the trade-off between the FTS time cost and controller energy consumption in fractional-order systems, whereas previous research in this area has mostly focused on integer-order systems. Finally, the validity of the theoretical results is demonstrated through numerical simulations.}, archive = {J_NEUCOM}, author = {Tian Lan and Baoxian Wang and Jigui Jian and Kai Wu}, doi = {10.1016/j.neucom.2025.130872}, journal = {Neurocomputing}, month = {10}, pages = {130872}, shortjournal = {Neurocomputing}, title = {Trade-off analysis between finite-time synchronization and energy consumption for fractional-order two-layer neural networks}, volume = {651}, year = {2025}, }
• (2025). Homomorphic activation and argmax via bootstrapping and applications in privacy-preserving neural networks. NEUCOM, 651, 130862. (www)

  @article{ , abstract = {Privacy-preserving neural networks (PPNN) aim to provide inference as a service without breaching users’ privacy, which have many potential applications in the smart healthcare field. One promising technology to realize PPNN is fully homomorphic encryption (FHE), which supports homomorphic operations over the encrypted data. However, FHE-enabled PPNN is facing two major challenges. First, most FHE schemes only support arithmetic operations, but not non-linear operations (e.g., ReLU, Argmax). Second, most FHE-based PPNN schemes avoid using bootstrapping for the sake of efficiency, which, however, significantly limits the depth of neural networks. In order to execute non-linear operations over the encrypted data and extend them to neural networks with arbitrary depth, we adopt a bit-wise fully homomorphic encryption scheme (i.e. TFHE with bootstrapping), so that we are able to realize non-linear homomorphic activation functions and argmax directly. First, based on the gate bootstrapping and programmable bootstrapping in TFHE, Full-Domain Programmable Bootstrapping(FD_PBS) is realized, which provides a framework for computing arbitrary linear and non-linear functions in bootstrapping. Next, non-linear homomorphic activation functions are achieved by designing the corresponding test polynomial in FD_PBS, and non-linear homomorphic argmax function is advanced by homomorphic comparison using gate bootstrapping. Last, based on homomorphic activation and argmax, privacy-preserving neural networks with arbitrary depth are proposed, which yield high efficiency by performing activation during bootstrapping instead of running an extra activation layer. Experiments show that we can homomorphically evaluate non-linear activation and argmax directly; homomorphic activation enables the inference accuracy over the ciphertext to be close to that over the plaintext; and homomorphic argmax improves inference time by returning only the class with the largest predicted probability rather than all classes and their probabilities to the user.}, archive = {J_NEUCOM}, author = {Peng Zhang and Hengrui Lu and Tianbin Ding}, doi = {10.1016/j.neucom.2025.130862}, journal = {Neurocomputing}, month = {10}, pages = {130862}, shortjournal = {Neurocomputing}, title = {Homomorphic activation and argmax via bootstrapping and applications in privacy-preserving neural networks}, volume = {651}, year = {2025}, }
• (2025). Text-centric disentangled representation interaction network for multimodal sentiment analysis. NEUCOM, 651, 130857. (www)

  @article{ , abstract = {With the rise of short video content, Multimodal Sentiment Analysis (MSA) has gained significant attention as a research hotspot. However, the issue of heterogeneity has emerged as a major challenge in fusing these three modalities. While some recent studies have attempted to reduce this problem of heterogeneity by disentangling the modalities, they overlooked two critical issues. First, their approach treats all three modalities equally during disentanglement, overlooking the central role of the text modality in MSA. As the primary carrier of semantic and emotional information, the text modality serves as the backbone for sentiment interpretation and multimodal fusion. Additionally, after disentangling the modalities, they do not effectively leverage the unique features of each modality, relying instead on simple concatenation and Transformer to combine similar and dissimilar features. To fully harness the potential of text modality and the dissimilar features between the modalities, we propose a Text-centric Disentangled Representation Interaction Network (TDRIN), consisting of two main modules. In the Disentangled Representation Learning (DRL) module, we decompose representations from different modalities into separate sub-spaces centered around text modality, aiming to capture similar and dissimilar features among the modalities. Meanwhile, we utilize various constraints to learn better features and improve predictions. Additionally, to more effectively balance the similar and dissimilar features, we design the Disentangled Representation Fusion Network (DRFN) module, which fuses disentangled representations with text modality as the center, fully exploiting the correlations among disentangled representations. Extensive experiments on the CMU-MOSI, CMU-MOSEI, and CH-SIMS datasets demonstrate that TDRIN outperforms state-of-the-art methods across various metrics. Specifically, the F1 score surpasses the best-performing baseline by 3.19%, 0.96%, and 1.43% on the three datasets, respectively. Ablation studies further confirm the effectiveness of each module. Therefore, TDRIN effectively reduces the heterogeneity between modalities, resulting in improved performance in MSA tasks.}, archive = {J_NEUCOM}, author = {Jingming Hou and Nazlia Omar and Sabrina Tiun and Saidah Saad and Qian He}, doi = {10.1016/j.neucom.2025.130857}, journal = {Neurocomputing}, month = {10}, pages = {130857}, shortjournal = {Neurocomputing}, title = {Text-centric disentangled representation interaction network for multimodal sentiment analysis}, volume = {651}, year = {2025}, }
• (2025). A survey of data augmentation in named entity recognition. NEUCOM, 651, 130856. (www)

  @article{ , abstract = {Data augmentation (DA), initially prominent in Computer Vision (CV), has been successfully adapted to Natural Language Processing (NLP), proving effective in mitigating data scarcity problems in the context of few-shot settings or scenarios where deep learning techniques may underperform. Moreover, the primary goal of DA is to expand and diversify training datasets by different methods, enabling models to generate more diverse and high-quality sythetic data for training the NER models. This survey explored DA techniques in the context of Named Entity Recognition (NER), including linguistic features and four categories of data augmentation methods. Furthermore, we reviewed commonly used datasets in DA tasks, discussed some potential practical applications, and examined key challenges and future directions in DA for NER. These findings serve as a valuable reference for learners and offer insights for researchers. As an essential and cost-effective approach, DA alleviates data scarcity and overfitting in the NER models by facilitating the integration of diverse augmentation methods.}, archive = {J_NEUCOM}, author = {Yi Huang and Yuhan Gao and Chengjuan Ren}, doi = {10.1016/j.neucom.2025.130856}, journal = {Neurocomputing}, month = {10}, pages = {130856}, shortjournal = {Neurocomputing}, title = {A survey of data augmentation in named entity recognition}, volume = {651}, year = {2025}, }
• (2025). Learning nonseparable sparse regularizers via multivariate activation functions. NEUCOM, 651, 130853. (www)

  @article{ , abstract = {Sparse regularization is a widely embraced technique in high-dimensional machine learning and signal processing. Existing sparse regularizers, however, are predominantly hand-crafted and often separable, making them less adaptable to data and potentially hindering performance. In this paper, we present a novel approach aiming at learning nonseparable (multivariate) sparse regularizers. We leverage the proximal gradient algorithm to transform the challenge of acquiring nonseparable sparse regularizers into the task of learning multivariate activation functions. We further establish the necessary conditions that these activation functions should satisfy. Our contribution culminates in the introduction of MAF-SRL, a deep network designed to learn multivariate activation functions within existing deep learning frameworks. To our knowledge, this research marks the first endeavor to learn nonseparable sparse regularizers. Extensive experiments conducted on benchmark datasets underscore the superiority of regularizers learned through MAF-SRL. They exhibit significantly enhanced performance in terms of both accuracy and sparseness compared to existing sparse regularizers.}, archive = {J_NEUCOM}, author = {Xin Xu and Zhouchen Lin}, doi = {10.1016/j.neucom.2025.130853}, journal = {Neurocomputing}, month = {10}, pages = {130853}, shortjournal = {Neurocomputing}, title = {Learning nonseparable sparse regularizers via multivariate activation functions}, volume = {651}, year = {2025}, }
• (2025). ML-GAN: Multi-level text-driven fine-grained image generation using generative adversarial network. NEUCOM, 651, 130851. (www)

  @article{ , abstract = {Text-to-image generation aims to convert input text into semantically accurate and visually realistic images. Existing methods typically generate a general outline of an image using sentence-level text and then refine it with word-level text. However, this approach of jumping directly from coarse-grained to fine-grained text features makes it challenging for the model to refine images accurately. To address this issue, we propose a Multi-level Text-driven Fine-grained Image Generation using Generative Adversarial Networks (ML-GAN). The model leverages different hierarchical levels of text information to construct and optimize generated images progressively. We design a Dual-level Text Parallel Fusion Module (DPFM) and a Triple-level Text Parallel Fusion Module (TPFM) to precisely adjust and optimize image details by utilizing text information at different levels. Additionally, to enhance semantic consistency between text and generated images, we introduce a Cross-modal Attention Fusion Module in the discriminator to improve its ability to recognize text-image matching, thereby guiding the generator to produce images that better match text content. Compared to baseline models, our proposed model achieves improvements of 19.10% and 12.99% in FID scores on the CUB and COCO datasets, respectively. This validates the effectiveness of the multi-level text-to-image transformation approach in enhancing the quality of generated images.}, archive = {J_NEUCOM}, author = {Hong Zhao and He Wang and Yongjuan Yang}, doi = {10.1016/j.neucom.2025.130851}, journal = {Neurocomputing}, month = {10}, pages = {130851}, shortjournal = {Neurocomputing}, title = {ML-GAN: Multi-level text-driven fine-grained image generation using generative adversarial network}, volume = {651}, year = {2025}, }
• (2025). DCB-VIM: An ensemble learning based filter method for feature selection with imbalanced class distribution. NEUCOM, 651, 130848. (www)

  @article{ , abstract = {Feature selection aims to improve predictive performance and interpretability in the analysis of datasets with high dimensional feature spaces. Imbalanced class distribution can make the process of feature selection more severe. Robust methodologies are essential for dealing with this case. Therefore, we present a filter method based on ensemble learning, in which each classifier is built on randomly selected subspaces of features. Variable importance measure is computed based on a class-wise procedure within each classifier, and a feature weighting procedure is subsequently applied. The performance of classifiers is considered in the combination phase of the ensemble learning. Different choices of hyperparameters consisting of the subspace size and the number of classification trees are investigated through simulation studies for determining their effects on the predictive performance. The efficiency of the proposed method is evaluated with respect to predictive performance by different selection strategies based on real data analysis in the presence of class imbalance.}, archive = {J_NEUCOM}, author = {Nayiri Galestian Pour and Soudabeh Shemehsavar}, doi = {10.1016/j.neucom.2025.130848}, journal = {Neurocomputing}, month = {10}, pages = {130848}, shortjournal = {Neurocomputing}, title = {DCB-VIM: An ensemble learning based filter method for feature selection with imbalanced class distribution}, volume = {651}, year = {2025}, }
• (2025). Federated continual learning: Concepts, challenges, and solutions. NEUCOM, 651, 130844. (www)

  @article{ , abstract = {Federated Continual Learning (FCL) has emerged as a robust solution for collaborative model training in dynamic environments, where data samples are continuously generated and distributed across multiple devices. This survey provides a comprehensive review of FCL, focusing on key challenges such as heterogeneity, model stability, communication overhead, and privacy preservation. We explore various forms of heterogeneity and their impact on model performance. Solutions to non-IID data, resource-constrained platforms, and personalized learning are reviewed in an effort to show the complexities of handling heterogeneous data distributions. Next, we review techniques for ensuring model stability and avoiding catastrophic forgetting, which are critical in non-stationary environments. Privacy-preserving techniques are another aspect of FCL that have been reviewed in this work. This survey has integrated insights from federated learning and continual learning to present strategies for improving the efficacy and scalability of FCL systems, making it applicable to a wide range of real-world scenarios.}, archive = {J_NEUCOM}, author = {Parisa Hamedi and Roozbeh Razavi-Far and Ehsan Hallaji}, doi = {10.1016/j.neucom.2025.130844}, journal = {Neurocomputing}, month = {10}, pages = {130844}, shortjournal = {Neurocomputing}, title = {Federated continual learning: Concepts, challenges, and solutions}, volume = {651}, year = {2025}, }
• (2025). Mol-l2: Transferring text knowledge with frozen language models for molecular representation learning. NEUCOM, 651, 130837. (www)

  @article{ , abstract = {How to integrate abundant chemical text descriptions to produce expressive molecular representations is a compelling challenge. In this study, we propose a deep network architecture called Mol-L2, which aims to leverage powerful language models to transfer chemical text knowledge and enhance molecular representation learning. The main novelty of this work is the use of a two-stage training pipeline to align text and chemical spaces, where stage 1 pre-trains the language model using a specially constructed multi-objective loss, and stage 2 fine-tunes on molecular captioning. Subsequently, the output of the language model encoder is converted into a fixed-length text-enhanced embedding via a lightweight mapping network. Furthermore, a dedicated encoder containing information propagation of specific functional groups is designed to generate molecular initial representation and integrated with the text-enhanced embeddings using a weighted fusion module. Finally, the enhanced molecular representation is utilized for various downstream tasks through an additional output layer. The performance of the proposed Mol-L2 is tested on several standard benchmarks for molecular machine learning, including molecular properties prediction, drug-target interaction (DTI), and drug-drug interaction (DDI). Through comprehensive experiments, we demonstrate the merits and state-of-the-art performance of the Mol-L2 framework. Take blood–brain barrier penetration prediction, for instance, where Mol-L2 achieves the smallest prediction error, while the best comparison method is 91.8%, an improvement of 3.1%.}, archive = {J_NEUCOM}, author = {Maotao Liu and Qun Liu and Xu Gong and Yunsong Luo and Guoyin Wang}, doi = {10.1016/j.neucom.2025.130837}, journal = {Neurocomputing}, month = {10}, pages = {130837}, shortjournal = {Neurocomputing}, title = {Mol-l2: Transferring text knowledge with frozen language models for molecular representation learning}, volume = {651}, year = {2025}, }
• (2025). Emotion recognition systems with electrodermal activity: From affective science to affective computing. NEUCOM, 651, 130831. (www)

  @article{ , abstract = {Affective computing is an interdisciplinary field that aims to automatically recognize and interpret emotions. Recent research has focused on using physiological signals (e.g., electrodermal activity) to improve emotion recognition. However, the theoretical emotion models that underlie these systems have received comparatively little attention. We conducted a systematic review and meta-analysis on electrodermal-activity-based emotion-recognition systems. Our findings suggest that arousal prediction models outperform valence prediction models, supporting our preregistered hypothesis. This correlates with arousal's association with autonomic nervous system activity and its direct link to electrodermal activity. We also observed a mismatch between the machine-learning approaches most often used—chiefly classification models—and the predominantly dimensional emotion frameworks adopted in the literature. Specifically, although dimensional affective models are increasingly popular, there has not been a parallel rise in regression models that would better reflect the continuous nature of the underlying data. We conclude that a comprehensive understanding of affective states requires consideration of both psychological and computational perspectives in affective computing research.}, archive = {J_NEUCOM}, author = {Tomás Ariel D'Amelio and Lorenzo Ariel Galán and Emmanuel Alesandro Maldonado and Agustín Ariel Díaz Barquinero and Jerónimo Rodríguez Cuello and Nicolás Marcelo Bruno and Enzo Tagliazucchi and Denis-Alexander Engemann}, doi = {10.1016/j.neucom.2025.130831}, journal = {Neurocomputing}, month = {10}, pages = {130831}, shortjournal = {Neurocomputing}, title = {Emotion recognition systems with electrodermal activity: From affective science to affective computing}, volume = {651}, year = {2025}, }
• (2025). Improving distant supervised relation extraction with entity enhanced res-BiLSTM and self-rectified gate. NEUCOM, 651, 130830. (www)

  @article{ , abstract = {Distantly supervised relation extraction (DSRE) aims to extract relational facts between entity pairs from large-scale unstructured texts by aligning sentences with external knowledge bases. However, it is inevitable to introduce a large amount of noisy data in the process of labeling the data automatically by machine. Previous studies typically focus on removing noise instances by different types of attention techniques, ignoring the ineffectiveness of one-sentence and noisy bags. Existing sentence encoders typically rely on either CNNs for local features or LSTM/BERT models for global context, which fail to capture multi-scale semantic information and structural sentence details. To address this problem, we propose an Entity Pair Focused Denoising (namely EPFD) model for DSRE. Unlike traditional sentence embedding methods, our model considers multi-scale information via an entity-enhanced Res-BiLSTM module, which includes the global information, the local information, as well as the entity pair information all together in the same framework. To reduce the impact of noisy sentences, an element-wise gate obtained from each sentence is adopted to rectify the sentence itself and overcome the failure of attention mechanism. Experimental results on two widely used datasets show the significant superiority of our proposed approach over the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Wenbin Zhang and Zechen Meng and Mankun Zhao and Tianyi Xu and Jian Yu and Mei Yu}, doi = {10.1016/j.neucom.2025.130830}, journal = {Neurocomputing}, month = {10}, pages = {130830}, shortjournal = {Neurocomputing}, title = {Improving distant supervised relation extraction with entity enhanced res-BiLSTM and self-rectified gate}, volume = {651}, year = {2025}, }
• (2025). Front-door causal attention for unbiased panoptic scene graph generation. NEUCOM, 651, 130823. (www)

  @article{ , abstract = {Panoptic Scene Graph Generation (PSG) aims to comprehensively parse visual scenes into structured graphs, serving as an intermediate representation for downstream vision tasks. However, it suffers from severe bias due to long-tail predicate distributions, where the long-tail prior distorts visual feature learning. Existing methods that rely on re-sampling or re-weighting strategies fail to model the causal mechanism behind the data, thus they cannot address the root cause of bias: the spurious correlations between visual features and predicate labels caused by imbalanced data distribution. Causal intervention, as a method for identifying confounders from a causal inference perspective, is particularly effective in eliminating such implicit biases. In this paper, we propose a causal intervention framework based on Front-Door Attention(FDA) to achieve unbiased prediction. Specifically, we decouple the causal effect into Local Causal Attention (LCA) and Global Causal Attention (GCA): LCA models object-level interactions within a single entity pair, while GCA leverages a global dictionary compressed from the entire training set to eliminate dataset bias. The two attention modules work together to eliminate the visual feature bias caused by long-tail distribution. Experiments on PSG dataset demonstrate that our method achieves state-of-the-art performance, with significant improvements in mean recall (mR@50/mR@100) by 5.07% and 5.16% respectively, validating its effectiveness in mitigating long-tail distribution bias.}, archive = {J_NEUCOM}, author = {Shuang Liang and Long Zhang and Baihua Liu and Shuwei Yan and Hongming Zhu}, doi = {10.1016/j.neucom.2025.130823}, journal = {Neurocomputing}, month = {10}, pages = {130823}, shortjournal = {Neurocomputing}, title = {Front-door causal attention for unbiased panoptic scene graph generation}, volume = {651}, year = {2025}, }
• (2025). Robust observer and neurodynamics based distributed formation control of multiple mobile robots. NEUCOM, 651, 130818. (www)

  @article{ , abstract = {Formation control has been an interesting and ongoing problem in the cooperation of multiple robotic systems. This paper designs a novel bioinspired, observer-based formation control scheme to address challenges, including kinematic constraints, velocity jumps, incomplete state information, robustness against disturbances, and control smoothness under measurement noises. First, a distributed estimator is developed using a cascade design to provide the leader’s estimates without requiring derivative information. Next, a bioinspired state observer is introduced to estimate both unavailable states and disturbances accurately, enabling improved formation performance with smooth state measurements. Then, a neural-dynamic-based kinematic controller is proposed to effectively handle velocity jumps and kinematic constraints. Followed by the kinematic control, a nonlinear dynamic controller is designed to ensure robustness against disturbances. The proposed approach is validated through extensive simulation studies, demonstrating its effectiveness and confirming the theoretical results.}, archive = {J_NEUCOM}, author = {Zhe Xu and Tao Yan and Simon X. Yang and S. Andrew Gadsden and Mohammad Biglarbegian}, doi = {10.1016/j.neucom.2025.130818}, journal = {Neurocomputing}, month = {10}, pages = {130818}, shortjournal = {Neurocomputing}, title = {Robust observer and neurodynamics based distributed formation control of multiple mobile robots}, volume = {651}, year = {2025}, }
• (2025). Recurrent spiking neural networks as models of the entorhinal–hippocampal system for path integration: Grid cells and beyond. NEUCOM, 651, 130814. (www)

  @article{ , abstract = {Grid cells in the mammalian medial entorhinal cortex (MEC) play a pivotal role in coding spatial information and integrating self-motion, functioning as a spatial metric through multiscale periodic representations. To investigate the properties and functions of these grid codes, mechanistic models employ continuous attractor neural networks (CANNs) with hand-tuned connectivity and dynamics, while normative models use recurrent neural networks (RNNs) for path integration where hexagonal grid patterns emerge spontaneously. In this work, we develop recurrent spiking neural networks (RSNNs) with biologically realistic structures for the path integration task, generating more biologically plausible representations resembling those in the entorhinal–hippocampal system. Leveraging various spiking neuron models including a novel adaptive neuron model, the RSNNs achieve accurate and generalizable path integration performance comparable to prior normative models. Besides, the RSNNs exhibit the inherent formation of multimodular hexagonal grid patterns with more biologically plausible grid scale ratios, as well as toroidal topology and low-dimensional neural dynamics consistent with biological observations. Through experiments and analyses, the path-integrating RSNNs offer new insights into the fundamental mechanisms underlying mammals’ exceptional navigation abilities, paving the way for future research in biologically inspired navigation systems.}, archive = {J_NEUCOM}, author = {Ruilan Gao and Changjian Jiang and Yu Zhang}, doi = {10.1016/j.neucom.2025.130814}, journal = {Neurocomputing}, month = {10}, pages = {130814}, shortjournal = {Neurocomputing}, title = {Recurrent spiking neural networks as models of the entorhinal–hippocampal system for path integration: Grid cells and beyond}, volume = {651}, year = {2025}, }
• (2025). DeInfoReg: A decoupled learning framework with information regularization for better training throughput. NEUCOM, 651, 130813. (www)

  @article{ , abstract = {This paper introduces Decoupled Supervised Learning with Information Regularization (DeInfoReg), a novel approach that transforms a long gradient flow into multiple shorter ones, thereby mitigating the vanishing gradient problem. Integrating a pipeline strategy, DeInfoReg enables model parallelization across multiple GPUs, significantly improving training throughput. We compare our proposed method with standard backpropagation and other gradient flow decomposition techniques. Extensive experiments on diverse tasks and datasets demonstrate that DeInfoReg achieves superior performance and better noise resistance than traditional BP models and efficiently utilizes parallel computing resources. The code for reproducibility is available at: https://github.com/ianzih/Decoupled-Supervised-Learning-for-Information-Regularization/ .}, archive = {J_NEUCOM}, author = {Zih-Hao Huang and You-Teng Lin and Hung-Hsuan Chen}, doi = {10.1016/j.neucom.2025.130813}, journal = {Neurocomputing}, month = {10}, pages = {130813}, shortjournal = {Neurocomputing}, title = {DeInfoReg: A decoupled learning framework with information regularization for better training throughput}, volume = {651}, year = {2025}, }
• (2025). An investigation into creating counterfactual examples for non-linear support vector machines. NEUCOM, 651, 130809. (www)

  @article{ , abstract = {Counterfactual explanations provide interpretable insights into model decisions by identifying minimal changes to an instance that would lead to a different classification. This paper proposes a method for generating counterfactual explanations for non-linear Support Vector Machines (SVMs). Unlike prior approaches that rely on heuristic optimization or gradient-based methods, our approach leverages high-confidence examples as reference points, ensuring that counterfactuals are both realistic and reliably classified by the model. Our method guarantees the generation of a valid counterfactual for any given instance under mild conditions. We demonstrate its effectiveness through experiments on real-world tabular and image datasets, showing that it produces meaningful and interpretable counterfactuals across different domains under proximity and plausibility metrics.}, archive = {J_NEUCOM}, author = {Luca Bergamin and Fabio Aiolli}, doi = {10.1016/j.neucom.2025.130809}, journal = {Neurocomputing}, month = {10}, pages = {130809}, shortjournal = {Neurocomputing}, title = {An investigation into creating counterfactual examples for non-linear support vector machines}, volume = {651}, year = {2025}, }
• (2025). Two-stage model re-optimization and application in face recognition. NEUCOM, 651, 130805. (www)

  @article{ , abstract = {Face recognition technology has achieved mature performance in various fields and has been widely applied on edge devices with limited computational resources. However, in practical application scenarios, the deployment platforms are often predetermined, and the final deployed models are the result of trade-offs between the computational and storage capabilities of the platforms. These models are difficult to adjust once determined. Therefore, how to improve the performance of existing face recognition models without affecting the entire deployment process remains a topic of practical significance and research value. Additionally, with the rapid development of deep learning, there exists a vast array of high-performance open-source pre-trained models on the Internet. How to utilize these models efficiently is also a subject worthy of investigation. To address the aforementioned issues, this paper introduces a novel training framework called two-stage model re-optimization (TSMR). TSMR enhances face recognition performance by leveraging knowledge distillation, model re-parameterization and adversarial learning techniques. In particular, TSMR improves performance without increasing inference latency, modifying network architecture, or introducing additional computational cost. It is a versatile and effective framework that can be applied to any CNN-based recognition network. The experimental results show the effectiveness of TSMR in improving the performance of lightweight face recognition models. Across multiple datasets, including CALFW, CPLFW, YTF and AgeDB-30, TSMR achieves an average accuracy improvement ranging from 1% to 2%. Notably, when TSMR is applied with FDFaceNet as the baseline, it achieves an impressive accuracy of 95.58% on the LFW dataset, surpassing the state-of-the-art performance in lightweight face recognition networks.}, archive = {J_NEUCOM}, author = {Jianyu Qian and Shiyi Mu and Hengjie Lu and Shugong Xu}, doi = {10.1016/j.neucom.2025.130805}, journal = {Neurocomputing}, month = {10}, pages = {130805}, shortjournal = {Neurocomputing}, title = {Two-stage model re-optimization and application in face recognition}, volume = {651}, year = {2025}, }
• (2025). Individual fair fuzzy C-means clustering via density-adaptive spectral regularization. NEUCOM, 651, 130794. (www)

  @article{ , abstract = {Traditional clustering methods often overlook the importance of treating similar individuals fairly, which can result in biased or inconsistent groupings. In this work, we introduce a new approach called Individual Fair Fuzzy C-Means (IFCM), which aims to ensure that people or items with similar characteristics are grouped together in a consistent and equitable way. Our method achieves this by incorporating a fairness constraint into the clustering process, encouraging similar individuals to receive similar assignments while maintaining overall clustering quality. Through experiments on a variety of datasets, we demonstrate that IFCM can substantially improve fairness in clustering outcomes without sacrificing reliability. This work highlights the growing importance of fairness in data analysis and provides an accessible solution for real-world applications where equitable treatment is essential. To support reproducibility and further research, we have made our code publicly available at: https://github.com/Yanttk0319/IFCM .}, archive = {J_NEUCOM}, author = {Boyang Yan and Zhijing Yang and Junjie Zheng and Yiding Tang and Hui Zhang}, doi = {10.1016/j.neucom.2025.130794}, journal = {Neurocomputing}, month = {10}, pages = {130794}, shortjournal = {Neurocomputing}, title = {Individual fair fuzzy C-means clustering via density-adaptive spectral regularization}, volume = {651}, year = {2025}, }
• (2025). Persistent spectral based machine learning method for autism spectrum disorder classification. NEUCOM, 651, 130780. (www)

  @article{ , abstract = {Background: Autism spectrum disorder (ASD) is a widespread and intricate neurodevelopmental condition. The increasing prevalence of ASD creates a very significant burden on both society and families. Functional magnetic resonance imaging (fMRI) contributes to a deeper understanding of ASD while also facilitating the development of early diagnosis and effective treatment strategies. This study aims to provide new and more reliable tools for early diagnosis of ASD and gain deeper insights into its neural mechanisms through the combination of topology and persistent spectral theory with functional connectivity. Methods: We proposed a persistent spectral machine learning model based on the simplicial complex for characterizing the functional connectivity in the Autism Brain Imaging Data Exchange I dataset. Simplicial complexes were used to characterize the functional connectivity with coefficients no less than 0.3. We arranged a filtration value for each simplex and persistent Laplacian matrices were calculated through a filtration process. The corresponding persistent attributes, after removing covariates, were used as inputs of classifiers. Results: Achieving an accuracy of 87.5%, our model outperformed other models that applied functional connectivity, similar sample sizes and the same preprocessing pipelines. We found that the numbers and distribution of connected components and loops of the global functional connectivity are important for classification. Conclusions: This study provided a feature extraction method based on persistent spectral theory for ASD research. Our model offers a different perspective on the research of related conditions and has great and notable potential in diagnosis.}, archive = {J_NEUCOM}, author = {Xudong Zhang and Liyuan Ma and Yaru Gao and Yunge Zhang and Fengling Li and Fengchun Lei}, doi = {10.1016/j.neucom.2025.130780}, journal = {Neurocomputing}, month = {10}, pages = {130780}, shortjournal = {Neurocomputing}, title = {Persistent spectral based machine learning method for autism spectrum disorder classification}, volume = {651}, year = {2025}, }
• (2025). Dual-model spiking neural network for remote sensing image classification using mutual knowledge distillation. NEUCOM, 651, 130721. (www)

  @article{ , abstract = {With the increasing demand for low-power models in edge applications of remote sensing image classification, Spiking Neural Networks (SNNs) have garnered widespread attention due to their low-power neuromorphic computing characteristics. However, existing SNNs typically require longer time steps to achieve effective information processing, resulting in a decrease in inference speed. Therefore, how to reduce performance loss while decreasing time steps has become a significant challenge. To address this issue, this paper proposes a new framework called Long–Short Step Dual-Model Mutual Distillation Joint Training. In this framework, two spiking neural networks with different time steps are collaboratively trained, mutually distilling to reduce the model’s reliance on time steps, thereby alleviating the performance loss during the inference phase caused by using shorter time steps. Additionally, this paper introduces a smooth random time step mechanism, which gradually reduces the time step of the short-step model in a smooth manner to minimize the information loss caused by the reduction of time steps. Experimental results show that, at the same time step, our method reduces accuracy loss by nearly 3% compared to a single model.}, archive = {J_NEUCOM}, author = {Zhen Cao and Ju Liu and Qi Sun and Biao Hou and Hao Zhu and Lei Guo and Yifan Ge and Yulong Wang and Licheng Jiao}, doi = {10.1016/j.neucom.2025.130721}, journal = {Neurocomputing}, month = {10}, pages = {130721}, shortjournal = {Neurocomputing}, title = {Dual-model spiking neural network for remote sensing image classification using mutual knowledge distillation}, volume = {651}, year = {2025}, }
• (2025). Embedding similarity guided license plate super resolution. NEUCOM, 651, 130657. (www)

  @article{ , abstract = {Super-resolution (SR) techniques play a pivotal role in enhancing the quality of low-resolution images, particularly for applications such as security and surveillance, where accurate license plate recognition is crucial. This study proposes a novel framework that combines pixel-based loss with embedding similarity learning to address the unique challenges of license plate super-resolution (LPSR). The introduced pixel and embedding consistency loss (PECL) integrates a Siamese network and applies contrastive loss to force embedding similarities to improve perceptual and structural fidelity. By effectively balancing pixel-wise accuracy with embedding-level consistency, the framework achieves superior alignment of fine-grained features between high-resolution (HR) and super-resolved (SR) license plates. Extensive experiments on the CCPD and PKU dataset validate the efficacy of the proposed framework, demonstrating consistent improvements over state-of-the-art methods in terms of PSNR, SSIM, LPIPS, and optical character recognition (OCR) accuracy. These results highlight the potential of embedding similarity learning to advance both perceptual quality and task-specific performance in extreme super-resolution scenarios.}, archive = {J_NEUCOM}, author = {Abderrezzaq Sendjasni and Mohamed-Chaker Larabi}, doi = {10.1016/j.neucom.2025.130657}, journal = {Neurocomputing}, month = {10}, pages = {130657}, shortjournal = {Neurocomputing}, title = {Embedding similarity guided license plate super resolution}, volume = {651}, year = {2025}, }
• (2025). Dual-markov dynamic event-triggered synchronization of neural networks under asynchronous cyber attacks. NEUCOM, 651, 130597. (www)

  @article{ , abstract = {This paper investigates the synchronization of master–slave neural networks under asynchronous cyber attacks. First, a novel asynchronous cyber attack strategy is proposed, governed by two independent Markov jump processes. Next, a dual-Markov dynamic event-triggered mechanism is introduced, which saves communication resources and enhances defense against attacks. Additionally, a novel asymmetric Lyapunov–Krasovskii function is proposed, combining the reciprocal convex combinatorial inequality and the zero-equation method to reduce the conservatism of the stability criteria. An adaptive Markov jump event-triggered controller is then designed to achieve synchronization of the master–slave neural networks. Finally, two synchronization criteria are derived, and two sets of values are provided to validate these criteria.}, archive = {J_NEUCOM}, author = {Zou Yang and Jun Wang and Kaibo Shi and Xiangxiang Wang and Yiqian Tang}, doi = {10.1016/j.neucom.2025.130597}, journal = {Neurocomputing}, month = {10}, pages = {130597}, shortjournal = {Neurocomputing}, title = {Dual-markov dynamic event-triggered synchronization of neural networks under asynchronous cyber attacks}, volume = {651}, year = {2025}, }
• (2025). BFP: Balanced filter pruning via knowledge distillation for efficient deployment of CNNs on edge devices. NEUCOM, 650, 130946. (www)

  @article{ , abstract = {Model pruning can reduce the computational cost of convolutional neural networks (CNNs), which enables CNNs to be deployed on edge devices with limited computational resources. However, most existing CNN pruning methods rely on a single global indicator to evaluate the importance of filters, ignoring local feature redundancy, which can easily lead to the loss of key information and affect the performance recovery and generalization of the pruning model. In light of this circumstance, a novel balanced filter pruning (BFP) method is proposed in this paper, connecting global measurement with focused attention. First, the method utilizes the BN layer scaling coefficient to perform global channel evaluation and mines local information redundancy through feature map correlation to achieve dynamic balance between structure compression and information preservation. Next, the weighting of the above two indicators is used as a balanced indicator for assessing the importance of the filters, and pruning is performed according to the set pruning rate. Finally, knowledge distillation is used to compensate for the loss of performance caused by the pruning network, which makes the method show better application prospects in scenarios such as edge computing. The effectiveness of the proposed method is validated on two image classification datasets. For example, for the ResNet-50 on the ImageNet dataset, BFP achieves a 59.2 % reduction in float-point-operations (FLOPs) and a 47.1 % reduction in parameters, and the Top-1 accuracy and Top-5 accuracy of the model only lose 0.52 % and 0.35 %, respectively.}, archive = {J_NEUCOM}, author = {Haikun Zhang and Yajun Liu}, doi = {10.1016/j.neucom.2025.130946}, journal = {Neurocomputing}, month = {10}, pages = {130946}, shortjournal = {Neurocomputing}, title = {BFP: Balanced filter pruning via knowledge distillation for efficient deployment of CNNs on edge devices}, volume = {650}, year = {2025}, }
• (2025). Low-rank kernel consistent multi-view subspace clustering. NEUCOM, 650, 130944. (www)

  @article{ , abstract = {Multi-view subspace clustering aims to enhance data analysis by integrating information from diverse perspectives. However, existing methods often fail to simultaneously leverage both consistent and complementary information across different views, and most methods ignore the nonlinear structure in the data. In this study, we propose a novel low-rank kernel consistent multi-view subspace clustering algorithm named LRKCMS. Specifically, we factorize the coefficient matrices into three distinct factors to maintain rank consistency across different views. Concurrently, we explore the self-representative coefficient matrices of each view to fully harness their complementary information. To mitigate the influence of noise and outliers in the underlying data, we apply the Schatten p-norm to regularize the low-rank coefficient matrices. Additionally, to fully exploit the intrinsic nonlinear characteristics of the data, we introduce a kernel mapping term that adaptively solves the low-rank kernel mapping, ensuring that the data in the resulting kernel feature space are both low-rank and self-expressive. We evaluate the performance of LRKCMS on 10 benchmark datasets and compare it against 12 state-of-the-art methods. The results demonstrate that LRKCMS achieves superior clustering performance compared to these advanced multi-view clustering methods. The source code is available at https://github.com/wzhangwhu/LRKCMS .}, archive = {J_NEUCOM}, author = {Wei Zhang and Yue Yu and Xiaoying Zheng and Juan Shen and Yuanyuan Li and Shiqi Wang and Zizhu Fan}, doi = {10.1016/j.neucom.2025.130944}, journal = {Neurocomputing}, month = {10}, pages = {130944}, shortjournal = {Neurocomputing}, title = {Low-rank kernel consistent multi-view subspace clustering}, volume = {650}, year = {2025}, }
• (2025). Finite-time and fixed-time bipartite synchronization of signed networks with mixed delays. NEUCOM, 650, 130942. (www)

  @article{ , abstract = {This paper investigates finite-time (FET) and fixed-time (FDT) bipartite synchronization of signed networks with time-varying and distributed delays (mixed delays) using quantized control strategies. The communication links between adjacent nodes can be either positive or negative, representing the signed nature of the network. Assuming a balanced network structure, sufficient conditions for FET and FDT bipartite synchronization are derived through coordinate transformations, norm analytical techniques, and differential inequalities. Finally, three simulation results are provided to validate the efficacy of the theoretical findings.}, archive = {J_NEUCOM}, author = {Yao Tan and Junjian Huang and Wei Zhang and Junren Wang and Shiping Wen and Tingwen Huang}, doi = {10.1016/j.neucom.2025.130942}, journal = {Neurocomputing}, month = {10}, pages = {130942}, shortjournal = {Neurocomputing}, title = {Finite-time and fixed-time bipartite synchronization of signed networks with mixed delays}, volume = {650}, year = {2025}, }
• (2025). EiCoM: Multi-modal knowledge graph completion with enhanced information completeness. NEUCOM, 650, 130940. (www)

  @article{ , abstract = {Knowledge graph completion (KGC) is essential for improving the integrity and quality of knowledge graphs (KGs), while multi-modal data offers a richer information source. However, multi-modal knowledge graph completion (MMKGC) faces challenges in information incompleteness and modality imbalance, limiting its performance and applicability. To address the above issues, this paper proposes a novel MMKGC method, EiCoM. Firstly, EiCoM proposes a cross-modal semantic alignment and collaboration module to efficiently align and interact with features across modalities, enhancing semantic consistency across multiple modalities. Then, EiCoM introduces a modality adversarial generation module based on the generative adversarial network (GAN) framework to generate supplementary modality information, alleviating information incompleteness and optimizing modality imbalance in MMKGC. Finally, EiCoM designs a context-aware adaptive fusion module with a dynamic weight allocation mechanism that adjusts the contribution weights of each modality based on the semantic information of a target triple and its context. Experimental results show that EiCoM achieves significant performance improvements on five public datasets, outperforming 19 state-of-the-art MMKGC baselines. Additionally, a series of experiments, including ablation studies, modal noise injection, a case study, and the analysis of the modality adversarial generation module, comprehensively validate the effectiveness of the EiCoM’s core modules and its completeness enhancement across different scenarios.}, archive = {J_NEUCOM}, author = {Lianhong Ding and Mengxiao Li and Peng Shi and Juntao Li and Ruiping Yuan}, doi = {10.1016/j.neucom.2025.130940}, journal = {Neurocomputing}, month = {10}, pages = {130940}, shortjournal = {Neurocomputing}, title = {EiCoM: Multi-modal knowledge graph completion with enhanced information completeness}, volume = {650}, year = {2025}, }
• (2025). Joint domain transfer elasticity metric network for cross-domain small sample fault diagnosis. NEUCOM, 650, 130936. (www)

  @article{ , abstract = {Inadequate target domain samples are successfully addressed by transfer learning, but it suffers when there are also inadequate source domain samples. Meta-learning offers a novel solution to the small-sample problem by enhancing model generalization through multi-task learning. However, current meta-transfer research has not adequately considered the domain distribution differences resulting from changes in operating conditions and the problem of insufficient source domain samples, while also facing the challenge of class confusion caused by feature complexity and fuzzy class boundaries. To overcome these difficulties, this study proposes a joint domain transfer elasticity metric network (JDTEMN) to solve the cross-domain small sample fault diagnosis problem. First, a joint cross-domain transfer structure is introduced, utilizing maximum mean square difference (MMSD) with domain discriminators to match source and target domain features. This approach leverages sample mean and variance for accurate feature matching and incorporates adversarial training to capture domain-specific features, enabling more effective transfer. Second, the metric-based meta learning framework is integrated with the joint cross-domain transfer structure to address the small sample problem. To resolve class confusion in the metric network, an elasticity factor is introduced to reduce intraclass spacing and increase interclass spacing, enabling effective cross-domain small-sample fault diagnosis. To enhance the network’s learning of fault features, the CBAM attention mechanism is incorporated into JDTEMN, further improving diagnostic accuracy. Finally, the effectiveness of JDTEMN is validated through experiments on public and private bearing datasets, demonstrating superior diagnostic performance compared to other methods.}, archive = {J_NEUCOM}, author = {Zhiwu Shang and Xiaolong Du and Cailu Pan and Fei Liu and Ziyu Wang}, doi = {10.1016/j.neucom.2025.130936}, journal = {Neurocomputing}, month = {10}, pages = {130936}, shortjournal = {Neurocomputing}, title = {Joint domain transfer elasticity metric network for cross-domain small sample fault diagnosis}, volume = {650}, year = {2025}, }
• (2025). Evolutionary semi-supervised regression for blast furnace stability prediction. NEUCOM, 650, 130935. (www)

  @article{ , abstract = {The stability of blast furnace operating state is a major concern for practitioners in the field. However, due to the difficulty in obtaining labels required for this task, many research efforts have resorted to indirectly reflecting the thermal state and stability of blast furnaces through the silicon content in hot metal. To address the challenge, this paper proposes an Evolutionary Semi-Supervised Regression (ESSR) method, which combines the self-training framework and evolutionary computation to build an accurate stability prediction model. Starting from an initial supervised model, ESSR achieves joint optimization of pseudo-labeled data and the semi-supervised model by incrementally updating the latter. Furthermore, a method for training the initial model based on knowledge extraction is proposed to improve its prediction accuracy. Experiments show that ESSR enables the semi-supervised model to continuously acquire useful new knowledge from unlabeled data, and the obtained model significantly outperforms other comparative models. The initial model training method, on the other hand, improves the accuracy of the initial model for the stability prediction problem and can provide a better starting point for subsequent semi-supervised learning. Compared to the standard supervised model, the semi-supervised model trained through the proposed methods achieves approximately 23 % performance improvement in the stability prediction problem.}, archive = {J_NEUCOM}, author = {Tenghui Hu and Xianpeng Wang and Xiangman Song and Xiaoling Zhang}, doi = {10.1016/j.neucom.2025.130935}, journal = {Neurocomputing}, month = {10}, pages = {130935}, shortjournal = {Neurocomputing}, title = {Evolutionary semi-supervised regression for blast furnace stability prediction}, volume = {650}, year = {2025}, }
• (2025). Multi-scale convolutional sparse attention transformer: A lightweight fault diagnosis model for rotating machinery. NEUCOM, 650, 130934. (www)

  @article{ , abstract = {Prolonged operation of rotating machinery often results in mechanical faults, posing challenges to the reliability and efficiency of industrial systems. Transformer models can enhance diagnostic accuracy by effectively capturing global dependencies within signals. However, they face the limitations due to high computational costs and overfitting with limited data. Consequently, this study proposes the multi-scale convolutional sparse attention Transformer (MCSAT), a lightweight and efficient model for fault diagnosis in rotating machinery. MCSAT incorporates a pyramid convolution patch embedding module that uses multi-scale convolutional kernels to extract fine-grained features and perform multi-level fusion for 1D time-series vibration signals. A depthwise separable convolution layer is introduced to reduce spatial and channel redundancy, minimizing parameter count and computational load while maintaining high performance. Furthermore, a dynamic sparse attention Transformer block is employed, replacing conventional multi-head attention with bi-level routing attention to dynamically focus on the most informative features, thereby reducing memory usage and computational complexity. Extensive experiments on SQV and XJTU datasets demonstrate the superiority of MCSAT. The proposed model achieves a diagnostic accuracy of 99.69 % on SQV dataset and maintains high performance under limited sample conditions on XJTU gearbox dataset, with accuracy ranging from 97.94 % to 99.75 % for training sample sizes between 10 and 100 per class. Additionally, MCSAT requires only 0.065 million parameters and 8.607 million FLOPs. These results highlight the potential of MCSAT for real-time fault diagnosis in resource-constrained industrial environments. The implementation code is available at: https://github.com/ZHANG-JiXiang/Lightweight-Transformer-for-fault-diagnosis .}, archive = {J_NEUCOM}, author = {Jixiang Zhang and Mengjian Zhang and Deguang Wang and Ming Yang and Chengbin Liang}, doi = {10.1016/j.neucom.2025.130934}, journal = {Neurocomputing}, month = {10}, pages = {130934}, shortjournal = {Neurocomputing}, title = {Multi-scale convolutional sparse attention transformer: A lightweight fault diagnosis model for rotating machinery}, volume = {650}, year = {2025}, }
• (2025). Leveraging inter-chunk interactions for enhanced retrieval in large language model-based question answering. NEUCOM, 650, 130931. (www)

  @article{ , abstract = {Retrieving external knowledge and prompting large language models with relevant information is an effective paradigm to enhance the performance of question-answering tasks. However, in multi-document question answering tasks, existing studies often evaluate the relevance of each paragraph to the question independently. This approach can lead to issues such as missing context and ambiguous references in the knowledge retrieved, ultimately hindering the reasoning capabilities of large language models. To address this limitation, we propose a novel retrieval framework called IIER in this study, which leverages Inter-chunk Interactions to Enhance Retrieval. In IIER, we transform independent documents into a unified Chunk-Interaction Graph by constructing edges that integrate structural, keyword, and semantic information between chunks. This graph effectively completes potential reasoning paths between chunks and provides additional contextual information to guide enhanced retrieval. Additionally, we design the Evidence Chain Retriever, which iteratively searches the graph for supporting evidence by jointly encoding semantic and topological information. Through this mechanism, the retriever is able to generate prompts enriched with comprehensive context and reasoning evidence chains, thereby improving the accuracy of large language models’ reasoning. Extensive experiments conducted on multiple datasets demonstrate that IIER significantly improves retrieval performance and enhances the reasoning capabilities of large language models in MDQA tasks. Specifically, IIER outperforms strong baseline models guided by large language models, achieving accuracy improvements of 9.6 % on the 2WikiMQA and 6.0 % on the HotpotQA.}, archive = {J_NEUCOM}, author = {Tiezheng Guo and Chen Wang and Yanyi Liu and Jiawei Tang and Pan Li and Sai Xu and Qingwen Yang and Xianlin Gao and Zhi Li and Yingyou Wen}, doi = {10.1016/j.neucom.2025.130931}, journal = {Neurocomputing}, month = {10}, pages = {130931}, shortjournal = {Neurocomputing}, title = {Leveraging inter-chunk interactions for enhanced retrieval in large language model-based question answering}, volume = {650}, year = {2025}, }
• (2025). Brain-inspired computing based on deep learning for human-computer interaction: A review. NEUCOM, 650, 130928. (www)

  @article{ , abstract = {The continuous development of artificial intelligence has a profound impact on biomedicine and other fields, providing new research ideas and technical methods. Brain-inspired computing is an important intersection between multimodal technology and biomedical field. Focusing on the application scenarios of decoding text and speech from brain signals in human-computer interaction, this paper presents a comprehensive review of the brain-inspired computing models based on deep learning (DL), tracking their evolution, application value, challenges and potential research trends. We first review its basic concepts and development history, and divide its evolution into two stages: recent machine learning and current deep learning, emphasizing the importance of each stage in the research of brain-inspired computing for human-computer interaction. In addition, the latest progress of deep learning in different tasks of brain-inspired computing for human-computer interaction is reviewed from five perspectives, including datasets and different brain signals, and the application of key technologies in the model is elaborated in detail. Despite significant advances in brain-inspired computational models, challenges remain to fully exploit their capabilities, and we provide insights into possible directions for future academic research. For more detailed information, please visit our GitHub page: https://github.com/ultracoolHub/brain-inspired-computing .}, archive = {J_NEUCOM}, author = {Bihui Yu and Sibo Zhang and Lili Zhou and Jingxuan Wei and Linzhuang Sun and Liping Bu}, doi = {10.1016/j.neucom.2025.130928}, journal = {Neurocomputing}, month = {10}, pages = {130928}, shortjournal = {Neurocomputing}, title = {Brain-inspired computing based on deep learning for human-computer interaction: A review}, volume = {650}, year = {2025}, }
• (2025). A novel context-adaptive multi-scale defect detection method for object surface defects. NEUCOM, 650, 130927. (www)

  @article{ , abstract = {In industrial processes, detecting surface defects is crucial for enhancing product quality and production efficiency. However, detecting defects at smaller scales remains challenging. A new end-to-end multi-scale surface defect detection method, termed as context-adaptive multi-scale defect detection (CAMDD), is proposed to address this issue. Firstly, a novel feature pyramid network, named the context and texture-extractor module with feature pyramid network (CTM-FPN), is introduced, incorporating a context extractor, texture extractor, and multi-scale residual atrous spatial pyramid pooling (MR-ASPP) to enhance the representation capability of defect features at different scales. Secondly, an adaptive threshold soft non-maximum suppression (ATS-NMS) post-processing algorithm is proposed to address the issue of interweaving defects of different scales, which may incorrectly suppress information about small defects, thereby reducing the model’s false negative rate. By integrating the two modules mentioned above into existing object detection methods, the CAMDD method is formed. Experimental results on two steel plate datasets and a dataset of tiny transparent lenses demonstrate that the proposed CAMDD method has advantages in accuracy for detecting defects of various scales.}, archive = {J_NEUCOM}, author = {Hongbin Zhang and Jianhua Liu and Dong Pan and Zhaohui Jiang and Jinzong Dong}, doi = {10.1016/j.neucom.2025.130927}, journal = {Neurocomputing}, month = {10}, pages = {130927}, shortjournal = {Neurocomputing}, title = {A novel context-adaptive multi-scale defect detection method for object surface defects}, volume = {650}, year = {2025}, }
• (2025). Compatible anti-forggetting network for re-indexing free lifelong person re-identification. NEUCOM, 650, 130925. (www)

  @article{ , abstract = {Re-Indexing Free Lifelong Person Re-Identification (RFL-ReID) focuses on addressing the issues of high storage costs and privacy risks in traditional lifelong person re-identification. Unlike conventional lifelong learning settings, RFL-ReID imposes strict restrictions on the use of historical data during model training. Existing approaches mainly rely on knowledge distillation to tackle this challenge. However, due to inevitable information loss during the distillation process and significant domain shifts across tasks—such as differences in viewpoints, resolutions, illumination conditions, and background contexts—the effectiveness of continuous distillation is severely limited. To address these limitations, we propose a novel network architecture—the Compatible Anti-Forggetting Network (CAF). To compensate for the reduced learning capacity caused by restricted access to historical data, we design a deep feature fusion module, which enables the model to extract richer and more multi-dimensional feature representations within a single learning phase, significantly enhancing its generalization ability under the re-indexing-free setting. Furthermore, to mitigate the negative impact of domain discrepancies on knowledge transfer, we introduce an adaptive network constraint module. This module dynamically optimizes the model update path, effectively regulating parameter changes and ensuring a smooth transition between old and new knowledge, thereby improving overall stability and robustness. Experimental results on multiple lifelong person re-identification benchmark datasets demonstrate that our method achieves substantial improvements over current state-of-the-art approaches, with an average mAP gain of 5.1 % and a Rank-1 accuracy improvement of 5.8 %, validating the superior memory retention and adaptation capabilities of CAF in long-term sequential learning tasks.}, archive = {J_NEUCOM}, author = {Boxu Ren and Yongping Xie}, doi = {10.1016/j.neucom.2025.130925}, journal = {Neurocomputing}, month = {10}, pages = {130925}, shortjournal = {Neurocomputing}, title = {Compatible anti-forggetting network for re-indexing free lifelong person re-identification}, volume = {650}, year = {2025}, }
• (2025). Curriculum learning-driven PIELMs for fluid flow simulations. NEUCOM, 650, 130924. (www)

  @article{ , abstract = {This paper presents two novel, physics-informed extreme learning machine (PIELM)-based algorithms for solving steady and unsteady nonlinear partial differential equations (PDEs) related to fluid flow. Although single-hidden-layer PIELMs outperform deep physics-informed neural networks (PINNs) in speed and accuracy for linear PDEs, their extension to nonlinear problems remains challenging. To address this, the two algorithms introduced leverage the same curriculum learning strategy that reformulates nonlinear PDEs as a sequence of increasingly complex linear PDEs. Additionally, our approach enables a physically interpretable initialization of network parameters by leveraging Radial Basis Functions (RBFs). The performance of the proposed algorithms is validated on two benchmark incompressible flow problems: the viscous Burgers’ equation and lid-driven cavity flow. To the best of our knowledge, this is the first work to extend PIELM to solving Burgers’ shock problem as well as lid-driven cavity flow up to a Reynolds number of 100. As a practical application, we employ PIELM to predict blood flow in a stenotic vessel. The results confirm that PIELM efficiently handles nonlinear PDEs, positioning it as a promising alternative to PINNs for both linear and nonlinear PDEs.}, archive = {J_NEUCOM}, author = {Vikas Dwivedi and Bruno Sixou and Monica Sigovan}, doi = {10.1016/j.neucom.2025.130924}, journal = {Neurocomputing}, month = {10}, pages = {130924}, shortjournal = {Neurocomputing}, title = {Curriculum learning-driven PIELMs for fluid flow simulations}, volume = {650}, year = {2025}, }
• (2025). Waveformer: A high-precision time-frequency domain model for long-term prediction of electrical power consumption. NEUCOM, 650, 130923. (www)

  @article{ , abstract = {Accurate long-term demand forecasts are critical for energy management and system efficiency in the power industry. This paper proposes Waveformer, a high-accuracy time-frequency domain model for long-term electricity consumption forecasting. Waveformer intricately decomposes data within both time and frequency domains, isolating seasonal components and trend elements. Data is transformed into a finely-grained patch input in the time domain through patch embedding, utilizing an attention mechanism to elucidate spatial-temporal relationships. Concurrently, wavelet transform is employed in the frequency domain for noise reduction and sequence reconstruction, with linear mapping extracting pivotal features. Combining the wavelet transform, and the patch attention technique optimizes the feature extraction and time series analysis of long-period power data. Experimental results on multiple power datasets show that Waveformer significantly improves long-term prediction accuracy compared to traditional models. This model provides an efficient forecasting tool for the power industry and demonstrates its broad applicability in different power systems.}, archive = {J_NEUCOM}, author = {Yizhi Cao and Yi Xie and Xianze Zheng and Zijian Tian and Xinggao Liu}, doi = {10.1016/j.neucom.2025.130923}, journal = {Neurocomputing}, month = {10}, pages = {130923}, shortjournal = {Neurocomputing}, title = {Waveformer: A high-precision time-frequency domain model for long-term prediction of electrical power consumption}, volume = {650}, year = {2025}, }
• (2025). Efficient strip-shaped convolutional attention network via feature-based semi-online distillation for image super-resolution. NEUCOM, 650, 130922. (www)

  @article{ , abstract = {Convolutional Neural Networks (CNNs) have achieved remarkable advancements in lightweight single-image super-resolution (SISR) tasks but still face inherent limitations. Due to limited model capacity, lightweight architectures struggle to establish sufficiently large receptive fields for reconstruction demands, which constrains their ability to model long-range dependencies. Additionally, such limited capacity hinders the learning of complex image features during training, further restricting reconstruction accuracy. To address these challenges, we propose the Strip-shaped Convolutional Attention Network (SCAN), which incorporates two key innovations. First, we propose a novel strip-shaped convolution attention module, which significantly expands the network’s receptive field by cascading strip-shaped convolutions in different directions while maintaining low computational complexity, effectively addressing the limited long-range modeling capability of traditional square convolutional kernels. Second, we propose a novel semi-online distillation framework. By leveraging intermediate layer features from a pre-trained teacher model, this framework guides SCAN to dynamically learn more efficiently while improving reconstruction accuracy. Extensive experiments demonstrate that SCAN, augmented with the semi-online distillation framework, achieves superior reconstruction performance and faster inference speed compared to state-of-the-art methods, despite using fewer parameters. Notably, a lightweight variant of SCAN (SCAN-S) achieves a PSNR improvement of up to 0.10 dB over the champion models of the NTIRE 2023 and 2024 Efficient Super-Resolution Competition tracks for 4 × upscaling, while utilizing only 30 % of their parameters. The code is publicly available at: https://github.com/xuke17-2627902/SCAN .}, archive = {J_NEUCOM}, author = {Ke Xu and Lulu Pan and Guohua Peng and Lingxiao Li and Le Lei}, doi = {10.1016/j.neucom.2025.130922}, journal = {Neurocomputing}, month = {10}, pages = {130922}, shortjournal = {Neurocomputing}, title = {Efficient strip-shaped convolutional attention network via feature-based semi-online distillation for image super-resolution}, volume = {650}, year = {2025}, }
• (2025). SAMOccNet:Refined SAM-based surrounding semantic occupancy perception for autonomous driving. NEUCOM, 650, 130918. (www)

  @article{ , abstract = {Towards a more comprehensive perception and understanding of the surrounding environment for autonomous driving systems, semantic occupancy perception has gradually become a promising yet challenging research direction in autonomous driving. However, current occupancy-based perception algorithms cannot accurately capture the visual details in complex autonomous driving scenarios. To address this issue, we propose a novel method called Refined SAM-based Surrounding Semantic Occupancy Perception Network (SAMOccNet). This method introduces the Segment Anything Model (SAM) into the occupancy prediction task for the first time, leveraging SAM’s powerful ability to capture detailed information such as boundaries, textures, and regions, thereby improving prediction accuracy and generalization. To better achieve this, two key components are introduced in SAMOccNet: a dual-branch feature extraction network and a residual attention fusion module. The former extracts voxel features containing global context information and local detail information in multi-view images by designing a global voxel feature extraction branch and a SAM-based local voxel feature extraction branch, respectively. The latter is based on the residual operation and multi-attention mechanism to deeply fuse the two voxel features, and adaptively select the most effective features for occupancy prediction by dynamically adjusting the weights of the model. Furthermore, we progressively refine the initial coarse-grained occupancy predictions using high-resolution queries to generate fine-grained occupancy predictions, thereby enabling a more comprehensive and accurate reconstruction of complex 3D scenes. We conducted extensive ablation studies and experiments on the nuScenes-Occupancy dataset. The results demonstrate that SAMOccNet achieved the best performance with an IoU of 21.7 %, outperforming existing state-of-the-art methods by 1.6 %–8.72 %, thereby validating the effectiveness and advancement of the proposed approach.}, archive = {J_NEUCOM}, author = {Qifan Tan and Wenzhuo Liu and Han Bi and Lening Wang and Lei Yang and Yicheng Qiao and Zhuo Zhao and Yanhuan Jiang and Qiannan Guo and Huaping Liu and Zhiwei Li and Cheng Qiu}, doi = {10.1016/j.neucom.2025.130918}, journal = {Neurocomputing}, month = {10}, pages = {130918}, shortjournal = {Neurocomputing}, title = {SAMOccNet:Refined SAM-based surrounding semantic occupancy perception for autonomous driving}, volume = {650}, year = {2025}, }
• (2025). A reinforcement learning-assisted genetic programming algorithm for team formation problem considering person-job matching. NEUCOM, 650, 130917. (www)

  @article{ , abstract = {Efficient team formation is crucial for successful completion of new projects in a company. To address the team formation problem considering person-job matching (TFP-PJM), a 0-1 integer programming model with intuitionistic fuzzy numbers is proposed that considers both job matching and team members’ willingness to communicate for improved efficiency. A reinforcement learning-assisted genetic programming (RL-GP) algorithm is introduced as it can flexibly combine heuristic rules to solve complex TFP-PJMs. Before each generation’s population search, the agent selects from four population search modes based on information obtained, named ensemble population strategy, to balance exploration and exploitation. Furthermore, a k-Nearest Neighbor-based surrogate model is used to evaluate individual-generated formation plans, speeding up the algorithm learning process. Comparison experiments demonstrate RL-GP’s overall performance and effectiveness of improved strategies within the algorithm. The hyper-heuristic rules obtained through efficient learning can be utilized as decision-making aids when forming project teams. This study reveals the advantages of reinforcement learning methods, ensemble strategies, and the surrogate model applied to the GP framework. The diversity and intelligent selection of search patterns, along with fast adaptation evaluation, are distinct features that enable RL-GP to be deployed in real-world enterprise environments.}, archive = {J_NEUCOM}, author = {Yangyang Guo and Hao Wang and Lei He and Witold Pedrycz and Ponnuthurai Nagaratnam Suganthan and Lining Xing and Yanjie Song}, doi = {10.1016/j.neucom.2025.130917}, journal = {Neurocomputing}, month = {10}, pages = {130917}, shortjournal = {Neurocomputing}, title = {A reinforcement learning-assisted genetic programming algorithm for team formation problem considering person-job matching}, volume = {650}, year = {2025}, }
• (2025). A survey of automated essay scoring: Challenges, advances, and future. NEUCOM, 650, 130916. (www)

  @article{ , abstract = {Automated essay scoring, one of the crucial educational applications in the natural language processing community, seeks to score the quality of essays through computer technology automatically, which has garnered increasing attention due to its diverse range of research challenges and potential utility. In this paper, we provide a brief introduction to the recent advances in automated essay scoring tasks, including detailed datasets and evaluation metrics, as well as traditional and neural approaches, and discuss future trends. This survey aims to provide researchers at the intersection of artificial intelligence and educational assessment with the most up-to-date information on automated essay scoring research, as well as fundamental support for researchers just entering the field, and also provide cutting-edge ideas for industrial applications.}, archive = {J_NEUCOM}, author = {Jingbo Sun and Tianbao Song and Weiming Peng and Jihua Song}, doi = {10.1016/j.neucom.2025.130916}, journal = {Neurocomputing}, month = {10}, pages = {130916}, shortjournal = {Neurocomputing}, title = {A survey of automated essay scoring: Challenges, advances, and future}, volume = {650}, year = {2025}, }
• (2025). Representation discrepancy bridging method for remote sensing image-text retrieval. NEUCOM, 650, 130915. (www)

  @article{ , abstract = {Remote Sensing Image-Text Retrieval (RSITR) aims to bridge the semantic gap between heterogeneous modalities and plays a vital role in various geospatial applications. As a lower dimensional and more concise modality than image, the text modality is more discriminative and may dominate the optimization process. The nonnegligible imbalanced cross-modal optimization remains a bottleneck to enhancing the model’s performance. To address this issue, this study proposes a Representation Discrepancy Bridging (RDB) method for the RSITR task. On the one hand, a Cross-Modal Asymmetric Adapter (CMAA) is designed to enable modality-specific optimization and improve feature alignment. The CMAA comprises a Visual Enhancement Adapter (VEA) and a Text Semantic Adapter (TSA). VEA mines fine-grained image features through Differential Attention (DA) mechanism, while TSA identifies key textual semantics through Hierarchical Attention (HA) mechanism. On the other hand, this study extends the traditional single-task retrieval framework to a dual-task optimization framework and develops a Dual-Task Consistency Loss (DTCL). The DTCL improves cross-modal alignment robustness through an adaptive weighted combination of cross-modal, classification, and exponential moving average consistency constraints. Experiments on RSICD and RSITMD datasets show that the proposed RDB method achieves a 6 %–11 % improvement in mR metrics compared to state-of-the-art Parameter-Efficient Fine-Tuning (PEFT) methods and a 1.15 %–2 % improvement over the fully fine-tuned GeoRSCLIP model.}, archive = {J_NEUCOM}, author = {Hailong Ning and Siying Wang and Tao Lei and Xiaopeng Cao and Huanmin Dou and Bin Zhao and Asoke K. Nandi and Petia Radeva}, doi = {10.1016/j.neucom.2025.130915}, journal = {Neurocomputing}, month = {10}, pages = {130915}, shortjournal = {Neurocomputing}, title = {Representation discrepancy bridging method for remote sensing image-text retrieval}, volume = {650}, year = {2025}, }
• (2025). Cross-material stress-strain prediction: A Seq2Seq transfer approach with small punch data. NEUCOM, 650, 130914. (www)

  @article{ , abstract = {Accurate determination of stress-strain relationships in metallic materials has traditionally relied on extensive, costly experiments, especially when dealing with cross-material scenarios where data for certain materials are limited. To address these challenges, this study proposes a cross-material sequence-to-sequence (Seq2Seq) framework that predicts stress-strain relationships from Small Punch Test (SPT) data. The model transfers knowledge from high-strength steel to aluminum alloy with limited data. This domain selection reflects industrial practice, where high-strength steel has long been widely used with abundant data available, whereas aluminum alloys often suffer from data scarcity. To enhance feature extraction, raw SPT data are transformed into Gramian Angular Field (GAF) representations, and integrated with a high-order moment matching (HoMM) algorithm and a cross-attention algorithm to mitigate domain discrepancies between source and target materials. Experimental validation on both simulated datasets and real-world samples confirms the robustness and accuracy of the proposed approach. The model achieves an average mean absolute error (MAE) of 4.53 ± 3.43 MPa and a mean absolute percentage error (MAPE) of 0.91 % ± 1.04 % on high-strength steel. For real aluminum alloys, the prediction errors are as low as 2.86 %, 8.63 %, and 8.66 %. Additional ablation studies highlight the critical contributions of each model component to improving generalization across materials. The proposed framework offers a practical and accurate solution for cross-material stress-strain prediction, even when only limited data are available.}, archive = {J_NEUCOM}, author = {Zhengni Yang and Rui Yang and Weijian Han and Wenyuan Kang and Jingyi Zhang and Chao Tong and Jingyu Kong and Xiaohan Chen}, doi = {10.1016/j.neucom.2025.130914}, journal = {Neurocomputing}, month = {10}, pages = {130914}, shortjournal = {Neurocomputing}, title = {Cross-material stress-strain prediction: A Seq2Seq transfer approach with small punch data}, volume = {650}, year = {2025}, }
• (2025). DVVDNet: A dual-view vertebra detection network with anatomical constraints for automatic scoliosis assessment. NEUCOM, 650, 130913. (www)

  @article{ , abstract = {Scoliosis is a common spinal deformity, with adolescent idiopathic scoliosis being the most prevalent form. It significantly affects adolescents’ health and quality of life. Accurate assessment of spinal anatomy is essential for diagnosis and personalized treatment planning. Recent advances in deep learning have improved automated scoliosis evaluation, offering promising tools for clinical decision support. However, current methods face challenges such as vertebral feature loss in single-view X-rays due to occlusion, and reduced accuracy caused by image noise and low contrast. To address these issues, a deep learning-based dual-view vertebra detection network is proposed, which effectively integrates anterior-posterior and lateral spine radiographs by leveraging their complementary anatomical information. The network incorporates a gated multi-scale attentive fusion module to selectively enhance informative features while suppressing redundancy, thereby improving the representation capability of dual-view fusion. Additionally, a vertebral anatomy constraint loss is introduced, which incorporates both geometric morphology and structural characteristics of vertebrae to regularize model predictions, enhancing output stability and clinical applicability. The method is evaluated on the public AASCE challenge dataset and a private hospital dataset. Experimental results show high agreement with radiologists’ assessments, demonstrating the model’s accuracy and robustness. This study presents an automated, high-precision tool for scoliosis assessment, supporting more efficient diagnosis and individualized treatment planning.}, archive = {J_NEUCOM}, author = {Rui Hu and Xinwu Shu and Guoxiong Hu and Xiao Zhang}, doi = {10.1016/j.neucom.2025.130913}, journal = {Neurocomputing}, month = {10}, pages = {130913}, shortjournal = {Neurocomputing}, title = {DVVDNet: A dual-view vertebra detection network with anatomical constraints for automatic scoliosis assessment}, volume = {650}, year = {2025}, }
• (2025). Machine learning models for wearable-based human activity recognition: A comparative study. NEUCOM, 650, 130911. (www)

  @article{ , abstract = {Lately, Human Activity Recognition (HAR) has seen increased interest in healthcare, due to its ability to identify Activities of Daily Living (ADL), which are crucial for supporting independence and quality of life of elderly individuals. Network architectures in combination with wearable devices enable applications, such as remote monitoring, accident prevention, and rehabilitation support. However, real-world implementation remains hindered by challenges such as sensor configuration and placement, model complexity, and the generalizability across different activity types. In this paper, we aim to bridge the gap between HAR research and its design challenges for the practical adoption in healthcare. We systematically evaluate nine models, ranging from traditional machine learning to complex deep learning architectures across two datasets - namely, OPPORTUNITY and CogAge. We focus separately on the classification of state and behavioural activities, which must be carefully selected for different healthcare contexts, requiring different modeling strategies. The assessed models include Support Vector Machines (SVM) with hand-crafted features, straightforward neural networks like Convolutional Neural Networks (CNNs), Gated Recurrent Unit (GRU) and Long Short-Term Memory Networks (LSTMs), and more intricate models such as Transformer Networks, hybrid Convolutional LSTMs with and without attention. Our results demonstrate strong performance in recognizing state activities (94.10 %–96.48 % Average F1-Score). Behavioural activities were best recognized with the GRU model (79.91 % AF1-Score) on OPPORTUNITY and with the SVM with hand-crafted features on CogAge (69.23 % AF1-Score). Beyond insights into model performance, we provide a discussion on key design considerations for wearable-based HAR systems in healthcare.}, archive = {J_NEUCOM}, author = {Gabriela Ciortuz and Hawzhin Hozhabr Pour and Muhammad Tausif Irshad and Muhammad Adeel Nisar and Xinyu Huang and Sebastian Fudickar}, doi = {10.1016/j.neucom.2025.130911}, journal = {Neurocomputing}, month = {10}, pages = {130911}, shortjournal = {Neurocomputing}, title = {Machine learning models for wearable-based human activity recognition: A comparative study}, volume = {650}, year = {2025}, }
• (2025). Adaptive neural network finite-time optimal control for unmanned surface vehicle system. NEUCOM, 650, 130906. (www)

  @article{ , abstract = {This article investigates the adaptive neural network (NN) optimal control design problem for unmanned surface vehicle (USV) systems by finite-time control theory. A new adaptive finite-time NN optimal control policy is developed, which is composed of a NN adaptive feed-forward controller and an optimal error feedback controller. The former is constructed by using backstepping recursive control design algorithm and the latter is designed by using adaptive dynamic programming (ADP) theory. It is demonstrated that developed finite-time optimal control strategy is able to ensure the USV system is stable in a finite-time interval and achieve optimal control performance. Moreover, it can handle the computational complexity problem existing in previous finite-time optimal control methods. Comparison and simulation results illustrate the validity and superiority of the developed optimal control concept.}, archive = {J_NEUCOM}, author = {Jiaming Zhang and Wenjun Zhang and Shaocheng Tong}, doi = {10.1016/j.neucom.2025.130906}, journal = {Neurocomputing}, month = {10}, pages = {130906}, shortjournal = {Neurocomputing}, title = {Adaptive neural network finite-time optimal control for unmanned surface vehicle system}, volume = {650}, year = {2025}, }
• (2025). Pyramid hierarchical envelope generation structure based collaborative semantic unsupervised domain adaptation. NEUCOM, 650, 130902. (www)

  @article{ , abstract = {Semantic unsupervised domain adaptation (SUDA) is a crucial research area within domain adaptation. One of the key challenges in this field is determining which samples should be selected for adaptation or transfer. However, existing methods often overlook the information between the original samples and class centroids, a connection that is certified important for domain adaptation and classification. Additionally, it is also essential to explore how to conduct domain adaptation with the assistance of the information layer. To solve the problem, a Pyramid Hierarchical Envelope Generation Structure based Collaborative Semantic Unsupervised Domain Adaptation method (PHEGS_CSUDA) is proposed. Firstly, inspired by the knowledge pyramid theory, a pyramid hierarchical envelope generation structure is designed to represent the distribution characteristics of the data, ranging from a coarse-grained domain envelope to fine-grained original samples. Based on the three representative layers of this structure, which include class envelopes, granular envelopes and original samples, three distinct SUDA models are conducted. Subsequently, a bagging-like decision fusion strategy is employed to integrate the results from the three SUDA models. Finally, the traditional single-layer SUDA approach based on individual original samples is extended into a collaborative three-layer SUDA framework that incorporates different types of samples. Experimental results on benchmark datasets demonstrate the superiority and effectiveness of the proposed PHEGS_CSUDA method. Our codes are publicly available at https://github.com/cquleepf/PHEGS .}, archive = {J_NEUCOM}, author = {Pufei Li and Pin Wang and Yongming Li and Yinghua Shen and Witold Pedrycz}, doi = {10.1016/j.neucom.2025.130902}, journal = {Neurocomputing}, month = {10}, pages = {130902}, shortjournal = {Neurocomputing}, title = {Pyramid hierarchical envelope generation structure based collaborative semantic unsupervised domain adaptation}, volume = {650}, year = {2025}, }
• (2025). Deep contrastive variational subspace clustering. NEUCOM, 650, 130901. (www)

  @article{ , abstract = {Subspace clustering aims to group data points that lie in multiple low-dimensional subspaces within a high-dimensional space. Traditional shallow methods rely on linear optimization and self-expressive properties but struggle with complex, high-dimensional, and limited data samples. In contrast, Deep subspace clustering methods leverage neural networks to learn feature representations, but often specialize in specific domains, such as images, and require extensive training datasets. Furthermore, most approaches rely solely on raw features, making it difficult to capture meaningful structures in data with few examples. To address these challenges, we propose Deep Contrastive and Variational Subspace Spectral Clustering (DCSSC), a novel method that integrates relational information through a similarity matrix, enhancing clustering performance. DCSSC consists of three stages: (1) constructing a neighborhood graph for data representation, (2) training a contrastive variational autoencoder (cVAE) to learn embeddings, and (3) generating an enhanced distance matrix for spectral clustering. Our method is computationally efficient and performs well with minimal training data. Extensive experiments on simulated and 19 real-world datasets from diverse domains, including text, images, and bioinformatics, demonstrate the superiority of DCSSC over state-of-the-art deep and shallow clustering methods. Code is available at https://github.com/marcosd3souza/DCSSC .}, archive = {J_NEUCOM}, author = {Marcos de Souza Oliveira and Sergio Ricardo de Melo Queiroz and Cleber Zanchettin and Francisco de Assis Tenório de Carvalho}, doi = {10.1016/j.neucom.2025.130901}, journal = {Neurocomputing}, month = {10}, pages = {130901}, shortjournal = {Neurocomputing}, title = {Deep contrastive variational subspace clustering}, volume = {650}, year = {2025}, }
• (2025). Hierarchical graph sampling based minibatch learning with chain preservation and variance reduction. NEUCOM, 650, 130897. (www)

  @article{ , abstract = {Graph sampling-based Graph Convolutional Networks (GCNs) decouple sampling from forward and backward propagation during minibatch training, enhancing scalability with respect to layer depth and graph size. We propose HIS GCNs , a hierarchical importance sampling-based learning method. By constructing minibatches using sampled subgraphs, HIS GCNs focuses on the importance of both the core and periphery in a scale-free training graph. Specifically, it preserves the centrum of the core in most minibatches, which maintains connectivity between periphery nodes, and samples periphery edges without core node interference, which allows longer chains composed entirely of low-degree nodes remain within the same minibatch. HIS GCNs can maximize the discrete Ricci curvature (i.e., Ollivier-Ricci curvatures) of the edges in a subgraph, enabling preservation of important chains for information propagation. This approach can achieve a low node embedding variance and a high convergence speed. Diverse experiments on Graph Neural Networks (GNNs) with node classification tasks confirmed the superior performance of HIS GCNs in terms of both accuracy and training time.}, archive = {J_NEUCOM}, author = {Qia Hu and Bo Jiao}, doi = {10.1016/j.neucom.2025.130897}, journal = {Neurocomputing}, month = {10}, pages = {130897}, shortjournal = {Neurocomputing}, title = {Hierarchical graph sampling based minibatch learning with chain preservation and variance reduction}, volume = {650}, year = {2025}, }
• (2025). CTHNE: Contrastive triplet learning for heterogeneous information networks with enhanced model fusion. NEUCOM, 650, 130896. (www)

  @article{ , abstract = {Heterogeneous information networks (HINs) are widely applied across various fields because of their rich semantics and complex relational structures. Recent advancements in contrastive learning have improved the representation learning of HINs through multi-perspective contrast. However, a key challenge in self-supervised learning for HINs is how to generate semantically consistent positive samples and effective negative samples that clearly distinguish semantic differences. We propose a novel contrastive learning architecture: C ontrastive T riplet Learning for H eterogeneous Information N etworks with E nhanced Model Fusion (CTHNE) . In contrast to traditional methods that focus solely on positive–negative sample contrast, CTHNE extends contrastive learning across both views and networks. CTHNE uses meta-paths to generate multiple subgraphs as multi-views, leveraging the complementary information between different views to maximize the consistency of representations between similar graphs. We evaluate CTHNE on five practical datasets, and the results demonstrate that our approach consistently and significantly outperforms state-of-the-art baselines across all datasets and tasks. Specifically, CTHNE achieves up to a 3.2 % improvement in node clustering and a 2.8 % improvement in node classification, demonstrating its effectiveness and robustness.}, archive = {J_NEUCOM}, author = {Taiyu Zhu and Shuai Zhu and Haoyu Zhang and Xuchu Jiang}, doi = {10.1016/j.neucom.2025.130896}, journal = {Neurocomputing}, month = {10}, pages = {130896}, shortjournal = {Neurocomputing}, title = {CTHNE: Contrastive triplet learning for heterogeneous information networks with enhanced model fusion}, volume = {650}, year = {2025}, }
• (2025). APThreshold: A software tool for quantifying and analyzing action potential threshold dynamics. NEUCOM, 650, 130894. (www)

  @article{ , abstract = {The action potential (AP) threshold controls spike initiation in single neurons and exhibits considerable variability with recent firing history and input properties. Spike threshold dynamics enables neurons to perform temporally precise coding. Quantifying and analyzing threshold dynamics usually requires multiple ad-hoc programming, that is, obtaining APs, identifying threshold voltage, calculating spike features, and data visualization. To facilitate analysis efficiency, we present the AP threshold dynamics analyzer (APThreshold), which is a Python-based software tool with a user-friendly graphic user interface (GUI). We implement the waveform curvature-based and ramp stimulation-based methods for threshold identification in the software. The GUI is designed as an interactive tool, which enables users to visualize the data in a point-and-click manner and customize the coordinate axes of each plot without programming. This endows APThreshold with the ability to examine the relationship between threshold voltage and spike features and analyze its underlying biophysical basis. We provide three example workflows of how to use our software to quantify the threshold dynamics in biophysical models or experimental recordings. We also compare the estimated threshold voltages of 15 recorded action potentials from the Allen Cell Types Database using APThreshold and two existing similar tools. The results validate the functionality and efficiency of the software in identifying spike threshold with the implemented methods. APThreshold could be applied as an effective tool to analyze the threshold variability across cell types.}, archive = {J_NEUCOM}, author = {Guosheng Yi and Pengzhan Zhang and Ruifeng Bai}, doi = {10.1016/j.neucom.2025.130894}, journal = {Neurocomputing}, month = {10}, pages = {130894}, shortjournal = {Neurocomputing}, title = {APThreshold: A software tool for quantifying and analyzing action potential threshold dynamics}, volume = {650}, year = {2025}, }
• (2025). ECF-DETR: Enhanced cross-layer fusion transformer for pollen detection with IoU and classification guided evaluation. NEUCOM, 650, 130892. (www)

  @article{ , abstract = {Pollen allergy is one of the most common seasonal diseases, often triggering a variety of symptoms that severely affect both the physical and mental health of individuals. Therefore, rapid and accurate pollen detection is of great importance for preventing allergic reactions and protecting public health. However, because of the complexity of the pollen sampling process, the captured images often contain various impurities such as plant debris and dust. In addition, pollen grains are typically small, irregular in shape, and exhibit significant individual differences, making it difficult for existing models to effectively extract both global and local features, which limits detection performance. The Transformer architecture, with its powerful long-range dependency modeling capabilities, offers a promising solution to these challenges. To address these issues, this paper introduces a Transformer-based pollen detection framework named ECF-DETR. This method tackles key challenges such as limited training data, high annotation costs, and the mismatch between classification confidence and bounding box precision by introducing two core components: the Enhanced Cross-layer Location Information Fusion (E-CLIF) mechanism and the IoU and Classification Guided Evaluation (IoCE) strategy. E-CLIF adopts a hybrid matching strategy to increase the number of positive samples and fuses multi-layer spatial features to alleviate data scarcity. Meanwhile, IoCE jointly considers classification scores and IoU values to effectively mitigate the inconsistency between classification and localization. Extensive experiments conducted on a self-constructed pollen dataset in Beijing demonstrate that the proposed ECF-DETR achieves an Average Precision ( A P ) of 78.8%, outperforming the baseline DETR with the Improved deNoising anchOr box (DINO) by 1.0%, and achieving a 0.3% gain over the advanced Align-DETR framework, respectively. These findings confirm the feasibility and effectiveness of Transformer-based methods for practical pollen detection applications.}, archive = {J_NEUCOM}, author = {Baokai Zu and Xu Li and Yafang Li and Hongyuan Wang and Jianqiang Li}, doi = {10.1016/j.neucom.2025.130892}, journal = {Neurocomputing}, month = {10}, pages = {130892}, shortjournal = {Neurocomputing}, title = {ECF-DETR: Enhanced cross-layer fusion transformer for pollen detection with IoU and classification guided evaluation}, volume = {650}, year = {2025}, }
• (2025). PFRNet: Progressive multi-scale feature fusion and refinement for RGB-D salient object detection. NEUCOM, 650, 130888. (www)

  @article{ , abstract = {RGB-D salient object detection, through the integration of multi-modal feature information, is adept at generating visually compelling saliency maps. Despite the advancement of various RGB-D salient object detection models, significant challenges such as detection omissions, inaccurate object localization, and false detections persist, particularly in multi-object environments or cluttered backgrounds. To address these issues, we introduce a Progressive Multi-Scale Feature Fusion and Refinement Network (PFRNet) based on an encoder–decoder architecture. During the feature encoding phase, we utilize a dual-stream Pyramid Vision Transformer as the encoder to extract RGB and depth features. Given that low-level features contain detailed spatial information while high-level features encapsulate semantic information, we adopt the Spatial Detail Aggregation Module (SDAM) and the Semantic Feature Enhancement Module (SFEM) to facilitate the cross-modal fusion of these features. In the feature decoding stage, we design a progressive decoder anchored by the Feature Focusing and Refinement Module (FFRM). This decoder incrementally concentrates and refines discriminative information from fused features at multiple scales, simultaneously eliminating redundant content to achieve precise prediction of salient objects. The experimental results show that PFRNet outperforms 14 existing RGB-D salient object detection models across six public datasets, while demonstrating the method’s strong generalization capabilities in RGB-T salient object detection tasks.}, archive = {J_NEUCOM}, author = {Zhengqian Feng and Wei Wang and Mingle Zhou and Wang Li and Yuan Gao and Jiachen Li and Gang Li}, doi = {10.1016/j.neucom.2025.130888}, journal = {Neurocomputing}, month = {10}, pages = {130888}, shortjournal = {Neurocomputing}, title = {PFRNet: Progressive multi-scale feature fusion and refinement for RGB-D salient object detection}, volume = {650}, year = {2025}, }
• (2025). Event-triggered finite-time adaptive consensus control for stochastic nonlinear multi-agent systems. NEUCOM, 650, 130881. (www)

  @article{ , abstract = {In this paper, the finite-time distributed adaptive consensus control problem is studied for a class of continuous nonlinear multi-agent systems (MASs) with stochastic perturbations, external disturbances and unknown dynamics. By utilizing backstepping control approach, Zeno free event-triggered mechanism, artificial neural networks (ANNs) and finite-time command filter method (integrated with compensation signals), innovative finite-time control inputs are developed to guarantee that the outputs of all followers continuously track the leader’s output within a finite-time. In design of control inputs, the event-triggered technique is adopted to minimize the communication load and decrease the frequency of the control input update in practice. Moreover, ANNs are used to estimate the unknown nonlinear functions, stochastic perturbations and unknown external disturbances. Furthermore, to deal with the complexity explosion problem (which is caused by traditional backstepping approach), a developed finite-time command filter technique is suggested and applied. With the assistance of the Lyapunov stability theory, it is theoretically proven that the closed-loop MAS (considering the discussed undesirable factors) has the semi-globally finite-time stability in probability (SGFSP). In conclusion, numerical simulations are performed and represented to demonstrate the effectiveness of the proposed control scheme for the considered consensus problem.}, archive = {J_NEUCOM}, author = {Fatemeh Sedghi and Mohammad Mehdi Arefi and Ali Abooee}, doi = {10.1016/j.neucom.2025.130881}, journal = {Neurocomputing}, month = {10}, pages = {130881}, shortjournal = {Neurocomputing}, title = {Event-triggered finite-time adaptive consensus control for stochastic nonlinear multi-agent systems}, volume = {650}, year = {2025}, }
• (2025). Towards a new perspective on multi-objective enhanced ensemble classification. NEUCOM, 650, 130878. (www)

  @article{ , abstract = {Accuracy and diversity are two critical metrics in ensemble classification, but they often conflict with each other. Existing methods for generating diverse classifiers achieve some degree of diversity but struggle to balance it with predictive performance. To further improve ensemble classification performance, current approaches employ multi-objective optimization for classifier selection. However, these methods focus solely on diversity and accuracy, relying on uniformly set weight vectors, which fail to ensure a comprehensive exploration of the solution space and an optimal trade-off. To address these limitations, we propose a novel Multi-Objective Enhanced Ensemble Classification (MOEEC) method. MOEEC begins by training a set of base classifiers through the randomization of Gaussian kernel parameters. We propose a multi-objective optimization method to optimize the ensemble members based on their diversity, accuracy, and similarity. The optimization method also allows for the continuous adjustment of sub-objective weights based on feedback during the optimization process, enabling more effective exploration and exploitation of the solution space. Extensive experimental results on various datasets demonstrate the effectiveness of the proposed method. Code for this paper is available on: https://github.com/XiaojianDing/MOEEC .}, archive = {J_NEUCOM}, author = {Xiaojian Ding and Xin Wang and Zihan Xu and Pengcheng Shi and Kaixiang Wang}, doi = {10.1016/j.neucom.2025.130878}, journal = {Neurocomputing}, month = {10}, pages = {130878}, shortjournal = {Neurocomputing}, title = {Towards a new perspective on multi-objective enhanced ensemble classification}, volume = {650}, year = {2025}, }
• (2025). H3-astronaut pose net: A multiview 3D astronaut pose estimation system for spacecraft. NEUCOM, 650, 130877. (www)

  @article{ , abstract = {Recent years have witnessed the materialization of intelligent space station concepts, with operational paradigms transitioning from mechanized to intelligent systems. This technological evolution relies on robust human–machine interaction systems, which necessitate precise positioning and joint posture measuring of astronauts within intracabin environments. We propose a multi-view voxel-based fusion network (Astronaut Pose Net, APN: comprising Human Centroid Net, Human Localization Net, and Human Joints Net). APN employs multi-dimensional CNNs to extract image features, regress human centroids, localize planar occupancy, and ultimately fuse multi-view data to estimate 3D human poses. To address the core issues of posture variation caused by microgravity environments in manned spacecraft cabins and quantization errors inherent in voxel-based algorithms, this solution constructs three key innovations: (1) Biomechanical Encoding Supervision: We develop a body segment parameters embedding to compute gaussian centroid distributions as supervisory signals for training the Human Centroid Net to simultaneously regress anatomical centroid coordinates and establish centroid-referenced planar occupancy maps. (2) Dual-Attention CNN: Aimming to enhance the feature perception capability of 2D convolutional neural networks, we proposal a Dual-Attention CNN for human planar occupancy tasks in Human Centroid Net and Human Localization Net. (3) Fusion Attention Scale-Aware (FASA) Module: A scale-aware module with fused attention mechanisms is proposed in Human Joints Net to improve multi-scale feature extraction during pose regression, effectively mitigating accuracy degradation caused by human scale variations when implementing body segment parameter embedding. We test our method on the CMU Panoptic dataset, and the results demonstrate that our method achieves significant accuracy improvement on the dataset ( A P 25 +4.59%), enabling precise capture of human posture information even under occlusion conditions. This work providing a referential solution for non-contact astronaut precise occupancy and pose estimation in intelligent space station architectures.}, archive = {J_NEUCOM}, author = {Ruiqi Xun and Zhang Zhang and Liang Chang}, doi = {10.1016/j.neucom.2025.130877}, journal = {Neurocomputing}, month = {10}, pages = {130877}, shortjournal = {Neurocomputing}, title = {H3-astronaut pose net: A multiview 3D astronaut pose estimation system for spacecraft}, volume = {650}, year = {2025}, }
• (2025). Optimizing FPGA-based YOLO series accelerators: A survey of techniques. NEUCOM, 650, 130874. (www)

  @article{ , abstract = {This review provides a comprehensive overview of hardware optimization techniques used to accelerate the You Only Look Once (YOLO) series of object detection models on Field-Programmable Gate Arrays (FPGAs). YOLO models are widely adopted for real-time applications due to their balance of accuracy and speed, but their computational demands necessitate innovative solutions for deployment in resource-constrained environments. This paper categorizes optimization techniques into two primary categories: model-level (e.g., quantization, pruning) and architectural-level (e.g., parallelism, pipelining). Furthermore, it describes the key performance metrics utilized in most previous works. Challenges such as trade-offs between latency, complexity, and accuracy are highlighted, along with promising avenues for future research. This survey serves as a valuable resource for designing efficient FPGA-based YOLO series accelerators, advancing the deployment of object detection in edge and embedded systems.}, archive = {J_NEUCOM}, author = {Chua Wee Heng and Chessda Uttraphan and Chew Chang Choon and Kok Boon Ching}, doi = {10.1016/j.neucom.2025.130874}, journal = {Neurocomputing}, month = {10}, pages = {130874}, shortjournal = {Neurocomputing}, title = {Optimizing FPGA-based YOLO series accelerators: A survey of techniques}, volume = {650}, year = {2025}, }
• (2025). Enhanced non-negative matrix factorization via adaptive weighted bipartite graph learning for clustering problems. NEUCOM, 650, 130871. (www)

  @article{ , abstract = {Non-negative matrix factorization (NMF)-based clustering models, widely employed in modern applications, typically consist of two principal stages: obtaining low-dimensional representation through NMF and applying clustering algorithms such as k-means to the representation. However, traditional NMF methods consider these stages as independent processes, which fail to fully exploit the cluster centroids to guide the basis vectors. This may limit the optimization of basis vectors and result in low-dimensional representations that are unsuitable for clustering tasks. To address this issue, we propose an enhanced NMF method via adaptive weighted bipartite graph learning (AWBG-NMF). First, the NMF objective function is reformulated using a bipartite graph framework, where a flexible bipartite graph and an original bipartite graph are constructed to model the relationship between basis vectors and cluster centroids. Second, a novel adaptive weighted bipartite graph learning strategy is introduced to exploit high-quality cluster centroids: (a) a weight matrix is incorporated to adaptively assign feature weights, which alleviates the impact of noise; (b) a more precise flexible bipartite graph is learned as pseudo-label information to guide the optimization of basis vectors. Additionally, a centroid correlation strategy is implemented to enhance the low-dimensional representation in NMF by exploiting latent correlations among basis vectors. Finally, an efficient optimization algorithm is developed, and its convergence is theoretically proven. Extensive experiments on eight real-world datasets demonstrate that AWBG-NMF significantly outperforms state-of-the-art methods in clustering performance.}, archive = {J_NEUCOM}, author = {Yulei Huang and Libo Liu}, doi = {10.1016/j.neucom.2025.130871}, journal = {Neurocomputing}, month = {10}, pages = {130871}, shortjournal = {Neurocomputing}, title = {Enhanced non-negative matrix factorization via adaptive weighted bipartite graph learning for clustering problems}, volume = {650}, year = {2025}, }
• (2025). Hierarchical motion magnification. NEUCOM, 650, 130869. (www)

  @article{ , abstract = {Video motion magnification amplifies small, imperceptible variations, making them perceptible. It has a wide range of applications, such as measuring respiratory signals, classifying micro-expressions, measuring micro-displacements from a distance etc. Moreover, its applications are continuously expanding. Existing handcrafted techniques provide less effective solutions, often delivering lower magnification and requiring the selection of complex hyper-parameter values that vary across videos. To tackle these issues, deep-learning-based approaches are being introduced. However, they are sensitive to noise-related distortions during motion magnification. In response to these challenges, we propose a more robust hierarchical network for video motion magnification. To mitigate distortions that may arise from the magnification of noise and other illumination changes, we introduce a multi-scale manipulator with an edge-based input before magnification. Additionally, we propose a novel contrastive learning-based loss to further enhance the robustness of the magnification process against noise. To improve texture quality, we introduce a multi-resolution magnification generation architecture with a magnification decoder block. It produces hierarchical magnification from lower resolution to higher resolution. We compare the results of the proposed network both qualitatively and quantitatively with the state-of-the-art methods on synthetic and natural videos. Also, we conduct an ablation study to analyze various modules of our method. The results demonstrate that our proposed base and lightweight models outperform the current state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Jasdeep Singh and Santosh Kumar Vipparthi and Subrahmanyam Murala and G. Sankara Raju Kosuru and Hasan Al-Marzouqi}, doi = {10.1016/j.neucom.2025.130869}, journal = {Neurocomputing}, month = {10}, pages = {130869}, shortjournal = {Neurocomputing}, title = {Hierarchical motion magnification}, volume = {650}, year = {2025}, }
• (2025). Historical residual-based stealthy attack against cyber–physical systems. NEUCOM, 650, 130868. (www)

  @article{ , abstract = {This paper investigates the design of stealthy attacks against remote state estimation in cyber–physical systems, where the attack stealthiness is characterized by the Kullback–Leibler divergence. The objective is to degrade the estimation performance while ensuring the attack remains stealthy. A novel historical residual-based attack model, which incorporates all available historical residuals while requiring only a single parameter to be designed at each step, is proposed, and the error covariance of the remote estimation under the attack is derived as the attack performance metric. Then, to maximize the attack performance, the worst-case stealthy attack strategy is analytically obtained. Compared with the existing results, the proposed attack strategy achieves the desired attack performance with fewer design parameters. Finally, simulations are provided to verify the effectiveness of the developed results.}, archive = {J_NEUCOM}, author = {Li-Wei Mao and Guang-Hong Yang}, doi = {10.1016/j.neucom.2025.130868}, journal = {Neurocomputing}, month = {10}, pages = {130868}, shortjournal = {Neurocomputing}, title = {Historical residual-based stealthy attack against cyber–physical systems}, volume = {650}, year = {2025}, }
• (2025). BFC-net: Boundary-frame cross graph attention network for partially spoofed audio localization. NEUCOM, 650, 130867. (www)

  @article{ , abstract = {Partially spoofed audio poses a significant cybersecurity threat, requiring advanced methods to precisely locate subtle spoofed segments. While some existing studies have achieved remarkable results, these methods fail to effectively leverage inter-frame differences and spoof boundaries. To address this issue, we propose a novel method called Boundary-Frame Cross Graph Attention Network (BFC-Net). Specifically, BFC-Net consists of three core modules: Large Selective Kernel Residual Block (LSK-ResBlock), Multi-level Feature Aggregation (MFA), and Boundary-Frame Cross Attention (BFCA). Initially, LSK-ResBlock enhances spectral and temporal representations in high-dimensional space. Then, MFA aggregates local and global features to achieve accurate boundary prediction. Finally, BFCA controls frame interactions based on boundary prediction. Meanwhile, it effectively integrates inter-frame differences and boundary features, enabling precise localization of spoofed audio segments. We systematically evaluated BFC-Net on the PartialSpoof dataset and ADD2023 Track 2 dataset. Experimental results demonstrate that BFC-Net achieved an EER of 2.73% on the PartialSpoof and a Score of 92.13% on ADD2023 Track 2, attaining state-of-the-art performance on both benchmarks. BFC-Net not only fills the gap in partially spoofed audio localization but also provides strong support for ensuring the security and authenticity of network information.}, archive = {J_NEUCOM}, author = {Yi Zhou and Zhaodong Xue and Lotfi Senhadji and Huazhong Shu and Jiasong Wu}, doi = {10.1016/j.neucom.2025.130867}, journal = {Neurocomputing}, month = {10}, pages = {130867}, shortjournal = {Neurocomputing}, title = {BFC-net: Boundary-frame cross graph attention network for partially spoofed audio localization}, volume = {650}, year = {2025}, }
• (2025). A density variation and inverse outliers clustering algorithm for datasets with large density differences. NEUCOM, 650, 130866. (www)

  @article{ , abstract = {The density peaks clustering (DPC) algorithm was introduced in 2014. It is a density-based method that quickly identifies density peaks. However, DPC still has certain limitations. It often struggles to determine accurate cluster centers in datasets containing both sparse and dense clusters. To overcome the problem, an innovative design enhances the algorithm to focus on sparse clusters. In this paper, we propose a density variation and inverse outliers clustering algorithm for datasets with large density differences (DVIOC), which aims to efficiently distinguish sparse and dense clusters while suppressing the occurrence of cascading errors. Firstly, we modify the local density calculation using a varying correlation coefficient formula to enhance the accuracy of the density computation. Then, the selected representative points are merged to determine the appropriate cluster centers. In a novel assignment strategy, the local outlier factor (LOF) is applied to identify normal points in reverse. These normal points constitute the main structure of the clusters and form microclusters. Ultimately, the residual points are assigned by dynamic fuzzy weighted, which uses an affiliation degree to assign each point into the appropriate clusters. A series of experiments on synthetic datasets and UCI datasets demonstrate the superiority of the DVIOC algorithm over other state-of-the-art algorithms. In particular, the algorithm demonstrates strong robustness to significant inter-cluster differences and outlier boundaries within the dataset.}, archive = {J_NEUCOM}, author = {Xiangli Li and Yalin Kou and Yongqi Mi}, doi = {10.1016/j.neucom.2025.130866}, journal = {Neurocomputing}, month = {10}, pages = {130866}, shortjournal = {Neurocomputing}, title = {A density variation and inverse outliers clustering algorithm for datasets with large density differences}, volume = {650}, year = {2025}, }
• (2025). A min–max optimization framework for sparse multi-task deep neural network. NEUCOM, 650, 130865. (www)

  @article{ , abstract = {Multi-task learning is a subfield of machine learning in which the data is trained with a shared model to solve different tasks simultaneously. Instead of training multiple models, we only need to train a single model with shared parameters to solve different tasks. By sharing parameters, multi-task learning significantly decreases the number of parameters and reduces computational and storage requirements. However, when applying multi-task learning to deep neural networks, model size remains a challenge, particularly for edge platforms. Compressing multi-task models while maintaining performance across all tasks is another significant challenge. To address these issues, we propose a min–max optimization framework for highly compressed multi-task deep neural network models, combined with weight pruning or dynamic sparse training strategies to improve training efficiency by reducing model parameters. Specifically, weight pruning leverages reweighted l 1 pruning method, enabling high pruning rates while preserving the performance across all tasks. Dynamic sparse training, on the other hand, initializes and updates the sparse masks of the network dynamically during the training process while maintaining the same number of weights, which typically encourages sparsity in the weight matrices with the advantage of reducing memory footprint and computational requirements. Our proposed min–max optimization framework can automatically adjust the learnable weighting factors between different tasks, ensuring optimization for the worst-performing task. Experimental results on NYUv2 and CIFAR-100 datasets demonstrate that the model incurs minor performance degradation after pruning with the min–max framework. Further analyses indicate the min–max framework has reliable performance and the difference from prior methods is statistically significant. The proposed dynamic sparse multi-task framework reaches around 2% overall precision improvement using min–max optimization compared with prior methods when the models are equally sparsed.}, archive = {J_NEUCOM}, author = {Jiacheng Guo and Lei Li and Huiming Sun and Minghai Qin and Hongkai Yu and Tianyun Zhang}, doi = {10.1016/j.neucom.2025.130865}, journal = {Neurocomputing}, month = {10}, pages = {130865}, shortjournal = {Neurocomputing}, title = {A min–max optimization framework for sparse multi-task deep neural network}, volume = {650}, year = {2025}, }
• (2025). Inverse simulation learning of quasi-nonlinear fuzzy cognitive maps. NEUCOM, 650, 130864. (www)

  @article{ , abstract = {This paper presents a learning algorithm designed to address the inverse simulation problems in Quasi-Nonlinear Fuzzy Cognitive Maps (qFCMs). The algorithm determines the initial conditions for qFCM models that produce the desired outputs specified by the modeler. The contribution rests on three main components, which aim to generate feasible and accurate solutions. Firstly, we employ a quasi-nonlinear reasoning rule designed to ensure that the qFCM model will always converge to multiple fixed points that depend on the initial conditions. By steering clear of unique fixed-point attractors, the proposed learning algorithm mitigates the risk of encountering unsolvable inverse simulation scenarios. Secondly, we provide the modeler with mathematically determined, yet comprehensive insights into the feasible activation space for each neural concept. These spaces represent intervals that outline the activation values that can be reached by the neuronal concepts, which allow the modeler to select realistic targets. Lastly, we formulate the inverse learning problem for qFCM models as an optimization task resolved with the aid of numerical optimizers that approximate the Jacobian and Hessian. Empirical studies illustrate the effectiveness of our proposal to produce accurate inverse solutions for real-world scenarios and synthetically generated models.}, archive = {J_NEUCOM}, author = {Gonzalo Nápoles and Jose L. Salmeron and Yamisleydi Salgueiro}, doi = {10.1016/j.neucom.2025.130864}, journal = {Neurocomputing}, month = {10}, pages = {130864}, shortjournal = {Neurocomputing}, title = {Inverse simulation learning of quasi-nonlinear fuzzy cognitive maps}, volume = {650}, year = {2025}, }
• (2025). Mitigating impacts of hyperedge heterogeneity on semi-supervised hypergraph contrastive learning. NEUCOM, 650, 130863. (www)

  @article{ , abstract = {Many real-world interactions exhibit complex, diverse, and high-order relationships. Hypergraphs provide a flexible and natural representation for such correlations, gaining increasing attention in the community. Recent studies have extended graph convolutional networks and contrastive learning to hypergraphs, but it encounters two issues posed by hyperedge heterogeneity: (i) message-passing based hypergraph neural networks let node mix different neighbor information within heterogeneous hyperedges, resulting in poor representation learning. (ii) hyperedge heterogeneity influences model’s discriminative capability leading to ambiguous classification boundary. To solve these issues, we propose a new semi-supervised contrastive hypergraph learning model to mitigate impacts of hyperedge heterogeneity. Specifically, we first propose a new measurement of hyperedge homogeneity based on label information entropy and we map representation difference between node and hyperedge to approximate it making it applicable during training. Secondly, we develop a hypergraph encoder with this approximated hyperedge homogeneity as adaptive message passing weight to relieve hyperedge heterogeneity. Thirdly, we utilize pseudo-labels to bring similar nodes closer in hypergraph data augmentation. Lastly, by combining quality metrics, discriminative metrics, and visual analysis, we demonstrate the effectiveness of the proposed model compared with state-of-the-art semi-supervised and unsupervised methods across six public datasets in node classification task and the experiment of different heterogeneity levels.}, archive = {J_NEUCOM}, author = {Kezhao Yin and Xiaolun Jing and Genke Yang and Jian Chu}, doi = {10.1016/j.neucom.2025.130863}, journal = {Neurocomputing}, month = {10}, pages = {130863}, shortjournal = {Neurocomputing}, title = {Mitigating impacts of hyperedge heterogeneity on semi-supervised hypergraph contrastive learning}, volume = {650}, year = {2025}, }
• (2025). 4K-HAZE: A dehazing benchmark with 4K resolution hazy and haze-free images. NEUCOM, 650, 130861. (www)

  @article{ , abstract = {Haze removal from images is a fundamental computer vision challenge with significant implications for various applications. Currently, there are numerous growing demands for 4K image processing methods on the mobile and IOT (Internet of Things) devices. Absence of large-scale 4K benchmark datasets hampers progress in this area, especially for the task of dehazing. The challenges of building an ultra-high definition (UHD) dehazed dataset are the lack of estimation methods for UHD depth maps and the migration strategy from synthetic to real domains for UHD with haze images in constrained computational resources. To address these problems, we develop a novel synthetic method to simulate 4K hazy images (including nighttime and daytime scenes) from clear images, which first estimates the scene depth, simulates the light rays and object reflectance, then migrates the synthetic images to real domains by using a GAN, and finally yeilds the hazy effects on 4K resolution images. We wrap these synthesized images into a benchmark called the 4K-HAZE dataset. Specifically, we design the CS-Mixer (an MLP-based model that integrates C hannel domain and S patial domain) to estimate the depth map of 4K clear images, the GU-Net to migrate a 4K synthetic image to the real hazy domain. The most appealing aspect of our approach is the capability to run a 4K image on a single GPU with 24G RAM in real-time (33fps). Additionally, this work presents an objective assessment of several state-of-the-art single image dehazing methods that were evaluated using 4K-HAZE dataset. At the end of the paper, we discuss the limitations of the 4K-HAZE dataset and its social implications.}, archive = {J_NEUCOM}, author = {Xin Su and Pengwen Dai and Chen Lv and Zhuoran Zheng}, doi = {10.1016/j.neucom.2025.130861}, journal = {Neurocomputing}, month = {10}, pages = {130861}, shortjournal = {Neurocomputing}, title = {4K-HAZE: A dehazing benchmark with 4K resolution hazy and haze-free images}, volume = {650}, year = {2025}, }
• (2025). HeLPNet: Facial structure-aware landmark coordinate regression with heatmap-guided local patch embedding. NEUCOM, 650, 130860. (www)

  @article{ , abstract = {Recent transformer-based coordinate regression methods, employing local patch embedding and landmark queries, have shown remarkable progress in face alignment. However, they still encounter challenges in capturing spatial information, such as the facial contour and face structure within local patches, compared to traditional CNN-based heatmap regression methods. To address this problem, we propose a simple yet effective model for face alignment, Heatmap-guided Local Patch Network (HeLPNet), which leverages abundant spatial information extracted from various types of heatmap regression outputs as guidance for local patch embedding. This mechanism enables a landmark query-based transformer to effectively reflect spatial information and learn the relationships between landmark queries. Furthermore, to enhance the effectiveness of point heatmap guidance, we propose tangent direction anisotropic (TDA) Gaussian point heatmap, which consists of anisotropic circular points that emphasize the on-edge pixels in each point. Extensive experiments demonstrate that HeLPNet achieves state-of-the-art (SOTA) performance levels on three benchmarks, WFLW, 300 W, and COFW, notably reaching a remarkably low failure rate of 1.60 on WFLW, outperforming the previous SOTA model by a substantial margin of 0.48.}, archive = {J_NEUCOM}, author = {Jeongmin Yu and Jongseong Bae and Ha Young Kim}, doi = {10.1016/j.neucom.2025.130860}, journal = {Neurocomputing}, month = {10}, pages = {130860}, shortjournal = {Neurocomputing}, title = {HeLPNet: Facial structure-aware landmark coordinate regression with heatmap-guided local patch embedding}, volume = {650}, year = {2025}, }
• (2025). LoRA2: Multi-scale low-rank approximations for fine-tuning large language models. NEUCOM, 650, 130859. (www)

  @article{ , abstract = {Fine-tuning large language models (LLMs) with high parameter efficiency for downstream tasks has become a new paradigm. Low-Rank Adaptation (LoRA) significantly reduces the number of trainable parameters for fine-tuning. Although it has demonstrated commendable performance, updating parameters within a single scale may not be the optimal choice for complex downstream tasks. In this paper, we extend the LoRA to multiple scales, dubbed as LoRA 2 . We first combine orthogonal projection theory to train two low-dimensional LoRAs in two mutually orthogonal planes. By multiplying two LoRAs, a high-dimensional LoRA is obtained, forming a multi-scale LoRA. Then, we improve the importance score algorithm, significantly reducing the computation required for parameter sensitivity scoring. By pruning singular values with lower importance scores, thereby enhancing adaptability to various downstream tasks. Extensive experiments are conducted on two widely used pre-trained models to validate the effectiveness of LoRA 2 . Results show that it significantly reduces the number of trainable parameters to just 0.72% compared to full fine-tuning on the DeBERTa-V3-base model, while still delivering highly impressive performance. Our code is available here: https://github.com/Godz-z/LoRA-2 .}, archive = {J_NEUCOM}, author = {Jia-Chen Zhang and Yu-Jie Xiong and Chun-Ming Xia and Dong-Hai Zhu and Hong-Jian Zhan}, doi = {10.1016/j.neucom.2025.130859}, journal = {Neurocomputing}, month = {10}, pages = {130859}, shortjournal = {Neurocomputing}, title = {LoRA2: Multi-scale low-rank approximations for fine-tuning large language models}, volume = {650}, year = {2025}, }
• (2025). A multi-level contrastive learning-based correction model for semi-supervised fetal echocardiography segmentation. NEUCOM, 650, 130858. (www)

  @article{ , abstract = {Fetal echocardiography is an important test for diagnosing fetal congenital heart disease (CHD). The apical four-chamber view (A4C) is the most basic and essential view for fetal heart examination. Segmentation of key anatomical structures based on the A4C view provides a basis for functional assessment and disease diagnosis. However, current fetal heart segmentation methods do not fully leverage unlabeled intermediate frames in ultrasound videos. Furthermore, the complex and variable structure of the fetal heart often causes category confusion, significantly affecting the precision and reliability of segmentation. In this paper, we propose a multi-level contrastive learning-based correction model (MLCL-CM) for semi-supervised segmentation of five key cardiac structures in fetal echocardiography. MLCL-CM improves segmentation performance on labeled keyframes by exploiting the large number of unlabeled intermediate frames. To alleviate misclassification problems, we construct positive and negative pairs in category- and pixel-level contrastive learning and introduce two uncertainty quantization techniques to guide the learning of more robust and fine-grained discriminative feature representations in global and local spaces. In addition, we integrate an adaptive correction model to achieve more stable and consistent predictions by dynamically weighting the inconsistent regions. Experimental results on actual A4C views with 100 and 300 fetuses show that the proposed MLCL-CM outperforms the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Qi Yu and Danyang Song and Jinzhu Yang and Zhitao Zhang and Song Sun and Haiyang Sun and Yangyang Li and Yao Liu and Yiqiu Qi and Mei Wei and Yiming Liu}, doi = {10.1016/j.neucom.2025.130858}, journal = {Neurocomputing}, month = {10}, pages = {130858}, shortjournal = {Neurocomputing}, title = {A multi-level contrastive learning-based correction model for semi-supervised fetal echocardiography segmentation}, volume = {650}, year = {2025}, }
• (2025). A novel fuzzy-rule-based deep fusion of hypergraph multi-modal for alzheimer’s disease detection. NEUCOM, 650, 130855. (www)

  @article{ , abstract = {Alzheimer’s disease (AD) is a neurodegenerative disorder that impairs cognitive and memory functions. Currently, multi-modal neuroimaging-based hypergraphs have become the crucial methods for AD detection. However, most existing multi-modal methods only focus on the overall contribution of modality destroying the balance of fine-grained features on collaborative expression in feature-level fusion, while this methods lack the interpretability of detection results and is difficult to reveal the multi-level pathological mechanism of AD accurately. To overcome these challenges, we propose a novel fuzzy-rule-based deep fusion of hypergraph multi-modal for Alzheimer’s disease detection. Specifically, we first design a hypergraph-based multi-modal feature selection and fusion algorithm that employs the group sparse regularization technique to jointly select features of the same dimension across modalities, while quantifying the relative contributions of fine-grained features as the weight allocation of feature-level fusion. Secondly, we present a first-order TSK fuzzy support vector machine to enhance interpretability, which uses the inference relationship between input and output to reveal the multi-level pathological mechanisms of AD following multi-modal deep fusion. Experimental results on the ADNI dataset demonstrate that the proposed method significantly improves AD detection performance.}, archive = {J_NEUCOM}, author = {Manman Yuan and Can Yin and Jun Hu and Yan Zhao and Jiacheng Wang and Huijia Li}, doi = {10.1016/j.neucom.2025.130855}, journal = {Neurocomputing}, month = {10}, pages = {130855}, shortjournal = {Neurocomputing}, title = {A novel fuzzy-rule-based deep fusion of hypergraph multi-modal for alzheimer’s disease detection}, volume = {650}, year = {2025}, }
• (2025). VMD-MSANet: A multi-scale attention network for stock series prediction with variational mode decomposition. NEUCOM, 650, 130854. (www)

  @article{ , abstract = {The stock market is a crucial component of the financial system. Accurate prediction of its price is essential for effective risk management and informed investment decision-making. However, the complex dynamics of the stock market, including multi-scale non-stationarity and complex stock-market interactions, pose significant challenges for prediction. To address these challenges, we introduce VMD-MSANet, a stock price prediction model that combines Variational Mode Decomposition (VMD) with a multi-scale attention mechanism. We employ VMD to decompose the stock price series into sub-components with distinct frequency components, and use the multi-scale attention mechanism to capture both short-term and long-term temporal patterns effectively. By incorporating external market factors, the model enhances its comprehensive understanding and adaptability to the market environment. Extensive experiments on the Chinese market demonstrate that VMD-MSANet achieves higher predictive accuracy and exhibits enhanced robustness and generalization compared to existing state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Yunzhu Chen and Neng Ye and Wenyu Zhang and Sijia Lv and Liwei Shao and Xiangming Li}, doi = {10.1016/j.neucom.2025.130854}, journal = {Neurocomputing}, month = {10}, pages = {130854}, shortjournal = {Neurocomputing}, title = {VMD-MSANet: A multi-scale attention network for stock series prediction with variational mode decomposition}, volume = {650}, year = {2025}, }
• (2025). Memory-VQA: Video quality assessment of UGC based on human memory system. NEUCOM, 650, 130852. (www)

  @article{ , abstract = {The rapid expansion of user-generated content (UGC) videos poses a significant challenge to video quality assessment (VQA) algorithms. Previous research in UGC-VQA has primarily aimed to capture temporal and spatial distortions in videos or replicate the subjective perception of videos by the human visual system (HVS). However, VQA is a protracted process that requires careful observation, comparison, and memorization of video content rather than instant assessment. This paper proposes the Memory-VQA method, drawing on insights from the human memory system (HMS). The method integrates five stages related to memory and is an early attempt to bring the process of memory formation to the VQA domain. At the outset of memory perception, theories of memory formation and reconstruction were used to segment the video into sequences with varying frame rates. Following segmentation, the video’s motion and spatial attributes were isolated during the memory encode phase, taking into account their influence on memory consolidation. To simulate memory loss, the encode features undergo a linear transformation, and their configurations are modified during the memory storage process. Subsequently, keyframes are integrated with previously stored features based on cue-driven theory, facilitating the extraction of spatio-temporal characteristics during the memory retrieval phase. Additionally, the introduction of a novel attention module enhances the nuanced perception of details in the memory retrieval process. In the final phase of the memory reconstruction module, subliminal effects integrate encode and retrieval characteristics to predict quality. The experimental results demonstrate that the proposed method achieves the best performance on five popular UGC-VQA datasets. The code is available at https://github.com/luoshui123/Memory-vqa .}, archive = {J_NEUCOM}, author = {Minjie Sun and Guangqian Kong and Xun Duan and Huiyun Long}, doi = {10.1016/j.neucom.2025.130852}, journal = {Neurocomputing}, month = {10}, pages = {130852}, shortjournal = {Neurocomputing}, title = {Memory-VQA: Video quality assessment of UGC based on human memory system}, volume = {650}, year = {2025}, }
• (2025). Question fuzzy-attention embedding graph-to-tree network for intelligent math word problem solver. NEUCOM, 650, 130850. (www)

  @article{ , abstract = {Designing intelligent solution models for Math Word Problems (MWPs) holds significant importance for intelligent education. The task of solving MWPs involves translating mathematical problem texts into equations. The introduction of artificial intelligence-generated content models has opened new avenues for advancing intelligent education through digital technology, albeit with new challenges. Current intelligent educational assistants often lack adequate mathematical computation and complex reasoning abilities, especially in dealing with specialized domain knowledge. To address this gap, the paper proposes a semantic document embedding method based on axiomatic fuzzy set membership. The method focuses on enhancing the model’s mathematical computation capabilities by processing semantic key information embedded in textual problems. The paper introduces an innovative framework for applying intelligent education systems, specifically targeting tasks like quantity relationship graphs in mathematical questions. Extensive experiments were conducted on two popular MWP datasets. Results show that the proposed model significantly outperforms existing baseline models on both benchmark datasets. Further detailed ablation studies demonstrate the effectiveness and interpretability of the proposed model. Moreover, the paper validates the model’s reasoning capabilities in practical subjects such as actuarial science and financial analysis.}, archive = {J_NEUCOM}, author = {Qi Lang and Minghao Yin and Zhiwei Wang and Xiaodong Liu and Shuang Liang}, doi = {10.1016/j.neucom.2025.130850}, journal = {Neurocomputing}, month = {10}, pages = {130850}, shortjournal = {Neurocomputing}, title = {Question fuzzy-attention embedding graph-to-tree network for intelligent math word problem solver}, volume = {650}, year = {2025}, }
• (2025). Kernel clustering with automatic variable weighting for interval data. NEUCOM, 650, 130849. (www)

  @article{ , abstract = {Symbolic Data Analysis (SDA) is a field associated with statistics and artificial intelligence that deals with multi-valued data, such as histograms, intervals, and lists. This type of data emerges as an alternative to conventional aggregation methods (mean, median, and mode) to account for variability. It is particularly useful when analyzing groups of individuals rather than single individuals. For example, when we have information about patients but aim to describe and analyze hospitals. Kernel functions are extensively employed in clustering algorithms because they perform better when there is no linear separability between clusters and/or the clusters are not hyperspherical. This paper proposes new clustering methods based on the Gaussian kernel that weigh the interval-valued variables automatically. These methods are particularly appropriate when there are non-informative variables or variables relevant to specific clusters. We introduce four global variants, in which each variable has the same weight across all clusters, and two local variants, where variable weights differ for each cluster. Although local methods offer greater flexibility in the weighting scheme, global methods are less susceptible to local minima. Experimental evaluation over simulated and real interval-valued datasets, compared to traditional clustering methods for interval data, demonstrated the effectiveness of the introduced algorithms. The source code and datasets are available at https://github.com/Natandradesa/Kernel-Clustering-for-Interval-Data .}, archive = {J_NEUCOM}, author = {José Nataniel Andrade de Sá and Marcelo Rodrigo Portela Ferreira and Francisco de Assis Tenorio de Carvalho}, doi = {10.1016/j.neucom.2025.130849}, journal = {Neurocomputing}, month = {10}, pages = {130849}, shortjournal = {Neurocomputing}, title = {Kernel clustering with automatic variable weighting for interval data}, volume = {650}, year = {2025}, }
• (2025). Combinations of generative adversarial network and reinforcement learning: A survey. NEUCOM, 650, 130847. (www)

  @article{ , abstract = {A generative adversarial network (GAN) consists of a generator and a discriminator, which extract features and compete through min–max optimization in a zero-sum two-player game. Reinforcement learning (RL) enables a decision maker (agent) to observe, learn, and select optimal actions within a dynamic environment over time. The combination of these two artificial intelligence approaches, referred to as GAN-RL, presents a powerful synergy that leverages their strengths while addressing their respective limitations. This paper provides the first comprehensive review that systematically examines GAN-RL, offering an in-depth analysis of its theoretical foundations, key methodologies, and diverse applications. Specifically, we categorize GAN-RL approaches based on five primary objectives: (a) generating comprehensive synthetic datasets for training; (b) minimizing the discrepancies between value (or cumulative reward) distributions, where the cumulative reward is defined as the total sum of rewards an agent accumulates over the course of its interactions with the environment; (c) predicting the next scenario; (d) predicting the gradient of the generator in discrete tasks; and (e) learning from demonstrations. Our analysis highlights how GAN-RL enhances system performance, including improvements in cumulative reward, convergence rate, and prediction accuracy. By identifying open challenges and future research directions, this review fills a critical gap in the literature and serves as a valuable resource for researchers exploring the intersection of GANs and RL.}, archive = {J_NEUCOM}, author = {Kok-Lim Alvin Yau and Yung-Wey Chong and Xiumei Fan and Faranak Nejati and Mohammad Kazem Chamran and Shalini Darmaraju}, doi = {10.1016/j.neucom.2025.130847}, journal = {Neurocomputing}, month = {10}, pages = {130847}, shortjournal = {Neurocomputing}, title = {Combinations of generative adversarial network and reinforcement learning: A survey}, volume = {650}, year = {2025}, }
• (2025). Automated neuropsychological scale assessment via LLMs: Patient simulation and prompting for scalable AD screening. NEUCOM, 650, 130846. (www)

  @article{ , abstract = {Alzheimer’s disease (AD) is a prevalent neurodegenerative disorder affecting the older adult population worldwide. Among the available cognitive screening methods, neuropsychological scale assessments (NSA) are widely employed in clinical practice to evaluate patients’ cognitive functions, particularly through speech-related tasks. However, manual evaluation of NSA is often labor-intensive and susceptible to subjective bias. While Automatic Speech Recognition (ASR) technology offers the potential to automate this process, it frequently faces notable challenges in handling dialectal variations, unclear pronunciations, and insufficient annotated data. These factors often lead to inaccuracies in transcribed speech and hinder the effective validation of assessment outcomes. To address these challenges, particularly the data bottleneck, we propose a novel dual-use paradigm for Large Language Models (LLMs). In this study, we designed an innovative agent framework employing LLMs not only for automated NSA scoring, thereby replacing manual assessment, but also for reverse-engineering clinical data by simulating patients with varied cognitive impairments. These synthetically generated NSA texts are then utilized in a closed-loop learning process as a form of “supervised learning” to iteratively refine the system prompts of an NSA evaluation agent. The prompt design explicitly incorporates the dialectal speech and pronunciation ambiguities, aiming to equip the NSA evaluation agent with evaluation and reasoning proficiencies comparable to those of experienced clinicians. The evaluations on a dataset of real-world clinical cases demonstrated a substantial 86% consistency between the NSA evaluation agent’s assessments and those of expert physicians, indicating the potential of LLM-driven approaches for accurate and scalable automated NSA in AD screening and beyond.}, archive = {J_NEUCOM}, author = {Meiwei Zhang and Yuwei Pan and Qiushi Cui and Yang Lü and Weihua Yu and Wenyuan Li}, doi = {10.1016/j.neucom.2025.130846}, journal = {Neurocomputing}, month = {10}, pages = {130846}, shortjournal = {Neurocomputing}, title = {Automated neuropsychological scale assessment via LLMs: Patient simulation and prompting for scalable AD screening}, volume = {650}, year = {2025}, }
• (2025). Conditional variational underwater image enhancement with kernel decomposition and adaptive hybrid normalization. NEUCOM, 650, 130845. (www)

  @article{ , abstract = {Enhancing underwater images has gained significant interest due to its wide range of applications in ocean engineering and marine robotics. However, underwater images often suffer from ambiguous degradations, making it difficult to construct a deterministic map between distorted underwater images and reference images. In addition, the distortions and artifacts affect the quality of underwater images in local details and global structures. To this end, we propose a novel network based on C onditional V ariational Autoencoder (CVAE) for U nderwater I mage E nhancement, named CVUIE. Specifically, to capture inherent uncertainties in underwater scenes and generate robust enhanced output, we propose a novel network structure that combines the CVAE with adversarial learning. Then, we present a Kernel Decomposition Attention (KDA) module to process and enhance features over a broader respective field. Moreover, to balance the complex details and global structures of enhanced images, we design a Probabilistic Adaptive Hybrid Normalization (PAHN) module. Evaluations conducted on multiple benchmark datasets prove that the proposed network performs qualitatively and quantitatively better than existing state-of-the-art methods. Real-world experiments have also demonstrated the promising future application prospects of our method.}, archive = {J_NEUCOM}, author = {Haopeng Zhang and Hongli Xu and Hao Liu and Xiaosheng Yu and Xiangyue Zhang and Chengdong Wu}, doi = {10.1016/j.neucom.2025.130845}, journal = {Neurocomputing}, month = {10}, pages = {130845}, shortjournal = {Neurocomputing}, title = {Conditional variational underwater image enhancement with kernel decomposition and adaptive hybrid normalization}, volume = {650}, year = {2025}, }
• (2025). Understanding beyond outputs: A novel knowledge distillation method using schur decomposition. NEUCOM, 650, 130841. (www)

  @article{ , abstract = {Recent advancements in knowledge distillation (KD) have highlighted its potential for enhancing model performance, especially in lightweight networks. Traditional approaches primarily focus on aligning the outputs of teacher and student networks through logit distillation or feature distillation at a numerical level. However, these methods often overlook the necessity for students to not only replicate correct answers but also to grasp the underlying logic and essence of the predictions, particularly in scenarios where student networks are significantly lighter than their teacher counterparts. In this paper, we propose a novel perspective on KD that emphasizes understanding over mere replication. We extract deeper insights from the logits by constructing Gram and covariance matrices, which capture the intricate relationships among features. Utilizing Schur decomposition, we uncover similarities in feature space distributions, enabling the student to learn not only the final outputs but also the essential interactions between features that contribute to those outputs. Moreover, we provide a plug-and-play loss function that can seamlessly integrate with existing state-of-the-art logit-based methods, leading to enhanced performance across various teacher and student network architectures. Extensive experiments on CIFAR-100 and ImageNet-2012 demonstrate the effectiveness of our method.}, archive = {J_NEUCOM}, author = {Chang-Ming Pan and Si-Bao Chen and Bo Jiang and Bin Luo}, doi = {10.1016/j.neucom.2025.130841}, journal = {Neurocomputing}, month = {10}, pages = {130841}, shortjournal = {Neurocomputing}, title = {Understanding beyond outputs: A novel knowledge distillation method using schur decomposition}, volume = {650}, year = {2025}, }
• (2025). Siamese neural network-based stationary feature extraction for nonstationary process monitoring. NEUCOM, 650, 130840. (www)

  @article{ , abstract = {Industrial processes often involve complex mechanisms, fluctuating operating conditions, and intricate procedural couplings that generate data exhibiting nonstationarity, non-Gaussianity, nonlinearity, and dynamic behavior. These characteristics present significant challenges for process monitoring and fault diagnosis. This paper introduces two innovative strategies based on Siamese neural networks to extract stationary features, thereby extending conventional stationary subspace analysis (SSA). The first strategy extracts time-invariant features while preserving critical information, and the second employs adversarial mechanisms to identify common features across different time periods. Compared with traditional SSA, the proposed methods effectively capture the dynamics, nonlinearity, and non-Gaussianity inherent in the data, substantially enhancing their applicability in complex industrial environments. The efficacy of these approaches is demonstrated using data from an ironmaking process, underscoring their potential for robust industrial applications.}, archive = {J_NEUCOM}, author = {Hanwen Zhang and Qingqing Liu and Weiwei Fan and Jun Shang and Linlin Li}, doi = {10.1016/j.neucom.2025.130840}, journal = {Neurocomputing}, month = {10}, pages = {130840}, shortjournal = {Neurocomputing}, title = {Siamese neural network-based stationary feature extraction for nonstationary process monitoring}, volume = {650}, year = {2025}, }
• (2025). AdaPAC: Prototypical anchored contrastive test time adaptation for domain generalization. NEUCOM, 650, 130839. (www)

  @article{ , abstract = {Test time adaptation (TTA) aims to address distribution shifts between training and testing data by adjusting a given model based on the testing sample. This method is particularly suitable for Domain Generalization (DG) scenarios, where it has access to online test data during inference. Most current TTA techniques modify the source model using the model’s predictions on the test data as pseudo-labels. However, under test-time domain shifts, the accuracy of these pseudo-labels is not guaranteed, potentially leading to performance declines after adaptation. Additionally, adapting all the model parameters using test samples can be harmful due to the presence of noisy samples. To tackle these issues, we propose AdaPAC, a prototypical anchored contrastive test-time adaptation approach. This method leverages knowledge from source data, specifically subclass prototypes, and incorporates them into prototypical contrastive learning within the representation space. By doing so, it effectively aligns the test and source distributions. AdaPAC categorizes test samples as reliable or unreliable based on their distances from the anchored prototypes and applies prototypical and instance-wise contrastive learning to them, respectively. Furthermore, we propose a reliable weight regularization for unreliable samples that encourages the model to update parameters that are sensitive to the distribution shift of reliable test samples. AdaPAC outperformed the current state-of-the-art DG and TTA methods across domain generalization benchmarks including VLCS, PACS, OfficeHome, and TerraIncognita.}, archive = {J_NEUCOM}, author = {Saeed Karimi and Hamdi Dibeklioğlu}, doi = {10.1016/j.neucom.2025.130839}, journal = {Neurocomputing}, month = {10}, pages = {130839}, shortjournal = {Neurocomputing}, title = {AdaPAC: Prototypical anchored contrastive test time adaptation for domain generalization}, volume = {650}, year = {2025}, }
• (2025). Joint Global–Local frames modeling to enhance semantic alignment for zero-shot long video editing. NEUCOM, 650, 130836. (www)

  @article{ , abstract = {Semantic alignment and time consistency have been significant challenges in the editing of long videos. Previous work addresses this by compressing spatial–temporal features into basis vectors to maintain temporal consistency throughout the DDIM inversion process. However, as the number of video frames increases, the capacity of these basis vectors to represent spatial information diminishes, leading to poor semantic alignment. To address this issue, we propose a method called global–local frames modeling, which retains temporal consistency while enhancing the semantic alignment between text and video frames during editing tasks. This method involves global frames modeling, local frames modeling, and spatial–temporal fusion as its three core components. Global frames modeling reduces the dimensionality of all frames in a long video to capture primary temporal features, while local frames modeling segments the long video into clips and performs probabilistic clustering on each clip individually to capture spatial visual semantic details. Spatial–temporal fusion is designed to achieve the integration of spatial and temporal features. Notably, our method improves the BLIP score by 1.95% when the number of basis vectors is the same as in the state-of-the-art method, while also reducing memory usage by over 50%. Our code is available at https://github.com/MarcYugo/global-local-frames-modeling .}, archive = {J_NEUCOM}, author = {Zewen Yu and Pengchong Qiao and Jie Chen and Xiaoqin Zhang}, doi = {10.1016/j.neucom.2025.130836}, journal = {Neurocomputing}, month = {10}, pages = {130836}, shortjournal = {Neurocomputing}, title = {Joint Global–Local frames modeling to enhance semantic alignment for zero-shot long video editing}, volume = {650}, year = {2025}, }
• (2025). Towards bandwidth efficient edge–cloud collaborative deep learning with data importance driven compression. NEUCOM, 650, 130835. (www)

  @article{ , abstract = {The proliferation of mobile devices, exponential growth in data, and rapid development in edge computing have given rise to a distributed computing architecture for deep learning training and inference, characterized by end–edge–cloud collaboration. Consequently, investigating deep learning frameworks based on edge–cloud collaboration holds paramount importance. We propose Data Importance-driven Compression (DIC), a bandwidth-efficient edge–cloud collaborative framework for training DNN-based classifiers. DIC introduces the novel concept of data importance for classification tasks in edge–cloud collaborative training, where the compression degree of uploaded data is dynamically determined by its assessed importance. This innovation effectively balances accuracy preservation and bandwidth consumption, thereby enhancing the overall performance of distributed training frameworks. Specifically, DIC introduces a lightweight classification-guided autoencoder that dynamically adjusts compression ratios per sample based on their importance, preserving critical features while minimizing bandwidth usage. This approach ensures the generation of more accurate advanced models on the cloud through collaborative learning. Extensive experimental results demonstrate that DIC achieves a reduction in bandwidth consumption by 3.6-7 × , with an accuracy loss of less than 0.6%. Furthermore, it outperforms existing compression methods by approximately 4% in classification accuracy under identical conditions, validating its effectiveness in preserving critical learning features during edge–cloud collaborative training.}, archive = {J_NEUCOM}, author = {Yalin Jiang and Peng Zhao and Cong Zhao and Jie Lin}, doi = {10.1016/j.neucom.2025.130835}, journal = {Neurocomputing}, month = {10}, pages = {130835}, shortjournal = {Neurocomputing}, title = {Towards bandwidth efficient edge–cloud collaborative deep learning with data importance driven compression}, volume = {650}, year = {2025}, }
• (2025). Decentralized neighboring information fusion for traffic network signal control. NEUCOM, 650, 130834. (www)

  @article{ , abstract = {Optimizing traffic network signal control strategies through Multi-Agent Reinforcement Learning (MARL) shows considerable promise in improving road network traffic efficiency. Current MARL algorithms for multi-intersection traffic network signal control predominantly rely on a server for collecting global information or aggregating local models. Such reliance suffers from the curse of dimensionality and vulnerability to single-point failures. In this study, we proposed D-NIF, a novel fully decentralized MARL algorithm to address multi-intersection TSC problems. A primary challenge in designing this algorithm is to enhance the efficiency of the traffic network signal control model without a central server, as communications are only allowed among connected neighbors. To address this challenge, we introduce the neighboring information fusion algorithm for effectively encoding the spatial coupling and temporal correlation in traffic flows. Additionally, we design the decentralized learning algorithm for aggregating local models without needing a central server. Our simulation analysis demonstrates that the introduced methods of neighboring information fusion and decentralized learning substantially enhance agents’ performance in traffic network signal control. Moreover, the proposed D-NIF algorithm surpasses existing state-of-the-art benchmarks.}, archive = {J_NEUCOM}, author = {Bo Liu and Xinyang Liu and Chang Chen and Jianwei Huang and Zhengtao Ding}, doi = {10.1016/j.neucom.2025.130834}, journal = {Neurocomputing}, month = {10}, pages = {130834}, shortjournal = {Neurocomputing}, title = {Decentralized neighboring information fusion for traffic network signal control}, volume = {650}, year = {2025}, }
• (2025). Co-speech video generation via motion transfer based on diffusion models. NEUCOM, 650, 130833. (www)

  @article{ , abstract = {Co-speech video generation aims to enhance the realism of virtual characters by synchronizing gestures and facial expressions with speech. However, existing methods face challenges such as blurred hand motion, occlusion, and inaccurate gestures, which compromise visual quality and temporal consistency. To address these issues, we propose a novel Motion Transfer-based Co-speech Video Generation (MTCVG) method that integrates an enhanced Thin Plate Spline (TPS) transformation and diffusion models. Specifically, our approach consists of two stages: First, we designed a motion transfer module that introduces a hand weighting generation network and a gradient optimization strategy to improve the TPS transformation. This module dynamically adjusts the significance of hand movements based on keypoint relevance, enabling more precise image warping and enhancing the clarity of hand gestures. Second, we incorporated a conditional motion diffusion module based on Denoising Diffusion Probabilistic Models (DDPM), which utilizes speech input to ensure smooth and consistent temporal alignment between motions and speech. Experiments on the PATS and BEAT2 datasets demonstrate that MTCVG surpasses existing methods in generating co-speech videos with more accurate gestures and enhanced appearance features, achieving state-of-the-art results in both qualitative and quantitative evaluations.}, archive = {J_NEUCOM}, author = {Jiye Zhang and Zhiye Chen and Ruidi Zheng and Ruoyu Liu and Xiuhua Jiang}, doi = {10.1016/j.neucom.2025.130833}, journal = {Neurocomputing}, month = {10}, pages = {130833}, shortjournal = {Neurocomputing}, title = {Co-speech video generation via motion transfer based on diffusion models}, volume = {650}, year = {2025}, }
• (2025). Anomaly detection in attributed networks via local multi-order contrastive learning and global topology awareness. NEUCOM, 650, 130829. (www)

  @article{ , abstract = {Detecting anomalies in attributed networks has become increasingly important in practical applications such as cybersecurity and e-commerce. Most existing methods primarily focus on the local graph dimension, relying on a single substructure to construct contrastive instance pairs. These approaches limit the representation of interactive relationships within the network, negatively impacting the performance of subsequent learning tasks. Moreover, current methods inadequately consider global topology information in attributed networks, which hinders their ability to capture abnormal topology patterns within the graphs. To address these limitations, we propose LMGTA ( L ocal M ulti-Order Contrastive Learning and G lobal T opology A wareness), a novel framework for anomaly detection in attributed networks. In the local dimension, LMGTA utilizes a specially designed data augmentation strategy to enhance local graph representations by enriching the network’s structural diversity and broadening interaction relationships within subgraphs. These enhanced subgraphs are then processed with a contrastive learning model to capture comprehensive local neighborhood information. At the global level, LMGTA effectively links node attributes with global topology through community structures, enabling the identification of topology-based anomalies. By integrating complementary local and global views, LMGTA achieves precise detection of anomalous nodes. Extensive experiments on eight real-world datasets demonstrate the effectiveness and superiority of LMGTA compared to state-of-the-art baseline methods.}, archive = {J_NEUCOM}, author = {Mark Junjie Li and Gen Zhao and Sunjie Huang and Qin Zhang and Jiang Liu and Meiting Li and Jun Li}, doi = {10.1016/j.neucom.2025.130829}, journal = {Neurocomputing}, month = {10}, pages = {130829}, shortjournal = {Neurocomputing}, title = {Anomaly detection in attributed networks via local multi-order contrastive learning and global topology awareness}, volume = {650}, year = {2025}, }
• (2025). A centroid-based fine-tuning method for out-of-scope classification. NEUCOM, 650, 130828. (www)

  @article{ , abstract = {Accurately detecting out-of-scope queries is a challenging task in task-oriented dialog systems. Most existing research focus on adding an outlier detector after classification or designing an open world classification to identify unknown intents. There is still a major performance gap on achieving high efficiency and accuracy based on above methods. In our research, we tend to solve this problem by constructing an out-of-scope class in the classification. We propose an explainable centroid-based fine-tuning method including a modified decision metric (MDM) and a centroid-based cosine loss (CCL) on Pre-trained Transformer models for optimization. This loss function builds on Copernican structure and assigns the same margin to each in-scope class to resolve an ambiguous configuration on out-of-scope detection. Moreover, cosine similarity is utilized to remove radial variations of centroids. Experimental results show that our proposed method achieves improvement compared to other baseline methods.}, archive = {J_NEUCOM}, author = {Xinyi Cai and Pei-Wei Tsai and Youwen Zhang and Jiao Tian and Kai Zhang and Ke Yu and Hongwang Xiao and Jinjun Chen}, doi = {10.1016/j.neucom.2025.130828}, journal = {Neurocomputing}, month = {10}, pages = {130828}, shortjournal = {Neurocomputing}, title = {A centroid-based fine-tuning method for out-of-scope classification}, volume = {650}, year = {2025}, }
• (2025). S2TCR: Cross complementary semantics and structure architectures within complex glyph shapes for scene traditional chinese recognition. NEUCOM, 650, 130826. (www)

  @article{ , abstract = {Scene text recognition (STR) has made significant progress due to its wide range of applications, but traditional Chinese (Trd-CN) is not included in these advances. Compared to simplified Chinese (Sim-CN), traditional Chinese characters have more complex and diverse structures, more numerous and dense strokes, posing greater challenges for character feature extraction and parsing. The complementarily semantic meaning and structural form of complex Trd-CN glyph can capture various aspects of the characters, which has not been addressed in previous research on simplified glyph. In this article, we propose an innovative scene traditional Chinese recognition method called S 2 T C R , which leverages multi-hierarchical feature interaction to enhance glyph representation learning through cross-complementary fusion of comprehensive semantics meaning and spatial structural form cues. S 2 T C R includes the Glyph Hierarchical Windows Partition (GHWP), Cross Complementary Semantics Attention (CSemA) and Cross Complementary Structure Attention (CStrA). Specifically, GHWP maintains the relative coherence of the complex glyph features. CSemA supplements semantic meaning information for low-hierarchical deficiency, while CStrA supplements structural form information for high-hierarchical deficiency. Experimental results verify the over 11.80% significant superiority of our proposed method on scene traditional Chinese recognition and also achieve state-of-the-art performance on several artistic and general Chinese datasets.}, archive = {J_NEUCOM}, author = {Haobo Yang and Zhichong Li and Xiangqian Wu}, doi = {10.1016/j.neucom.2025.130826}, journal = {Neurocomputing}, month = {10}, pages = {130826}, shortjournal = {Neurocomputing}, title = {S2TCR: Cross complementary semantics and structure architectures within complex glyph shapes for scene traditional chinese recognition}, volume = {650}, year = {2025}, }
• (2025). Analyzing and mitigating bias of facial attribute classifiers using ChatGPT. NEUCOM, 650, 130825. (www)

  @article{ , abstract = {Recent studies have demonstrated the potential of ChatGPT in facial attribute classification, achieving accuracies comparable to those of vision models. However, limited research has been conducted on the fairness of ChatGPT in this context. This paper investigates the fairness of ChatGPT in facial attribute classification and examines how its natural language explanations (linguistic features) can enhance both fairness and explainability in traditional vision model for facial attribute classification by proposing: (a) a novel training framework that integrates ChatGPT’s linguistic features with vision models to improve both classification accuracy and fairness, and (b) a mapping function between the feature spaces of vision and language models, enabling ChatGPT to provide natural language explanations of the vision model’s decision-making process. To evaluate our approach, we conduct two types of experiments: (1) face-based gender classification, where race is considered the protected attribute, and (2) gender-independent classification of 13 facial attributes, where gender is the protected attribute. Our experimental results on popular facial attribute datasets show that ChatGPT offers complementary strengths in obtaining fairer outcomes over vision classifiers for different facial attribute classification tasks across demographic attributes. Moreover, incorporating ChatGPT’s linguistic descriptions enhances vision model accuracy by 1.16% to 1.34%, reduces demographic performance disparities by 1.23% to 11.32%, and provides interpretable explanations, all without altering the architecture of the vision model.}, archive = {J_NEUCOM}, author = {Ayesha Manzoor and Bharath Krishnamurthy and Ajita Rattani}, doi = {10.1016/j.neucom.2025.130825}, journal = {Neurocomputing}, month = {10}, pages = {130825}, shortjournal = {Neurocomputing}, title = {Analyzing and mitigating bias of facial attribute classifiers using ChatGPT}, volume = {650}, year = {2025}, }
• (2025). Robust multi-expert deep learning framework for brain MRI classification with taguchi optimization. NEUCOM, 650, 130824. (www)

  @article{ , abstract = {Deep learning (DL) has significantly transformed image classification, particularly in medical imaging. However, challenges such as high computational costs, large parameter counts, noisy data, uncertainty in medical images, adversarial vulnerability, and complex hyperparameter optimization limit the deployment of convolutional neural networks (CNNs) in resource-constrained environments. This paper proposes a novel Multi-Expert Deep Learning framework for Brain MRI classification that integrates Denoising, Fuzzy logic, Adversarial training, and Taguchi optimization to address these issues. The framework comprises three expert models: (i) Denoising Expert, which enhances image quality by removing noise and achieves post-denoising Peak Signal-to-Noise Ratio (PSNR) of 58.09 dB with Structural Similarity Index (SSIM) of 0.9875; (ii) Fuzzy Logic Expert, which effectively handles uncertainty and achieves classification accuracy of 94.13% in low-intensity MRI images, and (iii) Adversarially Trained Expert, which strengthens robustness against adversarial perturbations, improving accuracy under FGSM and PGD attacks from 68.89% to 85.72% and 69.09% to 84.35%, respectively. These experts are fused through the proposed Multi-Expert Fusion (MEF) model, dynamically selecting the most relevant features for robust classification. The Taguchi optimization method is applied for efficient hyperparameter tuning, achieving state-of-the-art classification accuracy of 97.25% and 99.10% on Brain MRI datasets while significantly reducing computational complexity. The proposed model maintains a lightweight structure with only 738,100 parameters and size of 2.82 MB, efficient for real-world medical imaging applications. The proposed framework demonstrates superior robustness, computational efficiency, and classification accuracy compared to existing CNN-based models, making it well-suited for healthcare applications in resource-constrained environments.}, archive = {J_NEUCOM}, author = {Anshu Singh and Maheshwari Prasad Singh and Amit Kumar Singh}, doi = {10.1016/j.neucom.2025.130824}, journal = {Neurocomputing}, month = {10}, pages = {130824}, shortjournal = {Neurocomputing}, title = {Robust multi-expert deep learning framework for brain MRI classification with taguchi optimization}, volume = {650}, year = {2025}, }
• (2025). RecSal-net: Recursive saliency network for video saliency prediction. NEUCOM, 650, 130822. (www)

  @article{ , abstract = {Video saliency prediction, which emulates the selective visual processing mechanisms of the human visual system, has found widespread applications across various domains. However, existing models often struggle to effectively integrate spatiotemporal features, particularly due to challenges associated with multi-resolution feature processing. To address these limitations, we propose the Recursive Saliency Network (RecSal-Net), a model designed to enhance the performance of video saliency prediction. The model adopts the Video Swin Transformer as its backbone to extract rich spatiotemporal features. In addition, a Recursive Feature Pyramid structure is introduced to integrate multi-resolution features while minimizing information loss. To further improve feature representation, a top-down feature integration strategy is employed, transferring high-level semantic features to lower-level feature maps. This is complemented by iterative upsampling operations, which enrich both high- and low-resolution representations. Experimental results demonstrate that RecSal-Net outperforms state-of-the-art methods on the DHF1K, Hollywood-2, and UCF Sports datasets, achieving superior performance across key evaluation metrics such as AUC-J, CC, and NSS. These findings validate the model’s effectiveness in capturing long-range spatiotemporal dependencies and integrating multi-resolution features. Overall, our work underscores the potential of recursive feature modeling to advance future video saliency prediction frameworks. The code is available at https://github.com/affctivai/RecSal-Net .}, archive = {J_NEUCOM}, author = {ChaeEun Woo and SuMin Lee and Soo Min Park and Byung Hyung Kim}, doi = {10.1016/j.neucom.2025.130822}, journal = {Neurocomputing}, month = {10}, pages = {130822}, shortjournal = {Neurocomputing}, title = {RecSal-net: Recursive saliency network for video saliency prediction}, volume = {650}, year = {2025}, }
• (2025). Distributed nash equilibrium seeking for multi-coalition games with constrained NASVs: A hierarchical prescribed-performance algorithm. NEUCOM, 650, 130821. (www)

  @article{ , abstract = {This paper investigates the distributed Nash equilibrium seeking problem for multi-coalition games with constrained networked autonomous surface vehicles (NASVs), involving uncertain dynamics and external disturbances. A hierarchical framework centered around prescribed-performance principles is developed for the design of a distributed Nash equilibrium seeking algorithm. In the first layer, multiple consensus protocols are adopted for the NASVs to exchange information both within the same coalition and between different coalitions under a unified communication graph. For enabling the NASVs to strategize their expected states within a target zone, the second layer utilizes the gradient descent method incorporating the virtual information and a projection operator. As for the third layer, an adaptive neural network is employed to approximate nonlinear dynamics and external disturbances, while a buffer zone is established by means of the barrier function approach, making sure that the NASVs’ actual strategies can track individual anticipated states without violating the constraints. Besides, the convergence results can be hierarchically obtained through the utilization of nonsmooth Lyapunov analysis, and numerical simulations substantiate the efficacy of the proposed algorithm.}, archive = {J_NEUCOM}, author = {Wen-Jin Liu and Xiang-Yu Yao and Menghu Hua and Ju H. Park and Ming-Feng Ge}, doi = {10.1016/j.neucom.2025.130821}, journal = {Neurocomputing}, month = {10}, pages = {130821}, shortjournal = {Neurocomputing}, title = {Distributed nash equilibrium seeking for multi-coalition games with constrained NASVs: A hierarchical prescribed-performance algorithm}, volume = {650}, year = {2025}, }
• (2025). Pair-mode tensor wheel decomposition with randomized block krylov iteration. NEUCOM, 650, 130820. (www)

  @article{ , abstract = {The tensor wheel (TW) decomposition offers a unified framework for representing high-order tensors through a set of latent factors combined by multilinear operations in a wheel-like topology. This generalization encompasses several well-known models, including tensor ring, tensor train, Tucker, and CP decompositions. However, its baseline formulation can suffer from computational inefficiencies due to the high cost of updating latent factors, which involves large-scale operations. To address these limitations and enhance the model’s capability to capture orthogonally constrained features, we propose the pair-mode tensor wheel (PM-TW) decomposition. It retains the wheel-like topology but introduces a novel structure in which the ring factors jointly represent pairs of modes, leading to a more compact representation. This modification significantly alleviates the storage and computational complexities associated with high-dimensional data, especially within a practical range of parameter settings. The proposed topology enables more efficient computation of latent factors using a recursive SVD-based updating scheme, which can be further accelerated via randomization techniques. In particular, we extend this approach by incorporating randomized block Krylov iteration (rBKI), resulting in a scalable randomized PM-TW algorithm with quasi-linear computational complexity. Extensive experiments demonstrate that the proposed methods not only outperform existing models in capturing multilinear structures from tensors generated by various decomposition models, but also exhibit strong potential for tackling low-rank tensor completion tasks.}, archive = {J_NEUCOM}, author = {Rafał Zdunek}, doi = {10.1016/j.neucom.2025.130820}, journal = {Neurocomputing}, month = {10}, pages = {130820}, shortjournal = {Neurocomputing}, title = {Pair-mode tensor wheel decomposition with randomized block krylov iteration}, volume = {650}, year = {2025}, }
• (2025). Stability and associative memories of multi-layer memristive neural networks in the flux-charge domain. NEUCOM, 650, 130817. (www)

  @article{ , abstract = {The stability and associative memories of multi-layer memristive neural networks (MMNNs) are discussed in this article. According to a novel relation between charge and flux of memristor, several conditions are established to ensure μ -stability of MMNNs. Depending on given conditions, the estimation of attraction basin for stable equilibrium point (EP) in MMNNs is discussed by means of analysis. Compared with standard memristive neural networks (MNNs), the flux tends to the constant and the voltage disappears during the steady state of MMNNs, which facilitates low consumption and information storage. Furthermore, the dynamic associative memory of multi-layer images is designed based on the architecture and stability of MMNNs. Different from general associative memory, dynamic associative memory can store multi-layer images in different layers of MMNNs independently, and through the stability, the images of each layer recall the memory pattern. Two numerical examples are presented to verify the main results.}, archive = {J_NEUCOM}, author = {Fanghai Zhang and Peng Gao and Tingwen Huang and Zhigang Zeng}, doi = {10.1016/j.neucom.2025.130817}, journal = {Neurocomputing}, month = {10}, pages = {130817}, shortjournal = {Neurocomputing}, title = {Stability and associative memories of multi-layer memristive neural networks in the flux-charge domain}, volume = {650}, year = {2025}, }
• (2025). LexiSem: A re-ranker balancing lexical and semantic quality for enhanced abstractive summarization. NEUCOM, 650, 130816. (www)

  @article{ , abstract = {Sequence-to-sequence neural networks have recently achieved significant success in abstractive summarization, especially through fine-tuning large pre-trained language models on downstream datasets. However, these models frequently suffer from exposure bias, which can impair their performance. To address this, re-ranking systems have been introduced, but their potential remains underexplored despite some demonstrated performance gains. Most prior work relies on ROUGE scores and aligned candidate summaries for ranking, exposing a substantial gap between semantic similarity and lexical overlap metrics. In this study, we demonstrate that a second-stage model can be trained to re-rank a set of summary candidates, significantly enhancing performance. Our novel approach leverages a re-ranker that balance lexical and semantic quality. Additionally, we introduce a new strategy for defining negative samples in ranking models. Through experiments on the CNN/DailyMail, XSum and Reddit TIFU datasets, we show that our method effectively estimates the semantic content of summaries without compromising lexical quality. In particular, our method sets a new performance benchmark on the CNN/DailyMail dataset (48.18 R1, 24.46 R2, 45.05 RL) and on Reddit TIFU (30.37 R1,RL 23.87).}, archive = {J_NEUCOM}, author = {Eman Aloraini and Hozaifa Kassab and Ali Hamdi and Khaled Shaban}, doi = {10.1016/j.neucom.2025.130816}, journal = {Neurocomputing}, month = {10}, pages = {130816}, shortjournal = {Neurocomputing}, title = {LexiSem: A re-ranker balancing lexical and semantic quality for enhanced abstractive summarization}, volume = {650}, year = {2025}, }
• (2025). Lag consensus and lag h∞ consensus for multiagent systems subject to topology attacks based on PD control and PI control. NEUCOM, 650, 130815. (www)

  @article{ , abstract = {The proportional–derivative (PD) control and proportional–integral (PI) control approaches are utilized to tackle the lag and lag H ∞ consensus for multiagent systems (MASs) suffering topology attacks. A lag consensus criterion for the MAS is presented with the help of graph theory, matrix theory, and the devised PD control scheme. Moreover, a sufficient condition to guarantee the lag consensus of the MAS is also derived by exploiting Barbalat’s lemma, the properties of Kronecker product, and the designed PI control strategy. In addition, two lag H ∞ consensus criteria are acquired for the MAS under external disturbances by utilizing PD control and PI control approaches. Finally, two numerical examples are afforded to validate the devised control schemes.}, archive = {J_NEUCOM}, author = {Jin-Liang Wang and Xin-Yu Zhao and Yi-Ming Zhai}, doi = {10.1016/j.neucom.2025.130815}, journal = {Neurocomputing}, month = {10}, pages = {130815}, shortjournal = {Neurocomputing}, title = {Lag consensus and lag h∞ consensus for multiagent systems subject to topology attacks based on PD control and PI control}, volume = {650}, year = {2025}, }
• (2025). A SEIHR knowledge dissemination model with time-varying parameters based on physics-informed neural networks. NEUCOM, 650, 130812. (www)

  @article{ , abstract = {In recent years, researchers have considered the influence of different factors on the knowledge dissemination process, and they have made the model more complicated by investigating the model in depth from different angles. However, there is a lack of consideration of the knowledge-sharing willingness of the individuals who hold the knowledge. In addition, various knowledge dissemination models are frequently complicated by the interaction of several real factors and carry many unknown time-varying parameters. Therefore, this research provides a time-varying parameter for the susceptible–exposed–infected/hesitant-recovered (SEIHR) knowledge dissemination model that considers knowledge-sharing willingness based on the susceptible–exposed–infected–recovered (SEIR) model. The population can be divided into disseminators and non-disseminators based on the willingness of those who possess knowledge to share it, allowing for a more precise description of the knowledge dissemination process and the study of its dynamics. Meanwhile, this paper solves the forward and inverse problems of the SEIHR model by adding the consideration of time-varying parameters to a deep learning method called Physics-Informed Neural Networks (PINNs): the forward direction solves the set of Ordinary Differential Equations (ODEs) and predicts the short-term future data, and the inverse direction extracts the unknown time-varying parameters. The experimental findings demonstrate that motion trends of unknown time-varying parameters can be successfully extracted from the data by PINNs employing time-varying parameters. By comparing with the Maximum Likelihood Estimation (MLE) method, it can be seen that in addition to the higher accuracy in fitting time-varying parameters, the PINNs method also possesses the ability to predict the parameter changes in a short time in the future. Through a comparative analysis between PINNs with time-varying parameters, PINNs with constant parameters, and ordinary Multi-layer Perceptron (MLP), it is discovered that PINNs with time-varying parameters have better results in both solution and prediction.}, archive = {J_NEUCOM}, author = {Dan Xia and Boyang Diao and Qiusha Min and Jun Mei}, doi = {10.1016/j.neucom.2025.130812}, journal = {Neurocomputing}, month = {10}, pages = {130812}, shortjournal = {Neurocomputing}, title = {A SEIHR knowledge dissemination model with time-varying parameters based on physics-informed neural networks}, volume = {650}, year = {2025}, }
• (2025). ACNTrack: Agent cross-attention guided multimodal multi-object tracking with neural kalman filter. NEUCOM, 650, 130811. (www)

  @article{ , abstract = {Exploring and associating the complementary information from visible, thermal infrared, and low-light images is crucial for advancing Multimodal Multi-Object Tracking (MMOT). While previous studies have shown that efficient feature fusion modules can bolster tracking performance in complex environments, these methods often encounter constraints in global feature interaction and computational efficiency. We present a novel multimodal multi-object tracker based on a tracking-by-detection paradigm, comprising a multimodal detector and a data associator. A dual cross-attention feature fusion detection framework, predicated on an agent attention mechanism, is introduced to enhance feature interaction efficiency and effectively capture cross-modal complementary information. To more accurately capture detailed and complex information inherent in each modality, we propose a Feature Pyramid Shared Convolution (FPS-Conv) operation to supersede the Spatial Pyramid Pooling Fast (SPPF) operation within the detector. Additionally, a Neural Kalman Filter (NKF) is developed to augment the performance of the data associator, which dynamically adjusts process and observation noise in accordance with the current motion state. Our innovative fusion architecture significantly reduces computational complexity while maintaining high-quality feature interactions, and our proposed NKF demonstrates superior performance in handling diverse motion patterns compared to traditional fixed-parameter approaches. Experimental results validate these advantages, with our proposed method achieving state-of-the-art results on the KAIST, FLIR, and UniRTL test datasets and demonstrated competitive performance on the VT-MOT dataset.}, archive = {J_NEUCOM}, author = {Lian Zhang and Lingxue Wang and Yuzhen Wu and Mingkun Chen and Dezhi Zheng and Yi Cai}, doi = {10.1016/j.neucom.2025.130811}, journal = {Neurocomputing}, month = {10}, pages = {130811}, shortjournal = {Neurocomputing}, title = {ACNTrack: Agent cross-attention guided multimodal multi-object tracking with neural kalman filter}, volume = {650}, year = {2025}, }
• (2025). EmoVerse: Enhancing multimodal large language models for affective computing via multitask learning. NEUCOM, 650, 130810. (www)

  @article{ , abstract = {Affective computing is essential for applications such as human–computer interaction. While Multimodal Large Language Models (MLLMs) demonstrate impressive general capabilities, they face considerable challenges in affective computing, particularly in detecting subtle facial expressions and handling complex affective tasks, such as emotion reason inference and understanding emotions in multimodal long-context scenarios. Furthermore, there is a lack of a unified MLLM that can effectively handle a wide range of affective tasks. To address these challenges, we propose Emotion Universe (EmoVerse), a MLLM trained through a M ultistage M ultitask S entiment and E motion (M2SE) instruction tuning strategy. Through this training strategy, EmoVerse acquires the ability to deeply analyze the underlying reasons for affective states. Besides, to address the lack of multitask datasets in affective computing, we introduce the Affective Multitask (AMT) Dataset, which supports multimodal sentiment analysis, multimodal emotion recognition, facial expression recognition, emotion reason inference, and emotion cause-pair extraction tasks. Extensive experiments demonstrate that EmoVerse outperforms existing methods, achieving state-of-the-art results in affective tasks. The code is available at https://github.com/liaolea/EmoVerse .}, archive = {J_NEUCOM}, author = {Ao Li and Longwei Xu and Chen Ling and Jinghui Zhang and Pengwei Wang}, doi = {10.1016/j.neucom.2025.130810}, journal = {Neurocomputing}, month = {10}, pages = {130810}, shortjournal = {Neurocomputing}, title = {EmoVerse: Enhancing multimodal large language models for affective computing via multitask learning}, volume = {650}, year = {2025}, }
• (2025). BAST-mamba: Binaural audio spectrogram mamba transformer for binaural sound localization. NEUCOM, 650, 130804. (www)

  @article{ , abstract = {Accurate sound localization in reverberant environments is essential for human auditory perception. Recently, Convolutional Neural Networks (CNNs) have been used to model the binaural human auditory pathway. However, CNNs face limitations in capturing global acoustic features. To address this issue, we propose a novel end-to-end Binaural Audio Spectrogram Mamba Transformer (BAST-Mamba) model to predict sound azimuth in both anechoic and reverberant conditions. We explore two implementation modes: BAST-Mamba-SP and BAST-Mamba-NSP, which correspond to shared and non-shared parameter configurations, respectively. Our best model BAST-Mamba-SP, equipped with subtraction-based interaural integration and a hybrid loss function, achieves a state-of-the-art angular distance (AD) error of 0.89°and mean squared error of 0.0004, significantly outperforming baseline models. The model demonstrates generalization across acoustic environments, robust hemifield symmetry and high accurate real-time localization performance ( < 4°AD at 300 ms). Moderate noise augmentation at 30 dB SNR yields the strongest noise resilience. Explainability analyses highlight consistent frequency focus in the 2–3 kHz and 5.5–6.5 kHz bands, aligning with known neurophysiological cues. These results validate the potential of neurobiologically inspired Transformer for robust, high-precision sound localization and offer new insights into human sound localization.}, archive = {J_NEUCOM}, author = {Sheng Kuang and Jie Shi and Kiki van der Heijden and Siamak Mehrkanoon}, doi = {10.1016/j.neucom.2025.130804}, journal = {Neurocomputing}, month = {10}, pages = {130804}, shortjournal = {Neurocomputing}, title = {BAST-mamba: Binaural audio spectrogram mamba transformer for binaural sound localization}, volume = {650}, year = {2025}, }
• (2025). IMFR-net: Interval measurement and full recovery network for video compressive sensing. NEUCOM, 650, 130803. (www)

  @article{ , abstract = {Video compressive sensing (VCS) can reconstruct the video from a relatively small number of measurements, which has been extensively studied by many scholars. For VCS, how to exploit the spatiotemporal correlation in video for efficient sampling and reconstruction is the focus of the task. While the most existing methods adopt frame-by-frame observation strategies and have achieved notable improvements, the inherent potential of video signals remains underexplored. In this paper, we propose an Interval Measurement and Full Recovery Network (IMFR-Net) to utilize the spatiotemporal correlation for VCS. Technically, in measurement part, considering the huge redundancy in video, we propose an interval sensing strategy to measure the video frames at a certain temporal interval. In reconstruction part, we recover full frames including key frames and non-key frames. In detail, key frames are first reconstructed using spatial correlations, then non-key frames are synthesized by explicitly modeling the motion estimation and frame interpolation. Besides, the multi-level mechanism is designed to deal with different motions in video, which contributes to high-quality video reconstruction. Extensive experiments demonstrate that the proposed method achieves better quantitative and qualitative performance compared to the existing methods. Meanwhile, the proposed method allows video reconstruction at the frame rate of 48 fps on video sequence with 256 × 256 resolution.}, archive = {J_NEUCOM}, author = {Yong Zhang and Wanxin Zhang and Zhifu Zhao and Fu Li and Yi Niu and Jianan Li}, doi = {10.1016/j.neucom.2025.130803}, journal = {Neurocomputing}, month = {10}, pages = {130803}, shortjournal = {Neurocomputing}, title = {IMFR-net: Interval measurement and full recovery network for video compressive sensing}, volume = {650}, year = {2025}, }
• (2025). FedLDR: Federated optimization with label distribution-aware representations. NEUCOM, 650, 130802. (www)

  @article{ , abstract = {Federated optimization with label distribution-aware representations enhances the ability to capture and represent local label distribution variations, allowing for better model generalization across heterogeneous data. Traditionally, federated optimization with label representations allows collaborative model training across multiple clients without violating privacy, but it faces challenges in scenarios with skewed label distributions where clients have diverse local data distributions. This label distribution skew fails to capture the heterogeneity of local distributions, resulting in poor global model performance. To address these challenges, we propose federated optimization with label distribution-aware representations (FedLDR), which captures local label distributions and integrates them with the global model through a weighted approach during optimization. FedLDR enhances the ability to represent local label distribution variations, allowing for better model generalization across heterogeneous data. We evaluate FedLDR on five image classification datasets and compare it against seven baseline methods. The experimental results show that FedLDR delivers up to 25% accuracy improvement with only a minimal increase in computational cost (minimum of 0.1%). These findings demonstrate that FedLDR is an effective solution for federated learning environments with highly skewed label distributions. The FedLDR code is available at https://github.com/CST-FederatedLearning/FedLDR .}, archive = {J_NEUCOM}, author = {Suxia Zhu and Yu Yang and Guanglu Sun}, doi = {10.1016/j.neucom.2025.130802}, journal = {Neurocomputing}, month = {10}, pages = {130802}, shortjournal = {Neurocomputing}, title = {FedLDR: Federated optimization with label distribution-aware representations}, volume = {650}, year = {2025}, }
• (2025). GeoProspect: A domain-specific geological large language model with enhanced continual learning. NEUCOM, 650, 130801. (www)

  @article{ , abstract = {Developing domain-specific large language models (LLMs) in highly specialized fields such as geology presents three challenges: the discrepancy between general LLMs’ pretrained knowledge and the specialized knowledge structures of geology, the lack of domain-specific corpora, and the risk of catastrophic forgetting during continual learning. To tackle these challenges, we introduce GeoProspect, the first LLM designed specifically for the geological domain. Our methodology includes the creation of a high-quality geological corpus and the adaptation of the general-purpose Qwen2-7B model through continued pre-training and supervised instruction fine-tuning. We present two variants, GeoProspect-LoRA, a model fine-tuned using low-rank adaptation (LoRA) as an initial phase, and GeoProspect, which alleviates catastrophic forgetting during continual learning via a novel task-driven grouped mixture of LoRA experts (TG-MoLE) architecture. TG-MoLE integrates a dual-layer grouped routing strategy and a task-driven cross-group routing adjustment mechanism to optimize weight distribution and expert collaboration. To assess model performance, we develop GeoEval, the first benchmark tailored to geological LLMs. Experimental results demonstrate that our model significantly improves performance in geological knowledge understanding and mineral exploration question answering. Moreover, GeoProspect, equipped with TG-MoLE, maintains general task performance while achieving superior results in geology-specific applications.}, archive = {J_NEUCOM}, author = {Gang Wu and Haitao Wang and Kunyan Zhang and Shengguang Zhu and Jian Cui and Mengzhe Fan and Hailiang Wang and Hengliang Guo and Gubin Zhang and Dujuan Zhang and Haitao Wei and Shan Zhao}, doi = {10.1016/j.neucom.2025.130801}, journal = {Neurocomputing}, month = {10}, pages = {130801}, shortjournal = {Neurocomputing}, title = {GeoProspect: A domain-specific geological large language model with enhanced continual learning}, volume = {650}, year = {2025}, }
• (2025). Deep learning-based lane detection for intelligent driving: A comprehensive survey of methods, datasets, challenges and outlooks. NEUCOM, 650, 130795. (www)

  @article{ , abstract = {Lane detection plays a pivotal role in the environment perception system for autonomous vehicles, providing critical road information and aiding in vehicle localization. However, lane detection faces numerous challenges, including obstacle occlusion, external lighting conditions, and road structure variations. Recently, deep learning technologies with outstanding feature acquisitions and expression capabilities have achieved significant progress in environment perception for autonomous vehicles. This paper presents a comprehensive survey of deep learning-based lane detection methods. First, we review these methods from both 2D and 3D perspectives, and further categorize them into different categories, while introducing the representative works of each category. Furthermore, we summarize the datasets and evaluation metrics that are widely used for lane detection, along with a comparative analysis of the detection performance of different methods. Finally, we summarize the review works and offer a discussion on the existing challenges and future directions of deep learning-based lane detection.}, archive = {J_NEUCOM}, author = {Xin Luo and Yingping Huang and Jianyuan Cui and Kui Zheng}, doi = {10.1016/j.neucom.2025.130795}, journal = {Neurocomputing}, month = {10}, pages = {130795}, shortjournal = {Neurocomputing}, title = {Deep learning-based lane detection for intelligent driving: A comprehensive survey of methods, datasets, challenges and outlooks}, volume = {650}, year = {2025}, }
• (2025). Gradient surgery based on convolutional filters grouping for multi-task models in panoptic driving perception. NEUCOM, 650, 130792. (www)

  @article{ , abstract = {Multi-task learning has been widely applied in panoptic driving perception, where cross-task knowledge sharing can enhance the performance of multiple tasks. Typically, a multi-task learning model in panoptic driving perception consists of a shared encoder and several independent decoders, both based on CNNs. The optimization of the shared encoder by different tasks involves conflicting gradients, resulting in the overall optimization direction not being jointly optimal. It can even result in performance degradation on certain tasks, causing negative knowledge transfer. To overcome this limitation, we introduce “gradient surgery”, also known as Project Conflicting Gradients (PCGrad). However, different convolutional filters contribute to varying degrees in feature aggregation for each task. To align PCGrad for different convolutional filters that aggregate similar features, it is necessary to group the convolutional filters properly. Finding an appropriate granularity for grouping convolutional filters to perform PCGrad presents a challenge. Therefore, in this paper, we first propose an adaptive method for grouping convolutional filters based on gradient similarity. Following the grouping method, we introduce P CGrad based on C onvolutional F ilter G roups ( PCFG ) to better eliminate conflicting gradients across features in different tasks, enabling more effective knowledge sharing. In addition, we abstract PCFG as a general gradient update module and implement it in three open-source models in the field of panoptic driving perception. Extensive experiments conducted on the BDD100K dataset have demonstrated the effectiveness and versatility of our approach. By integrating PCFG, the performance of all models has been improved.}, archive = {J_NEUCOM}, author = {Rui Pan and Quan Yuan and Guiyang Luo and Bo Chen and Yilin Liu and Jinglin Li}, doi = {10.1016/j.neucom.2025.130792}, journal = {Neurocomputing}, month = {10}, pages = {130792}, shortjournal = {Neurocomputing}, title = {Gradient surgery based on convolutional filters grouping for multi-task models in panoptic driving perception}, volume = {650}, year = {2025}, }
• (2025). Component-coordinated and uncertainty-enhanced LoRA for few-shot source-free domain adaptive object detection. NEUCOM, 650, 130787. (www)

  @article{ , abstract = {We propose Component-Coordinate and Uncertainty-Enhanced LoRA (CULoRA), an innovative method for Few-shot Source-Free Domain Adaptive Object Detection, specifically designed for situations where only a limited number of unlabeled target samples are available and no access to source samples is granted. CULoRA operates within a teacher–student framework, utilizing Component Coordination (CoCo) to constrain its adapter parameters and Uncertainty Enhancement (UNE) to boost the learning process. These mechanisms ensure that the adapter maintains structural similarity with the pre-trained parameters without imposing overly restrictive constraints. CoCo selectively targets domain-invariant components for constraint, while UNE works to increase input diversity and bolster the model’s robustness. A series of experiments were carried out, and the results demonstrated that our method outperformed other methods on 6 representative benchmarks. In addition, we conducted ablation experiments to illustrate the effectiveness of each individual component of our proposed CULoRA. Overall, our proposed CULoRA provides a highly practical and effective solution for tackling the challenging Few-shot Domain Adaptive Source-Free Object Detection task, making it a significant contribution to the field.}, archive = {J_NEUCOM}, author = {Xudong Yao and Hao Liu and Xiaoshan Yang}, doi = {10.1016/j.neucom.2025.130787}, journal = {Neurocomputing}, month = {10}, pages = {130787}, shortjournal = {Neurocomputing}, title = {Component-coordinated and uncertainty-enhanced LoRA for few-shot source-free domain adaptive object detection}, volume = {650}, year = {2025}, }
• (2025). SFAG-DeepLabv3+: An automatic segmentation approach for coronary angiography images. NEUCOM, 650, 130781. (www)

  @article{ , abstract = {Automated segmentation of coronary angiography images is highly significant for computer-aided diagnosis of coronary heart disease. However, existing segmentation methods suffer from the problem of poor segmentation results caused by insufficient extraction and fusion of the features of the complex topological structure of blood vessels. In view of this, this paper proposes an automated segmentation method for coronary angiography images based on SFAG-DeepLabv3+. This method utilizes the Swin Transformer network to screen coronary angiography images and proposes a Filtering Smoothing Equalization (FSE) image enhancement method to improve the quality of angiography images. Furthermore, this paper proposes an improved automatic segmentation network for coronary arteries based on the DeepLabv3+. In the encoder section, an Adaptive hybrid Dilated convolution and double Pooling (ADP) module is proposed to enhance the ability to extract topological features of coronary blood vessels. Between the encoder and decoder, a Gaussian Context Spatial Fusion (GCSF) module is proposed to reduce information loss during the compression and decompression of information from the encoder to the decoder. In the decoder section, bicubic interpolation upsampling is employed to improve the continuity of the segmented blood vessel topology. To validate the effectiveness of the proposed method, experiments were conducted using both the ARCADE public dataset and a self-constructed CSH dataset. Experimental results demonstrate that the method proposed in this paper can perform effective feature extraction, fusion and correction on coronary angiography images, achieving average Dice coefficients of 0.9249 on the CSH dataset and 0.9156 on the ARCADE dataset.}, archive = {J_NEUCOM}, author = {Yinsheng Chen and Ying Zhang and Miaomiao Jiang and Jiahao Li and Xu Han and Kun Sun and Fan Wang and Jinwei Tian and Bo Yu}, doi = {10.1016/j.neucom.2025.130781}, journal = {Neurocomputing}, month = {10}, pages = {130781}, shortjournal = {Neurocomputing}, title = {SFAG-DeepLabv3+: An automatic segmentation approach for coronary angiography images}, volume = {650}, year = {2025}, }
• (2025). ISGM-fus: Internal structure-guided model for multispectral and hyperspectral image fusion. NEUCOM, 650, 130777. (www)

  @article{ , abstract = {Fusing a high-resolution multispectral image (HR-MSI) and a low-resolution hyperspectral image (LR-HSI) is an important way to reconstruct a high-spatial-resolution hyperspectral image (HR-HSI). Most of the deep learning-based and model-based fusion methods have achieved remarkable success in recent years. However, the former usually requires massive training samples, which is restricted because of the limited training samples. The latter usually shows limited performance due to the absence of a training phase. To solve this problem, in this paper, we propose an internal structure-guided model for HR-MSI and LR-HSI fusion (ISGM-Fus) to increase the fusion accuracy without requiring additional training samples. Firstly, we design two spatial internal structure autoregressive models to capture the internal spatial information from both the LR-HSI and HR-MSI, including a local autoregressive model and a non-local autoregressive model. Secondly, we design a split-band low-rank prior to capture the internal spectral information from the LR-HSI. Finally, we apply the spatial and spectral information learned to design a new fusion model to increase the spatial resolution of HSI. The experimental results on three sets of HSIs show the superiority of ISGM-Fus to some other traditional and advanced fusion methods. The source code is available at https://github.com/Q-J-M/ISSLM-Fus .}, archive = {J_NEUCOM}, author = {Cong Liu and Jinming Qian and Faming Fang}, doi = {10.1016/j.neucom.2025.130777}, journal = {Neurocomputing}, month = {10}, pages = {130777}, shortjournal = {Neurocomputing}, title = {ISGM-fus: Internal structure-guided model for multispectral and hyperspectral image fusion}, volume = {650}, year = {2025}, }
• (2025). Fdmclust: Functional data model-based clustering using approximation of probability density for a random function in a reproducing kernel hilbert space framework. NEUCOM, 650, 130768. (www)

  @article{ , abstract = {This paper proposes a new probability density approximation for functional random variables in the reproducing kernel Hilbert space (RKHS). Based on this approximation, a novel method for model-based clustering of functional data named Fdmclust is introduced. The previous study was based on the Gaussian assumption of the functional data due to the independence used for the covariance kernel. This assumption is not necessarily valid for general cases. The proposed method is applicable to both Gaussian and non-Gaussian functional data, utilizing the projection of functional data onto the Mercer kernel rather than the covariance kernel. To this end, PCA and ICA are used on the obtained projections for Gaussian and non-Gaussian functional random variables, respectively. The estimation of parameters is based on the EM-algorithm. The Fdmclust approach is evaluated on several simulated and real datasets, and the results confirm its efficiency.}, archive = {J_NEUCOM}, author = {Hanieh Saeidi and Mina Aminghafari and Mehdi Ashkartizabi}, doi = {10.1016/j.neucom.2025.130768}, journal = {Neurocomputing}, month = {10}, pages = {130768}, shortjournal = {Neurocomputing}, title = {Fdmclust: Functional data model-based clustering using approximation of probability density for a random function in a reproducing kernel hilbert space framework}, volume = {650}, year = {2025}, }
• (2025). Attention-enhanced re-activation CAMs for weakly supervised semantic segmentation. NEUCOM, 650, 130762. (www)

  @article{ , abstract = {Fully supervised semantic segmentation relies on pixel-level annotations, resulting in high labeling costs, whereas weakly supervised semantic segmentation (WSSS) lowers costs by leveraging image-level labels. However, WSSS faces key challenges: (1) Existing multi-class token mechanisms struggle to adequately capture and differentiate weakly activated object categories, leading to under-representation and increased misclassification risk in multi-object scenes; (2) Attention mechanism limitations, global average pooling loses local details, and self-attention averaging introduces noise, reducing accuracy; (3) Suboptimal pseudo-label quality, without pixel-level supervision, pseudo-labels lack precision in capturing both global structure and fine details. To address these challenges, this paper proposes a WSSS method leveraging Vision Transformer (ViT) and class-specific attention enhancement. First, a weak-activation-aware class-specific token mechanism is introduced, which assigns dedicated tokens to categories with weak activations, enhancing their representation and improving feature learning for under-represented classes in multi-object scenarios. Next, a class-specific attention enhancement module dynamically adjusts attention distribution, reducing misclassification in complex backgrounds. Furthermore, class activation maps are optimized using cross attention maps and block attention maps, improving pseudo-label quality and reducing missing targets. Experimental results on PASCAL VOC and MS COCO show that the proposed method enhances target localization, reduces category interference, and generates higher-quality pseudo-labels, leading to better overall segmentation performance. Source code is available at: https://github.com/LE-Li14/multi .}, archive = {J_NEUCOM}, author = {Longsheng Wei and Tangqiang Li and Fan Xu}, doi = {10.1016/j.neucom.2025.130762}, journal = {Neurocomputing}, month = {10}, pages = {130762}, shortjournal = {Neurocomputing}, title = {Attention-enhanced re-activation CAMs for weakly supervised semantic segmentation}, volume = {650}, year = {2025}, }
• (2025). Bidirectional segmentation-aware network for one-shot object detection. NEUCOM, 650, 130761. (www)

  @article{ , abstract = {One-shot object detection (OSOD) has recently emerged as a prominent research focus, owing to its capability to detect novel-class objects in target images using only a single query image as guidance. However, despite significant advances in existing query-guided frameworks, these models often struggle with identifying small objects and distinguishing objects from distractors due to the progressive loss of fine-grained details. Additionally, the semantic gap between query-target image pairs in cluttered scenes leads to missed detections stemming from insufficient contextual understanding. To address these limitations, we propose a novel unified framework, Bidirectional Segmentation-Aware Network (BSANet), which explores the reciprocal relationship between segmentation and detection tasks. The integration of the segmentation model, which excels at capturing fine-grained details, effectively complements and enhances detection performance. Specifically, we design an Instance-Guided Segmentation (IGS) module that performs segmentation guided by instance-level priors, aiming to introduce instance awareness to the pixel level. Furthermore, we develop a Segmentation-Aware Prediction (SAP) module to generate a segmentation-aware prototype enriched with fine-grained and homogeneous information, which is then matched with proposal features to facilitate instance-level score prediction in turn. Information from the two branches is bidirectionally fused, enhancing contextual understanding and sensitivity to fine-grained details. Extensive experiments demonstrate that our method achieves state-of-the-art performance on PASCAL VOC and COCO datasets in diverse evaluation settings.}, archive = {J_NEUCOM}, author = {Wenwen Zhang and Zhenghua Chen and Yongyi Su and Zhiyu Xiang and Hangguan Shan and Eryun Liu}, doi = {10.1016/j.neucom.2025.130761}, journal = {Neurocomputing}, month = {10}, pages = {130761}, shortjournal = {Neurocomputing}, title = {Bidirectional segmentation-aware network for one-shot object detection}, volume = {650}, year = {2025}, }
• (2025). Generalized variational autoencoders for learning disentangled representation. NEUCOM, 650, 130752. (www)

  @article{ , abstract = {The major goal of disentangled representation learning is to form representation space, which independently captures the underlying sources of variation, responsible for generating the data. A pioneering approach is suggested under the group of methods based on Autoencoders (AE), such as, Variational Autoencoders (VAE), β -Variational Autoencoders ( β -VAE), σ -VAE, Control-VAE, Dynamic-VAE and Learnable-VAE (L-VAE). These methods incorporate a disentanglement term, mostly expressed as a Kullback–Leibler Divergence and several hyperparameters and regularization terms in the loss function. These methods assume an equal disentanglement degree for the sources in different dimensions of the representation by using an empirically adjusted and fixed β parameter ( β = 0 or ≥ 1 ) across all dimensions. However, given the unobservable nature of the data-generating process and potential entanglements of different sources, we expect that distinct dimensions of the learned representation should exhibit varying degrees of disentanglement. In this study, we generalize the variational autoencoders and its variants by introducing a set of flexible weight functions and regularization terms for different dimensions. This generalization enables us to disentangle each latent dimension by learning the weight function of each dimension, independently. We also propose a special case of generalized VAE, called Multidimensional Learnable Variational Autoencoders, (mdL-VAE), which provide a better disentanglement-reconstruction trade-off without empirically tuning the hyperparameters of the loss function. The learned weight functions of mdL-VAE provide us useful insights about the degree of entanglement among the underlying factors of variation.}, archive = {J_NEUCOM}, author = {Hazal Mogultay Ozcan and Sinan Kalkan and Fatos T. Yarman Vural}, doi = {10.1016/j.neucom.2025.130752}, journal = {Neurocomputing}, month = {10}, pages = {130752}, shortjournal = {Neurocomputing}, title = {Generalized variational autoencoders for learning disentangled representation}, volume = {650}, year = {2025}, }
• (2025). Enhancing model explainability through S-GuISE: A spectral clustering-guided input sampling scheme for explanation. NEUCOM, 650, 130750. (www)

  @article{ , abstract = {Neural networks are extensively utilized in image-based applications due to their ability to capture significant features, learn patterns in pixel values, and efficiently process large volumes of information to yield highly accurate predictions. However, a predominant challenge arises when deep learning-based neural networks are deployed for vast amounts of data. Neural networks automatically identify relevant features from the data, but this process makes it difficult to determine which specific features contribute to the predictions of the model, presenting a challenge to interpretability. To address this, the concept of explainable artificial intelligence (XAI) has been adopted. However, traditional and state-of-the-art perturbation-based XAI methods tend to unintentionally highlight certain insignificant pixels. In this paper, we address this problem by designing a novel method, called Spectral Clustering-Guided Input Sampling for Explanations (S-GuISE), which effectively overcomes this issue and takes less computational time. To analyze the saliency map of the new method and demonstrate how it surpasses existing methods, we utilize evaluation metrics such as insertion and deletion. However, these metrics are computationally expensive and less reliable for superpixel-based saliency maps, which makes them primarily suited to evaluate pixel-based saliency. In response, we propose novel evaluation strategies for XAI, a largely unexplored area of research. We introduce two novel evaluation metrics, namely Dynamic Threshold-based Insertion (DTI) and Dynamic Threshold-based Deletion (DTD), that address the shortcomings of the existing metrics.}, archive = {J_NEUCOM}, author = {Arju Bano and Monidipa Das}, doi = {10.1016/j.neucom.2025.130750}, journal = {Neurocomputing}, month = {10}, pages = {130750}, shortjournal = {Neurocomputing}, title = {Enhancing model explainability through S-GuISE: A spectral clustering-guided input sampling scheme for explanation}, volume = {650}, year = {2025}, }
• (2025). Adapt & align: Continual learning with generative models’ latent space alignment. NEUCOM, 650, 130748. (www)

  @article{ , abstract = {Motivation: Neural networks suffer from abrupt loss in performance when retrained with additional data from different distributions. At the same time, training with additional data without access to the previous examples rarely improves the model’s performance. Methods: We propose Adapt & Align, a novel continual learning framework that leverages generative models to align their latent representations across tasks. The approach is divided into two phases: • Local Training: Train a generative model (e.g., a Variational Autoencoder (VAE) or a Generative Adversarial Network (GAN)) on the current task to capture task-specific features. • Global Training: Use a translator network to map these task-specific latent representations into a unified global latent space, thereby facilitating both forward and backward knowledge transfer. Results: Experiments on benchmark datasets (e.g., MNIST, Omniglot, CIFAR, CelebA) as well as real-world application for particle simulation at CERN demonstrate that Adapt & Align mitigates catastrophic forgetting and improves generation quality as indicated by metrics such as Fréchet Inception Distance (FID), distribution precision and recall, or accuracy for the downstream classification task. Ablation studies confirm the critical role of each component.}, archive = {J_NEUCOM}, author = {Kamil Deja and Bartosz Cywiński and Jan Rybarczyk and Tomasz Trzciński}, doi = {10.1016/j.neucom.2025.130748}, journal = {Neurocomputing}, month = {10}, pages = {130748}, shortjournal = {Neurocomputing}, title = {Adapt & align: Continual learning with generative models’ latent space alignment}, volume = {650}, year = {2025}, }
• (2025). Long-tailed learning for generalized category discovery. NEUCOM, 650, 130745. (www)

  @article{ , abstract = {Generalized Category Discovery (GCD) utilizes labeled samples of known classes to discover novel classes in unlabeled samples. Existing methods show effective performance on artificial datasets with balanced distributions. However, real-world datasets are always imbalanced, significantly affecting the effectiveness of these methods. To solve this problem, we propose a novel framework that performs generalized category discovery in long-tailed distributions. We first present a self-guided labeling technique that uses a learnable distribution to generate pseudo-labels, resulting in less biased classifiers. We then introduce a representation balancing process to derive discriminative representations. By mining sample neighborhoods, this process encourages the model to focus more on tail classes. We conduct experiments on public datasets to demonstrate the effectiveness of the proposed framework. The results show that our model exceeds previous state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Cuong Manh Hoang}, doi = {10.1016/j.neucom.2025.130745}, journal = {Neurocomputing}, month = {10}, pages = {130745}, shortjournal = {Neurocomputing}, title = {Long-tailed learning for generalized category discovery}, volume = {650}, year = {2025}, }
• (2025). FtdBLIP: Fusion-to-dual distillation for superior image–text retrieval of dual-encoder architecture. NEUCOM, 650, 130744. (www)

  @article{ , abstract = {The end-to-end visual-language pre-training model with dual-encoder architectures is suitable for various time-sensitive downstream tasks, especially for image–text retrieval, owing to its rapid and efficient modal fusion capabilities that offer low time complexity. However, the exclusive use of feature dot multiplication for modality fusion in the dual-encoder architecture may be less effective compared to fusion-encoder architectures like BLIP, which leverage cross-attention mechanisms for feature integration. Despite the widespread application of knowledge distillation, transferring knowledge between models with substantial architectural disparities presents a significant challenge. To address this issue, we employ the benchmark model BLIP, which encompasses both Fusion-encoder and Dual-encoder architectures. Subsequently, we introduce a fusion-to-dual distillation (FtdBLIP) method aimed at transferring the superior feature fusion capabilities of the fusion-encoder (teacher) to the dual-encoder (student). Based on the characteristics of BLIP, we construct two cross-architecture distillation methods: (1) We feed the image and text embeddings from the dual-encoder into the fusion encoder. The last layer of cross attention generates a similarity map to optimize the dot product map between the original image and text embeddings. (2) We utilize the fusion encoder’s output rank scores to guide the similarity scores of visual and textual features in the dot product of the dual encoder. Experimental results demonstrate that FtdBLIP effectively improves the cross-modal retrieval performance of the BLIP dual encoder architecture on the Flickr30K and MSCOCO benchmark datasets. Furthermore, compared with existing vision-language pre-trained models, FtdBLIP demonstrates superior text-to-image retrieval performance.}, archive = {J_NEUCOM}, author = {Ziyan Gong and Tianhan Gao}, doi = {10.1016/j.neucom.2025.130744}, journal = {Neurocomputing}, month = {10}, pages = {130744}, shortjournal = {Neurocomputing}, title = {FtdBLIP: Fusion-to-dual distillation for superior image–text retrieval of dual-encoder architecture}, volume = {650}, year = {2025}, }
• (2025). Multi-granular embedding optimization with spatial-channel adaptive tuning for perceptual image quality assessment. NEUCOM, 650, 130731. (www)

  @article{ , abstract = {Fine-tuning strategies such as adapter-tuning and prompt-tuning have demonstrated remarkable effectiveness in optimizing Vision Transformers (ViTs) for various downstream tasks. Inspired by these advancements, we present a multi-granular embedding optimization framework designed for perceptual image quality assessment (IQA). This method refines image quality representations by integrating learnable convolutional filter banks to adapt both the spatial granularity and channel dimensionality within the ViT embedding layers. We investigate spatial-tuning and channel-tuning strategies of the embeddings, examining their synergistic effects on quality representation capabilities across diverse distortion scenarios. To efficiently learn the image embedding representation, we employ a k-nearest neighbors-based evolutionary algorithm (EA) to optimize the tuning parameters, allowing the model to capture fine-grained local details and contextual information across spatial domains while learning enhanced representations within feature subspaces. Experimental results demonstrate that the proposed tuning method enhances quality evaluation performance, achieving state-of-the-art results across a diverse set of IQA application scenarios.}, archive = {J_NEUCOM}, author = {Xiaoqi Wang and Yun Zhang and Junqi Liu}, doi = {10.1016/j.neucom.2025.130731}, journal = {Neurocomputing}, month = {10}, pages = {130731}, shortjournal = {Neurocomputing}, title = {Multi-granular embedding optimization with spatial-channel adaptive tuning for perceptual image quality assessment}, volume = {650}, year = {2025}, }
• (2025). Adversarial defenses via vector quantization. NEUCOM, 650, 130703. (www)

  @article{ , abstract = {Adversarial attacks pose significant challenges to the robustness of modern deep neural networks in computer vision, and defending these networks against adversarial attacks has attracted intense research efforts. Among various defense strategies, preprocessing-based defenses are practically appealing since there is no need to train the network under protection. However, such approaches typically do not achieve comparable robustness as other methods such as adversarial training. In this paper, we propose a novel framework for preprocessing-based defenses, where a vector quantizer is used as a preprocessor. This framework, inspired by and extended from Randomized Discretization (“RandDisc”), is theoretically principled by rate–distortion theory: indeed, RandDisc may be viewed as a scalar quantizer, and rate–distortion theory suggests that such quantization schemes are inferior to vector quantization. In our framework, the preprocessing vector quantizer treats the input image as a collection of patches and finds a set of “representative patches” based on the patch distributions; each original patch is then modified according to the representative patches close to it. We present two lightweight defenses in this framework, referred to as patched RandDisc (pRD) and sliding-window RandDisc (swRD), where the patches are disjoint in the former and overlapping in the latter. We show that vector-quantization-based defenses have certifiable robust accuracy and that pRD and swRD demonstrate state-of-the-art performance, surpassing RandDisc by a large margin. Notably, the proposed defenses possess the “obfuscated gradients” property. Our experiments however show that pRD and swRD remain effective under the STE and EOT attacks, which are designed specifically for defenses with gradient obfuscation. These observations confirm that the power of our defenses lies in the effective exploitation of image structure via vector quantization.}, archive = {J_NEUCOM}, author = {Zhiyi Dong and Yongyi Mao}, doi = {10.1016/j.neucom.2025.130703}, journal = {Neurocomputing}, month = {10}, pages = {130703}, shortjournal = {Neurocomputing}, title = {Adversarial defenses via vector quantization}, volume = {650}, year = {2025}, }
• (2025). PaddingFlow: Improving normalizing flows with padding-dimensional noise. NEUCOM, 650, 130651. (www)

  @article{ , abstract = {Normalizing flow is a generative modeling approach with efficient sampling. However, Flow-based models suffer two issues: (1) If the target distribution is manifold, due to the mismatch between the dimensions of the latent target distribution and the data distribution, flow-based models might perform badly. (2) Discrete data might make flow-based models collapse into a degenerate mixture of point masses. To sidestep such two issues, we propose PaddingFlow, a novel dequantization method, which improves normalizing flows with padding-dimensional noise. To implement PaddingFlow, only the dimension of normalizing flows needs to be modified. Thus, our method is easy to implement and computationally cheap. Moreover, the padding-dimensional noise is only added to the padding dimension, which means that PaddingFlow can dequantize without changing data distributions. After implementing existing dequantization methods, the original data distributions are changed, which might degrade performance. We validate our method on the main benchmarks of unconditional density estimation, including five tabular datasets and four image datasets for Variational Autoencoder (VAE) models, and the Inverse Kinematics (IK) experiments which are conditional density estimation. The results show that PaddingFlow can perform better in all experiments in this paper, which means PaddingFlow is widely suitable for various tasks. The code is available at: https://github.com/AdamQLMeng/PaddingFlow .}, archive = {J_NEUCOM}, author = {Qinglong Meng and Chongkun Xia and Xueqian Wang and Bin Liang}, doi = {10.1016/j.neucom.2025.130651}, journal = {Neurocomputing}, month = {10}, pages = {130651}, shortjournal = {Neurocomputing}, title = {PaddingFlow: Improving normalizing flows with padding-dimensional noise}, volume = {650}, year = {2025}, }
• (2025). Meta-learning-based delayless subband adaptive filter using complex self-attention for active noise control. NEUCOM, 650, 130637. (www)

  @article{ , abstract = {Active noise control typically employs adaptive filtering to generate secondary noise, where the least mean square algorithm is the most widely used. However, traditional updating rules are linear and exhibit limited effectiveness in addressing nonlinear environments and nonstationary noise. To tackle this challenge, we reformulate the active noise control problem as a meta-learning problem and propose a meta-learning-based delayless subband adaptive filter with deep neural networks. The core idea is to utilize a neural network as an adaptive algorithm that can adapt to different environments and types of noise. The neural network is trained under noisy observations, implying that it recognizes the optimized updating rule without true labels. A single-headed attention recurrent neural network is devised with learnable feature embedding to update the adaptive filter weight efficiently, enabling accurate computation of the secondary source to attenuate the unwanted primary noise. In order to relax the time constraint on updating the adaptive filter weights, the delayless subband architecture is employed, which will allow the system to be updated less frequently as the downsampling factor increases. In addition, the delayless subband architecture does not introduce additional time delays in active noise control systems. A skip updating strategy is introduced to decrease the updating frequency further, enabling machines with limited resources to adopt our meta-learning-based model. Extensive multi-condition training ensures generalization and robustness against various types of noise and environments. Simulation results demonstrate that the proposed model achieves lower steady-state error than competitors.}, archive = {J_NEUCOM}, author = {Pengxing Feng and Hing Cheung So}, doi = {10.1016/j.neucom.2025.130637}, journal = {Neurocomputing}, month = {10}, pages = {130637}, shortjournal = {Neurocomputing}, title = {Meta-learning-based delayless subband adaptive filter using complex self-attention for active noise control}, volume = {650}, year = {2025}, }
• (2025). In the eyes of a language model: A comprehensive examination through eye-tracking data. NEUCOM, 650, 130617. (www)

  @article{ , abstract = {Cognitive signals, particularly eye-tracking data, offer a unique lens for understanding human sentence processing. Leveraging eye-gaze data from the English and Italian section of the Multilingual Eye-Movement Corpus (MECO), we designed a series of experiments aiming at exploring whether pre-trained neural language models (NLMs) encode patterns representative of human reading behavior and if directly incorporating this information through a fine-tuning process influences the cognitive plausibility of the model. Additionally, we sought to determine if such an impact persists through a downstream task. Our findings reveal that transformers encode eye-gaze-related information during pretraining and that explicitly integrating eye-tracking features increases model alignment with human attention. When investigating the effect of intermediate fine-tuning on eye-tracking data on the model’s performance on a downstream task, we observe that this intermediate step does not result in catastrophic forgetting, despite the very different nature of the considered downstream task. In addition, the attention mechanism of models undergoing intermediate fine-tuning remains closely aligned with human attention. In conclusion, our comprehensive evaluation of NLMs informed by human attention patterns offers great potential for advancing the growing field of eXplainable Artificial Intelligence (XAI). Grounding language models in real-world cognitive processes enables the creation of systems that not only replicate human language output but also align with the cognitive mechanisms behind reading and comprehension. This alignment with human behavior enhances model adaptability, interpretability, and effectiveness, fostering more human-centric, transparent, and reliable AI applications across various domains. 1}, archive = {J_NEUCOM}, author = {Luca Dini and Luca Moroni and Dominique Brunato and Felice Dell’Orletta}, doi = {10.1016/j.neucom.2025.130617}, journal = {Neurocomputing}, month = {10}, pages = {130617}, shortjournal = {Neurocomputing}, title = {In the eyes of a language model: A comprehensive examination through eye-tracking data}, volume = {650}, year = {2025}, }
• (2025). Distillation guided deep unfolding network with frequency hierarchical regularization for low-dose CT image denoising. NEUCOM, 650, 130535. (www)

  @article{ , abstract = {Computed tomography (CT) is an important means of medical diagnosis. Low-dose CT (LDCT) avoids the problem of high incidence of disease by reducing the radiation dose, but reduces the quality of images. Deep unfolding networks are used to improve LDCT image quality because of its excellent performance and theoretical interpretability. However, the training of the network is usually based on paired LDCT-NDCT images, and the input image is processed iteratively and independently stage-by-stage. There are problems of accumulation of information distortion and lack of feature interaction between different stages. To solve these problems, we propose a distillation guided deep unfolding network with frequency hierarchical regularization (FHRDUN). It uses the pre-trained image denoising network as the teacher network, and adopts the initially clean image attained by distilling teacher network along with the LDCT images to supervise the training of student network. We also design the noise estimation module (NUNet) and frequency hierarchical regularization module (FHRM). NUNet is devised for initial noise removal and to synthesize LDCT images with the denoised images obtained by the regularization module. FHRM achieves the information exchange across stages and the regularization of frequency-division subimage. Extensive experiments on Mayo, Piglet and Kaggle datasets demonstrate superior performance of FHRDUN compared with state-of-the-art methods. Compared with existing methods, the PSNR results for four different doses are increased on average by 0.28 dB on the Piglet dataset.}, archive = {J_NEUCOM}, author = {Linlin Yang and Hongying Liu and Yuanyuan Liu and Fanhua Shang and Liang Wan and Licheng Jiao}, doi = {10.1016/j.neucom.2025.130535}, journal = {Neurocomputing}, month = {10}, pages = {130535}, shortjournal = {Neurocomputing}, title = {Distillation guided deep unfolding network with frequency hierarchical regularization for low-dose CT image denoising}, volume = {650}, year = {2025}, }
• (2025). CUR-estimator: Towards reliable missing data imputation for aero-engine degradation process. NEUCOM, 649, 130876. (www)

  @article{ , abstract = {In the process of missing data imputation for aero-engine life cycle degradation time series, two primary challenges arise. First, due to variations in the duration of different flight missions, and the fact that the same flight phase may also vary in duration across different flights, the time intervals for collecting key samples are not always consistent. This variability increases the difficulty of evaluating the impact of individual flights on overall performance changes. Second, when using neural networks for imputing missing data, issues such as significant noise or extended periods of missing data may arise, leading to unreasonable imputation results. To address the challenges, this paper proposes a Constrained Unseen Recovery Estimator (CUR-Estimator) for imputing missing data in aero-engine life cycle degradation datasets. Firstly, the time interval information is encoded via a transformer-enhanced gate recurrent unit. The results are then combined with missing masks to adjust the hidden states and input weights for the missing segments, forming the Interval-Aware Temporal Imputation Network. Secondly, this paper uses statistical interpolation methods to constrain the imputation results of the neural network, limiting the range of imputation outcomes and thereby reducing the possibility of unreasonable outputs. As an example, the Piecewise Cubic Hermite Interpolating Polynomial is applied to constrain the Interval-Aware Temporal Imputation Network in handling time interval information. Finally, experiments were conducted using a simulation dataset and a real civil aero-engine dataset, which showed that the proposed method has high accuracy and strong stability.}, archive = {J_NEUCOM}, author = {Haoze Wu and Shisheng Zhong and Minghang Zhao and Xuyun Fu and Yongjian Zhang and Song Fu}, doi = {10.1016/j.neucom.2025.130876}, journal = {Neurocomputing}, month = {10}, pages = {130876}, shortjournal = {Neurocomputing}, title = {CUR-estimator: Towards reliable missing data imputation for aero-engine degradation process}, volume = {649}, year = {2025}, }
• (2025). MVC-former adaptation: A multi-view convolution transformer-based domain adaptation framework for cross-subject motor imagery classification. NEUCOM, 649, 130875. (www)

  @article{ , abstract = {Owing to individual difference, it is challenging to decode the target subject’s mental intentions by applying existing models in cross-subject Brain-Computer Interface tasks. The transfer learning methods have shown promising performance in this field, but they still suffer from poor presentations on temporal correlations characteristics of EEG signals. This paper proposes a multi-view convolution-transformer based domain adaptation framework for cross-subject motor imagery classification tasks. Firstly, to exploit the frequency diversity of EEG signals, we decomposed EEG signals into several overlapping frequency views and extracted frequency-related spatial and temporal features by parallel spatiotemporal convolution block. Subsequently, we use transformer blocks to extract long-range dependencies and narrow the marginal distribution between source and target domains. Eventually, a classifier and a domain discriminator were used for domain adaptation, and a mixed loss was employed to align conditional distributions. We conducted model validation on the BCI Competition IV 2a and 2b datasets and achieved average accuracies of 77.8 % and 80.1 %, respectively. The experimental results show that our proposed framework outperforms traditional deep adversarial domain adaptive methods.}, archive = {J_NEUCOM}, author = {Yining Liang and Ming Meng and Yunyuan Gao and Xugang Xi}, doi = {10.1016/j.neucom.2025.130875}, journal = {Neurocomputing}, month = {10}, pages = {130875}, shortjournal = {Neurocomputing}, title = {MVC-former adaptation: A multi-view convolution transformer-based domain adaptation framework for cross-subject motor imagery classification}, volume = {649}, year = {2025}, }
• (2025). Comparison of distance and reinforcement-learning rules in social-influence models. NEUCOM, 649, 130870. (www)

  @article{ , abstract = {Values are essential for decision-making in people and machines. When a decision is to be made, relevant information is obtained and then, the course of action that maximizes expected value is selected. For people, values have socio-cultural components, with individuals learning from each other. This learning leads the socio-cultural organization of values that includes grouping and polarization. To model this organization, sociologists and social psychologists use Agent-Based Social-Influence models. In standard form, these models use distance-based rules, that is, the degree by which a person influences another is a function of the distance between their values. In this article, we also introduce social-influence rules based on reinforcement learning. This is the mechanism that the brain and artificial intelligence use to learn to optimize values. We report computer simulations of the dynamics of multi-agent, Social-influence models, using either distance or reinforcement-learning rules. Our results show that both types of rules account for grouping and polarization. However, reinforcement-learning rules lead to a better accounting of the number of groups, their polarization, and the degree of individuality, that is, agents not belonging to clusters. Our simulations also reveal factors that influence these results. These factors include contrarians, the number of interacting agents, and the dimensional richness of the value space. Finally, our results with reinforcement-learning rules show complex dynamics of values, including group instability, tipping points, and phase transitions in spontaneous group formation. We discuss the consequences of these results for artificial-intelligence systems learning from each other through social-influence models.}, archive = {J_NEUCOM}, author = {Norberto M. Grzywacz}, doi = {10.1016/j.neucom.2025.130870}, journal = {Neurocomputing}, month = {10}, pages = {130870}, shortjournal = {Neurocomputing}, title = {Comparison of distance and reinforcement-learning rules in social-influence models}, volume = {649}, year = {2025}, }
• (2025). Semi-supervised fault diagnosis of bearings under noisy environments with limited labeled samples. NEUCOM, 649, 130843. (www)

  @article{ , abstract = {Bearings are critical components in mechanical and engineering systems, where accurate fault diagnosis is essential for ensuring operational reliability and safety. However, industrial applications often face significant challenges, including the scarcity of labeled samples and the presence of noise interference. To address these challenges, the study presents a semi-supervised bearing fault diagnosis method based on an improved Wasserstein Generative Adversarial Network (IWGAN). Initially, continuous wavelet transform (CWT) is applied to denoise the signals and extract key features. Subsequently, a limited number of labeled samples along with a large quantity of unlabeled samples are fed into the IWGAN. The study employs a hybrid module consisting of CNN and Transformer as the foundation for the generator (G). In the architecture, the model incorporates convolutional operations to simultaneously leverage attention-based global feature capture and the local feature extraction capabilities of CNN, effectively capturing both local and global features. The discriminator (D) incorporates a dynamic sparse attention mechanism (DSAM), which adaptively enhances the key features while suppressing redundant information. By integrating semi-supervised learning strategies with a novel loss function, the proposed method achieves a synergy between supervised and unsupervised learning paradigms. Experimental validation on two representative bearing fault datasets demonstrates that the proposed approach achieves superior diagnostic accuracy and robustness against noise, even under conditions of substantial noise interference and limited labeled samples. The study introduces a novel and scalable diagnostic framework, providing an effective solution to the challenges of industrial fault diagnosis in real-world scenarios.}, archive = {J_NEUCOM}, author = {Wenxuan Zhao and Jianxiong Gao and Qin Cheng and Yiping Yuan and Jianxing Zhou}, doi = {10.1016/j.neucom.2025.130843}, journal = {Neurocomputing}, month = {10}, pages = {130843}, shortjournal = {Neurocomputing}, title = {Semi-supervised fault diagnosis of bearings under noisy environments with limited labeled samples}, volume = {649}, year = {2025}, }
• (2025). Rapid rib lesion diagnosis based on improved YOLOv8 and bytetrack algorithms enhanced multi-planar reconstruction. NEUCOM, 649, 130842. (www)

  @article{ , abstract = {Rib fractures, a prevalent traumatic injury, pose significant risks of morbidity and mortality due to complications such as pneumothorax or organ damage. Accurate and timely detection is critical for effective clinical intervention and improved patient outcomes. Traditional computed tomography based manual rib fracture detection is labor-intensive, time-consuming, and prone to errors. Recent artificial intelligence driven tools have been introduced for fracture detection, but challenges remain in balancing diagnostic accuracy with computational efficiency. This study proposes an enhanced multi-planar reconstruction method based on improved YOLOv8 and Bytetrack algorithms. The enhanced YOLOv8-based detection framework is integrated with the global attention mechanism and the convolutional block attention module, double attention mechanisms to improve rib localization accuracy for automated multiplanar reconstruction. Then, it is paired with ByteTrack, a robust multi-object tracking algorithm, to establish a continuous rib tracking pipeline. The proposed architecture enhances the detection of subtle fracture patterns in low-contrast regions, potentially reducing diagnostic oversights in emergency radiology scenarios. Clinical systematic validation demonstrates superior rib tracking precision (Precision: 0.980, Recall: 0.974, mAP50: 0.985, F1: 0.977, 11.7 % improvement in F1-score versus baseline YOLOv8) and reduces reconstruction time by nearly 70 % compared to manual workflows, addressing the labor-intensive nature of manual multiplanar reconstruction and the suboptimal fracture detection sensitivity of existing artificial intelligence solutions. This technical advancement bridges a crucial gap between computational efficiency and diagnostic reliability in musculoskeletal imaging analysis.}, archive = {J_NEUCOM}, author = {Yue Zhou and Huasheng Zhuo and Donglin Wen and Yuran Gu and Zhiyong Liu and Gang Wu and Guanglan Liao and Tielin Shi}, doi = {10.1016/j.neucom.2025.130842}, journal = {Neurocomputing}, month = {10}, pages = {130842}, shortjournal = {Neurocomputing}, title = {Rapid rib lesion diagnosis based on improved YOLOv8 and bytetrack algorithms enhanced multi-planar reconstruction}, volume = {649}, year = {2025}, }
• (2025). A multimodal traffic scene understanding model integrated with optical flow maps. NEUCOM, 649, 130838. (www)

  @article{ , abstract = {Autonomous driving technology has rapidly developed in recent years, promising wide-ranging applications in the transportation industry. However, the decision-making processes of autonomous driving systems are nontransparent and nonexplainable, hindering the broader adoption and industrial deployment of this technology. To overcome these challenges, this article proposes an end-to-end traffic scene understanding model (TSUM) integrated with various proven performance enhancement techniques derived from research on multimodal large models. First, features from Farneback optical flow maps and raw scene videos are simultaneously extracted using a shared-parameter network, which effectively captures the temporal dynamics of traffic scenes. By incorporating scene perception and causal analysis at the task layer, an adaptive task complexity progression strategy is designed, which reduces the task complexity. Finally, we establish a guiding library with a diverse array of guiding phrases to enhance the generalization capabilities of the TSUM. The model is experimentally validated on the publicly available Berkeley Deep Drive-X dataset. Comparison studies confirmed that the TSUM outperforms state-of-the-art methods across multiple evaluation metrics.}, archive = {J_NEUCOM}, author = {Yuanhang Wang and Yonghua Zhou and Min Zhong and Yiduo Mei and Hamido Fujita and Hanan Aljuaid}, doi = {10.1016/j.neucom.2025.130838}, journal = {Neurocomputing}, month = {10}, pages = {130838}, shortjournal = {Neurocomputing}, title = {A multimodal traffic scene understanding model integrated with optical flow maps}, volume = {649}, year = {2025}, }
• (2025). Multimodal fusion techniques: Review, data representation, information fusion, and application areas. NEUCOM, 649, 130827. (www)

  @article{ , abstract = {Multimodality refers to a model's ability to process and integrate information from various sources, such as visual, audio, and textual data. In real-world scenarios, data is often collected across multiple modalities, necessitating effective techniques for their fusion. The task of fusion of these features is called multimodal fusion. While multimodal learning aims to combine complementary information from multiple modalities to form a unified representation or improve task performance, cross-modal learning emphasizes the mapping, alignment, or translation between modalities—such as generating images from textual descriptions or retrieving videos based on audio cues. This survey provides a comprehensive overview of unimodal and multimodal representation learning, identifying cross-modal correlations. Models like LSTM, GRU, and BERT are highlighted for semantic feature extraction, while architectures such as VGGNet, MobileNet, InceptionNet, and ResNet are discussed for visual features, along with recent advancements in Large Language Models (LLMs) and Vision Transformers (ViTs). The early, late, and intermediate fusion techniques for presenting a combined multimodal representation are discussed, with further analysis of the joint, coordinated, and encoder-decoder representation frameworks. Various conventional and contemporary fusion models, such as Attention networks, Transformers are discussed in depth, encompassing different application domains that revolve around text and visual data. In the contemporary scenario where big data is the trend and multimodality is a concern, the various aspects of volume, varsity, and variety of multimodal Big Data and its challenges have also been discussed and addressed. Various Vision-Languge Pre-trained (VLP) models and Multimodal Large Language Models (MLLMs) such as ALIGN, Florence, ALBEF, ViLT, and VLMo are highlighted. Key application areas, benchmark datasets, and evaluation metrics are outlined to provide a holistic understanding of the field. The survey concludes by exploring pressing challenges and outlining potential directions for future research in multimodal fusion.}, archive = {J_NEUCOM}, author = {Sakshini Hangloo and Bhavna Arora}, doi = {10.1016/j.neucom.2025.130827}, journal = {Neurocomputing}, month = {10}, pages = {130827}, shortjournal = {Neurocomputing}, title = {Multimodal fusion techniques: Review, data representation, information fusion, and application areas}, volume = {649}, year = {2025}, }
• (2025). Synchronization for coupling delayed complex dynamical networks in time-square-dependent looped-functionals via a new sampled-data control approach. NEUCOM, 649, 130819. (www)

  @article{ , abstract = {This paper proposes a novel sampled-data control (SDC) approach for addressing synchronization problems in complex dynamical networks (CDNs) with coupling delays in time-square-dependent looped-functionals (TSDLFs). The proposed method introduces a sampling-time-dependent control signal to achieve synchronization while considering coupled time-varying delays and uncertainties in CDNs. To ensure asymptotic stability, new Lyapunov functionals are constructed, including TSDLFs and two-sided looped functionals, effectively utilizing information over the intervals [ t k , t ] and [ t , t k + 1 ) . Sufficient conditions are derived to guarantee the asymptotic stability of the error states. A numerical example based on Chua’s circuits demonstrates the effectiveness and superiority of the proposed method.}, archive = {J_NEUCOM}, author = {Haripriya Muralikrishnan and N. Padmaja and E. Umamaheswari and Lakshmanan Shanmugam}, doi = {10.1016/j.neucom.2025.130819}, journal = {Neurocomputing}, month = {10}, pages = {130819}, shortjournal = {Neurocomputing}, title = {Synchronization for coupling delayed complex dynamical networks in time-square-dependent looped-functionals via a new sampled-data control approach}, volume = {649}, year = {2025}, }
• (2025). A spatiotemporal dynamic fusion network for surgical action recognition. NEUCOM, 649, 130808. (www)

  @article{ , abstract = {Surgical action recognition is crucial and tough in intelligent surgical robots, enabling these systems to accurately identify and interpret the ongoing actions within a surgical procedure. By recognizing the action state in real-time, the robot can provide immediate feedback and make adjustments to ensure the precision and safety of the surgery. However, it faces several challenges, such as the temporal complexity of surgical actions, the fine operation steps and subtle changes in surgical movements. Thus, a dynamic attention mechanism has been proposed to capture the temporal correlation between the current frame and the previous frames from video sequence. Furthermore, a spatiotemporal dynamic fusion network comprising two specialized modules has been proposed. The first module, Double Bi-level Routing Attention (DBRA), is designed to extract the most pertinent spatial and temporal features. While the other module, CNN-LSTM, is dedicated to delivering comprehensive spatiotemporal information. Experiments have been conducted on both a neurosurgical dataset Neuro67 and a public dataset Suturing to demonstrate the performance of the proposed method. The results indicate that the proposed method has achieved superior performance on the hard issue, achieving AP on Neuro67 and ACC on Suturing of 76.9% and 85.5%, leading by 6.7% and 1.2% respectively, with the modules effectively focusing on dependency relationships both within regions of a frame and across frames in video sequence.}, archive = {J_NEUCOM}, author = {Gui-Bin Bian and Yaqin Peng and Zhen Li and Qiang Ye and Ruichen Ma}, doi = {10.1016/j.neucom.2025.130808}, journal = {Neurocomputing}, month = {10}, pages = {130808}, shortjournal = {Neurocomputing}, title = {A spatiotemporal dynamic fusion network for surgical action recognition}, volume = {649}, year = {2025}, }
• (2025). SMFSwap: Student-aware multi-teacher knowledge distillation for fast face-swapping. NEUCOM, 649, 130807. (www)

  @article{ , abstract = {Recent face swapping frameworks have performed admirably in generating high-fidelity results. However, most of these methods require huge memory footprints and computational resources, making them difficult to apply in real-time applications. To overcome this problem, we propose a student-aware multi-teacher knowledge distillation for fast face swapping, called SMFSwap. Specifically, we use a compressed version of the existing network Simple Swap as the fast swap network ( FSN ) and introduce multi-teacher knowledge distillation to stabilize the training of FSN . In addition, we propose a novel multi-teacher weight assignment strategy for the face-swapping task without ground truth. It adaptively assigns different weights for each teacher’s result with the help of the matching degree between teacher and FSN results. During the training, teacher results that are close to the FSN result are assigned large weights to better guide the FSN training. Moreover, we also integrate teacher intermediate layer features to stabilize the knowledge transfer process. Extensive quantitative and qualitative experiments on various datasets demonstrate that the proposed SMFSwap achieves comparable results to the teacher network and other state-of-the-art face-swapping methods while having a lower model complexity. The code will be available at https://github.com/EifelTing/SMFSwap .}, archive = {J_NEUCOM}, author = {Yifeng Ding and Gaoming Yang and Shuting Yin and Ji Zhang and Xianjin Fang and Wencheng Yang}, doi = {10.1016/j.neucom.2025.130807}, journal = {Neurocomputing}, month = {10}, pages = {130807}, shortjournal = {Neurocomputing}, title = {SMFSwap: Student-aware multi-teacher knowledge distillation for fast face-swapping}, volume = {649}, year = {2025}, }
• (2025). A distribution feature extracting network with dual correlation for long sequence time-series forecasting. NEUCOM, 649, 130806. (www)

  @article{ , abstract = {Long-sequence time-series forecasting (LSTF) has broad real-world applications, and numerous methods have driven significant advancements in this field. However, challenges remain, such as dealing with distribution shifts and capturing the features from both global and local perspectives, such as trends and seasonal changes in time series. To address these issues, we propose an Auto/Cross-correlation Distribution Feature Extraction Network (ACDN), a linear complexity model that integrates two key modules and a mechanism: the Distribution Processing Module normalizes the input sequence and dynamically predicts the distribution of the forecasted sequence to capture the features of distribution shifts. The Auto/Cross-Correlation Module captures evolving trend components by computing both the autocorrelation of individual time series segments and the cross-correlation between different segments. The Subtle Feature Preservation Mechanism compensates for feature loss caused by dimensionality reduction in the encoder–decoder structure, ensuring critical fine-grained patterns are retained. Extensive experiments on nine datasets from diverse domains demonstrate the effectiveness of ACDN in multivariate LSTF tasks.}, archive = {J_NEUCOM}, author = {Jin Fan and Zehao Wang and Feiwei Qin and Huifeng Wu and Danfeng Sun and Jia Wu}, doi = {10.1016/j.neucom.2025.130806}, journal = {Neurocomputing}, month = {10}, pages = {130806}, shortjournal = {Neurocomputing}, title = {A distribution feature extracting network with dual correlation for long sequence time-series forecasting}, volume = {649}, year = {2025}, }
• (2025). Frequency-domain keyframe interpolation denoising for text and image-guided video editing acceleration with diffusion models. NEUCOM, 649, 130800. (www)

  @article{ , abstract = {In the field of video editing, diffusion models have become a dominant generative framework. Within this framework, the denoising process plays a crucial role in enhancing the visual quality of the final output. However, the computationally intensive nature of their traditional spatial-domain denoising process poses a significant bottleneck, especially when handling high-resolution, long-duration videos, leading to severely slow processing speeds. To address this crucial limitation and accelerate diffusion-based video generation and editing, we propose a novel Frequency-Domain Keyframe Interpolation Denoising ( FKID ) technique. Our approach drastically reduces the computational load by optimizing the denoising process within the diffusion framework. Instead of denoising every frame, we dynamically select keyframes based on their semantic relevance between video frames. This not only reduces the computational load of the denoising process but also ensures a stronger focus on the video’s crucial content. We then employ a Fourier transform-based interpolation algorithm to mitigate video incoherence resulting from the keyframe mechanism. By analyzing the frequency-domain relationship and time interval between adjacent keyframes, the algorithm dynamically adjusts inter-frame frequency components. This allows for the reconstruction of intermediate frames, ultimately improving the video’s visual appeal and temporal smoothness. Extensive experiments conducted on both synthetic and real datasets demonstrate that FKID achieves significant acceleration in processing high-resolution, long-duration video denoising compared to traditional diffusion-based methods. It offers a speedup of over three times compared to conventional algorithms, while simultaneously maintaining leading performance on aesthetic metrics such as CLIP and DINO.}, archive = {J_NEUCOM}, author = {Chenhao Pang and Chen Chen and Bin Song}, doi = {10.1016/j.neucom.2025.130800}, journal = {Neurocomputing}, month = {10}, pages = {130800}, shortjournal = {Neurocomputing}, title = {Frequency-domain keyframe interpolation denoising for text and image-guided video editing acceleration with diffusion models}, volume = {649}, year = {2025}, }
• (2025). Make full use of your data: On copy-based augmentation in speech anti-spoofing. NEUCOM, 649, 130799. (www)

  @article{ , abstract = {The generalization ability of speech anti-spoofing (SAS) models is critical for their effective deployment in security-critical applications. Recently, data augmentation has gained significant attention as an effective method to enhance model generalization. However, most existing augmentation techniques are either tailored to specific test sets or overlook spoof-related characteristics, limiting their generalization. The goal of this research is to develop a novel data augmentation framework that maximizes existing data usage and enhances model performance in unseen domains without complex design. We propose a copy-based augmentation framework, inspired by the theoretical error bound of domain generalization, which mitigates the distribution discrepancy between training and testing domains by applying transformations to existing speech signals or features. Within this framework, directly concatenating existing speech signals is problematic for SAS. To address this, we propose an optimized copy-paste augmentation (CpAug) method by concatenation and substitution policies, avoiding the introduction of forgery traces and enhancing intra-class diversity. Additionally, to further enhance generalization, we also propose mixed copy-paste augmentation (mCpAug), which integrates signal-level, generation-level, and lexical-level perturbations to better align the training data with unseen domains. Extensive cross-dataset evaluations demonstrate that the proposed methods outperform most augmentation strategies, exhibiting superior generalization across various datasets.}, archive = {J_NEUCOM}, author = {Linjuan Zhang and Baoning Niu and Kong Aik Lee and Longbiao Wang}, doi = {10.1016/j.neucom.2025.130799}, journal = {Neurocomputing}, month = {10}, pages = {130799}, shortjournal = {Neurocomputing}, title = {Make full use of your data: On copy-based augmentation in speech anti-spoofing}, volume = {649}, year = {2025}, }
• (2025). Cross-architecture knowledge distillation for speech enhancement: From CMGAN to unet. NEUCOM, 649, 130798. (www)

  @article{ , abstract = {Speech enhancement plays a crucial role in improving speech clarity and intelligibility in noisy environments, benefiting applications such as telecommunications, assistive hearing, voice-activated systems, and automatic speech recognition. Although transformer and conformer-based models achieve state-of-the-art performance, their complexity makes real-time deployment on resource-constrained devices impractical. In contrast, convolutional neural networks offer efficient on-device deployment due to their lower computational demands and better hardware compatibility. To bridge the gap between these architectures, knowledge distillation has emerged as a promising model compression technique, allowing knowledge transfer from a complex teacher model to a lightweight student model. However, conventional methods typically assume architectural similarity between teacher and student models, limiting their effectiveness in cross-architecture settings. To mitigate the challenges posed by heterogeneous architectures, we introduce an auxiliary teacher model as an intermediary between the primary teacher and the student that facilitates smoother knowledge transfer by aligning intermediate representations. Our experiments on the VoiceBank+DEMAND and LibriMix datasets demonstrate that this intermediary-based KD approach significantly enhances the student model’s performance, outperforming direct knowledge transfer. To the best of our knowledge, this is the first work exploring KD across heterogeneous architectures for the SE task, paving the way for efficient yet high-performing speech enhancement models.}, archive = {J_NEUCOM}, author = {Khanh Nguyen and Huy Tien Nguyen}, doi = {10.1016/j.neucom.2025.130798}, journal = {Neurocomputing}, month = {10}, pages = {130798}, shortjournal = {Neurocomputing}, title = {Cross-architecture knowledge distillation for speech enhancement: From CMGAN to unet}, volume = {649}, year = {2025}, }
• (2025). A novel lightweight deep attention network for automated nuclei segmentation in histopathology images. NEUCOM, 649, 130797. (www)

  @article{ , abstract = {Cell nuclei segmentation in histopathological images is essential for developing computer-aided diagnostic (CAD) systems for cancer diagnosis and prognosis. However, this task remains challenging due to poor staining, low contrast, irregular shapes, and overlapping nuclei in histology tissue images. Traditional encoder–decoder-based architectures often struggle to capture fine-grained spatial details while also being computationally expensive. To address these challenges, we propose a novel lightweight deep attention network comprising three residual blocks in the encoder, a bottleneck block, and three wavelet-driven attention blocks in the decoder. The residual blocks used in the encoder effectively extract high-level features, while the bottleneck block captures global multi-resolution features. A newly introduced wavelet-driven attention block in the decoder leverage high-frequency two-dimensional discrete wavelet transform coefficients, that captures the finer edge-level details that are often lost during the encoding process. Evaluated on two publicly available datasets, PanNuke and TNBC, the proposed architecture achieved five-fold cross-validation Jaccard Index scores of 74 . 47 ± 0 . 74 and 71 . 21 ± 0 . 82 , respectively, at a 95% confidence level. The proposed architecture has significantly fewer trainable parameters and a smaller model size than existing architectures without compromising its performance. To further validate its efficacy, the model was tested on the MonuSeg dataset as an independent cohort, achieving a Jaccard Index score of 64 . 17 ± 1 . 75 . Wilcoxon signed-rank and Scott-Knott ESD tests confirmed that the proposed architecture is statistically superior to existing models. Finally, Grad-CAM heatmaps revealed its superior focus on nuclei regions compared to conventional designs.}, archive = {J_NEUCOM}, author = {Rashika Bagri and Ankit Rajpal and Naveen Kumar}, doi = {10.1016/j.neucom.2025.130797}, journal = {Neurocomputing}, month = {10}, pages = {130797}, shortjournal = {Neurocomputing}, title = {A novel lightweight deep attention network for automated nuclei segmentation in histopathology images}, volume = {649}, year = {2025}, }
• (2025). VOS: Towards thermal infrared image colorization via view overlap strategy. NEUCOM, 649, 130793. (www)

  @article{ , abstract = {Currently, a significant challenge still exists in thermal infrared images colorization, as current methods struggle with translating texture naturally and achieving color accuracy. To overcome this challenge, we propose a View Overlap Strategy (VOS) for colorizing infrared images. The proposed VOS employs a dual-branch generator designed to translate different regions of the same object into colorization output, and it evaluates the generated overlapping regions through an Optimal Adversarial Strategy (OAS) to determine the best generator output results. To achieve whole-image colorization, a unique sliding mechanism is designed that gradually extends the colorized region over the entire infrared image, continually approximating the final colorization result during the dual-branch generator’s adversarial training. Extensive experiments on the FLIR dataset and KAIST dataset demonstrate that the proposed VOS can be applied within existing colorization adversarial networks, leading to superior performance metrics and visual quality. The color images generated through our proposed VOS present enhanced clarity and realism.}, archive = {J_NEUCOM}, author = {Weida Zhan and Yilin Wang and Yu Chen and Hang Yang and Guilong Zhao and Yingying Wang and Shujie Zhai and Tianyun Luan and Deng Han}, doi = {10.1016/j.neucom.2025.130793}, journal = {Neurocomputing}, month = {10}, pages = {130793}, shortjournal = {Neurocomputing}, title = {VOS: Towards thermal infrared image colorization via view overlap strategy}, volume = {649}, year = {2025}, }
• (2025). I2CKD: Intra- and inter-class knowledge distillation for semantic segmentation. NEUCOM, 649, 130791. (www)

  @article{ , abstract = {This paper proposes a new knowledge distillation method tailored for image semantic segmentation, termed Intra- and Inter-Class Knowledge Distillation (I2CKD). The key novelty lies in its dual focus on transferring both intra-class and inter-class knowledge between intermediate layers of the teacher (cumbersome model) and student (compact model). For knowledge extraction, we exploit class prototypes derived from feature maps. To facilitate knowledge transfer, we employ a triplet loss in order to minimize intra-class variances and maximize inter-class variances between teacher and student prototypes. Consequently, I2CKD enables the student to better mimic the feature representation of the teacher for each class, thereby enhancing the segmentation performance of the compact network. Extensive experiments on four segmentation datasets, i.e., Cityscapes, Pascal VOC, CamVid and ADE20K, using various teacher–student network pairs demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Ayoub Karine and Thibault Napoléon and Maher Jridi}, doi = {10.1016/j.neucom.2025.130791}, journal = {Neurocomputing}, month = {10}, pages = {130791}, shortjournal = {Neurocomputing}, title = {I2CKD: Intra- and inter-class knowledge distillation for semantic segmentation}, volume = {649}, year = {2025}, }
• (2025). Meta-guided dual path convolutional neural network for depression diagnosis with functional MR images. NEUCOM, 649, 130790. (www)

  @article{ , abstract = {Deep learning methods with functional magnetic resonance imaging (fMRI) data are successful in diagnosis of depression. However, developing robust models remains challenging due to small datasets and individual heterogeneity. Domain knowledge has the potential to enhance deep-learning-based diagnosis. Previous imaging studies reported abnormalities in brain regions, and meta-analysis can identify spatially convergent abnormal regions. In the present study, we proposed a meta-guided deep-learning framework integrating meta-analysis findings as domain knowledge. We designed the preprocessing of the meta-map for framework integration and developed meta convolutional block and meta dual-path block using the meta-map’s disease-associated regions as spatial guidance to learn ReHo features. Our framework achieved 76.9 % accuracy using a cohort of 385 subjects (192 healthy controls and 193 depressed patients). The effectiveness of the meta-map was comprehensively validated through extensive comparative experiments and systematic ablation studies. Experiments validated the meta-map’s effectiveness in addressing limited sample size and heterogeneity issues. This study introduces the findings of traditional literature beyond the given medical dataset, providing a more promising approach to addressing the problem of small-sized medical datasets for psychiatric disorders.}, archive = {J_NEUCOM}, author = {Jinjun Zhou and Si Zhou and Xinyi Wang and Junneng Shao and Qing Lu}, doi = {10.1016/j.neucom.2025.130790}, journal = {Neurocomputing}, month = {10}, pages = {130790}, shortjournal = {Neurocomputing}, title = {Meta-guided dual path convolutional neural network for depression diagnosis with functional MR images}, volume = {649}, year = {2025}, }
• (2025). LGL: Local guide local network for non-homogeneous image dehazing. NEUCOM, 649, 130789. (www)

  @article{ , abstract = {Non-homogeneous image dehazing (NHID) presents a significant challenge due to its uneven haze distribution. Information on areas with different densities contributes to the dehazing of other areas, necessitating long-range modeling capabilities in dehazing methodologies. Leveraging the global self-attention, Transformer architectures offer promise for NHID. However, pixel-level token embedding and the quadratic complexity of self-attention pose hurdles for effectively modeling non-homogeneous haze images. In this paper, we propose a novel method tailored for NHID, named Local Guide Local Network (LGL), which selects information with low similarity within each local window to guide the dehazing of other local windows. Specifically, we propose an Orthogonal Token Selector, optimized geometrically on the Stiefel manifold, to select unique tokens with top- k attention values. Subsequently, a Global Sparse Attention is performed among local windows to extract global features and exchange features of local areas based on the selected tokens. Finally, a Local Guided Feedforward Network is proposed to achieve region-specific dehazing. Extensive experimental evaluations validate the efficacy of our LGL for NHID, demonstrating superior performance over existing methods.}, archive = {J_NEUCOM}, author = {Zhongze Wang and Haitao Zhao and Jingchao Peng and Kaijie Zhao and Lujian Yao}, doi = {10.1016/j.neucom.2025.130789}, journal = {Neurocomputing}, month = {10}, pages = {130789}, shortjournal = {Neurocomputing}, title = {LGL: Local guide local network for non-homogeneous image dehazing}, volume = {649}, year = {2025}, }
• (2025). Multi-view unsupervised feature selection via global and local kernelized graph learning. NEUCOM, 649, 130786. (www)

  @article{ , abstract = {To cope with the impact of the nonlinear characteristics of real data on feature selection, we propose a new multi-view unsupervised feature selection method called Multi-view Unsupervised Feature Selection via Global and Local Kernelized Graph Learning (FSGLK). Our method leverages the self-representation property of samples to capture global structures through kernelized graph learning and utilizes the learned kernelized graph to construct high-order tensors for capturing high-order relationships between views. In addition, we employ a kernelized adaptive neighborhood strategy to enhance the model’s ability to capture the local structures of complex data. The constructed graph can more effectively capture both the local and global structures of multi-view data while eliminating redundant features in high-dimensional data. Symmetric non-negative matrix factorization is used to obtain low-dimensional representations, on which feature selection is performed. To flexibly control matrix row sparsity, the ℓ 2 , p -norm is introduced. Experimental evaluations on multiple benchmark datasets show that the proposed FSGLK method significantly outperforms existing methods in terms of clustering accuracy, consistency and information sharing.}, archive = {J_NEUCOM}, author = {Min Xu and Xijiong Xie and Yuqi Li and Guoqing Chao}, doi = {10.1016/j.neucom.2025.130786}, journal = {Neurocomputing}, month = {10}, pages = {130786}, shortjournal = {Neurocomputing}, title = {Multi-view unsupervised feature selection via global and local kernelized graph learning}, volume = {649}, year = {2025}, }
• (2025). UdpTrace: Utility-enhanced differential privacy scheme for trajectory data publishing. NEUCOM, 649, 130785. (www)

  @article{ , abstract = {Differential privacy is a popular method for preserving privacy in trajectory data publishing, allowing the utilization of user data while protecting sensitive information. However, trajectory data publishing under differential privacy often suffers from low utility and limits its value for meaningful analysis. In this paper, we propose UdpTrace, a utility-enhanced differential privacy scheme for trajectory data publishing. First, we introduce a method for discretizing geographic regions based on trajectory density. This method employs a coarse first-layer grid to enhance privacy protection and a fine second-layer grid to capture detailed trajectory information, thereby maintaining privacy while improving the application value of the data. Second, we build a semantic trip transfer matrix by aggregating the location label transfer probabilities and the cell transfers under the corresponding labels to maintain the semantic connection between the start and end regions. Third, we develop a novel sampling method to generate realistic intermediate trajectory points leveraging local transition patterns captured in the second-layer grid. Experimental results indicate that the proposed scheme effectively preserves privacy while significantly improving the utility of trajectory data.}, archive = {J_NEUCOM}, author = {Wei Sun and Kui Zhao and Gang Liang and Zhiwei Liang and Lingla Jiang}, doi = {10.1016/j.neucom.2025.130785}, journal = {Neurocomputing}, month = {10}, pages = {130785}, shortjournal = {Neurocomputing}, title = {UdpTrace: Utility-enhanced differential privacy scheme for trajectory data publishing}, volume = {649}, year = {2025}, }
• (2025). Regularizing classifier for few-shot object detection via feature augmentation and random labeling. NEUCOM, 649, 130784. (www)

  @article{ , abstract = {Few-shot object detection (FSOD) aims at using few-shot samples to train a detector for novel categories, which has received considerable attention in recent years. However, the scarcity of samples often leads to overfitting and poor generalization ability, resulting in unsatisfactory performance on novel classes. To address this issue, we propose a simple yet effective feature augmentation method for object detection network, which mitigates the problems of few-shot scenarios. We incorporate random labeling to reduce misplaced confidence caused by low-quality few-shot samples and further enhance the model’s generalization ability. Moreover, we derive an upper bound for the loss function of generated features, which can be used as an auxiliary regularization loss without explicitly augmenting features. Finally, a comprehensive approach for evaluating the model’s generalization ability is proposed. Extensive experiments on PASCAL VOC and MSCOCO datasets demonstrate the effectiveness of our method, achieving competitive or even superior performance compared to state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Yiming Peng and Shuxiao Li}, doi = {10.1016/j.neucom.2025.130784}, journal = {Neurocomputing}, month = {10}, pages = {130784}, shortjournal = {Neurocomputing}, title = {Regularizing classifier for few-shot object detection via feature augmentation and random labeling}, volume = {649}, year = {2025}, }
• (2025). Event extraction for visual: De-biasing with causality-guided attention mechanism. NEUCOM, 649, 130783. (www)

  @article{ , abstract = {The Transformer model has demonstrated remarkable performance in artificial intelligence, achieving unprecedented success across a wide range of tasks. At its core, the attention mechanism remains an indispensable component of advanced Transformer architectures, with its significance and contributions widely acknowledged by the AI community. However, as research deepens and the demand for solving more complex real-world tasks increases, researchers have gradually identified bottlenecks in the self-attention mechanism, with some hypotheses validated through experimental observations. Despite these advances, current studies on the limitations of the self-attention mechanism are primarily centered on empirical observations and experimental results, lacking rigorous theoretical foundations. This gap has led to limited interpretability and unclear rationales behind some findings, despite their potential significance. In this paper, we investigate the limitations of the attention mechanism in Transformer models and provide an in-depth analysis of its capability bottleneck within the data-driven research paradigm. We identify a distinct capability boundary of the traditional attention mechanism and present a comprehensive theoretical analysis based on causal sampling theory. To validate our theoretical findings, we evaluate them in the context of the challenging task of Video Situation Recognition (VSR), obtaining experimental results that align with our theoretical conclusions. We hope that the findings of this study will inspire new perspectives on the capabilities of attention mechanisms and contribute positively to the development of optimization strategies in attention mechanism design.}, archive = {J_NEUCOM}, author = {Chao Wang and Jianming Yang and Qing He}, doi = {10.1016/j.neucom.2025.130783}, journal = {Neurocomputing}, month = {10}, pages = {130783}, shortjournal = {Neurocomputing}, title = {Event extraction for visual: De-biasing with causality-guided attention mechanism}, volume = {649}, year = {2025}, }
• (2025). Unsupervised contrastive learning using out-of-distribution data for long-tailed dataset. NEUCOM, 649, 130779. (www)

  @article{ , abstract = {This work addresses the task of self-supervised learning (SSL) on a long-tailed dataset that aims to learn balanced and well-separated representations for downstream tasks such as image classification. This task is crucial because the real world contains numerous object categories, and their distributions are inherently imbalanced. Towards robust SSL on a class-imbalanced dataset, we investigate leveraging a network trained using unlabeled out-of-distribution (OOD) data that are prevalently available online. We first train a network using both in-domain (ID) and sampled OOD data by back-propagating the proposed pseudo semantic discrimination loss alongside a domain discrimination loss. The OOD data sampling and loss functions are designed to learn a balanced and well-separated embedding space. Subsequently, we further optimize the network on ID data by unsupervised contrastive learning while using the previously trained network as a guiding network. The guiding network is utilized to select positive/negative samples and to control the strengths of attractive/repulsive forces in contrastive learning. We also distil and transfer its embedding space to the training network to maintain balancedness and separability. Through experiments on four publicly available long-tailed datasets, we demonstrate that the proposed method outperforms previous state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Cuong Manh Hoang and Yeejin Lee and Byeongkeun Kang}, doi = {10.1016/j.neucom.2025.130779}, journal = {Neurocomputing}, month = {10}, pages = {130779}, shortjournal = {Neurocomputing}, title = {Unsupervised contrastive learning using out-of-distribution data for long-tailed dataset}, volume = {649}, year = {2025}, }
• (2025). DRIR-net: Dual-branch rotation invariant and robust network for 3D place recognition. NEUCOM, 649, 130778. (www)

  @article{ , abstract = {Place recognition using 3D point clouds (LiDAR) is crucial in autonomous driving, especially for simultaneous localization and mapping (SLAM) in challenging outdoor environments. Nowadays learning-based methods face significant challenges, such as substantial preprocessing time, difficulty in handling viewpoint rotation, and inability to resist noise interference. To address these issues, we propose a Dual-Branch Rotation Invariant and Robust Network (DRIR-Net) for 3D place recognition. In DRIR-Net, we leverage the characteristics of LiDAR scanning to segment the point cloud and employ a subregion downsampling method to reduce preprocessing time significantly. Utilizing this subregion structure, we design a local position encoding and feature extraction module to achieve robustness to data noise. Furthermore, we propose a global feature augmentation module for subregions and generate rotation-invariant global descriptors combined with symmetry functions. Additionally, we improve place recognition performance by fusing point and voxel features with subregion cross attention. Through extensive experiments, our method outperforms state-of-the-art approaches on the USyd, Oxford RobotCar, and KITTI datasets. Importantly, it effectively handles scenes with noise from rotated viewpoints and generalizes well across different environments.}, archive = {J_NEUCOM}, author = {Ming Liao and Xiaoguang Di and Maozhen Liu and Teng Lv and Xiaofei Zhang and Runwen Zhu}, doi = {10.1016/j.neucom.2025.130778}, journal = {Neurocomputing}, month = {10}, pages = {130778}, shortjournal = {Neurocomputing}, title = {DRIR-net: Dual-branch rotation invariant and robust network for 3D place recognition}, volume = {649}, year = {2025}, }
• (2025). Search to integrate multi-level heuristics with graph neural networks for multi-relational link prediction. NEUCOM, 649, 130776. (www)

  @article{ , abstract = {Multi-relational link prediction aims to forecast relationships among nodes within multi-relational graphs, with applications ranging from predicting drug interactions to completing knowledge graphs. Recent advancements have demonstrated that graph neural networks (GNNs), when augmented with heuristic information, significantly enhance the performance in this domain. However, existing approaches are limited by their uniform application of a single heuristic level across various datasets and often overlook the synergy between heuristic information and GNN architecture. Inspired by the successes of neural architecture search (NAS), this paper proposes a novel strategy that seeks to integrate multi-level heuristics with GNNs for enhanced multi-relational link prediction. This strategy involves a new framework that incorporates heuristic information at both global and local levels into GNNs, facilitating a cohesive application of these heuristics within the architecture. Building upon this framework, we have developed an extensive search space that includes widely-used heuristics and operations prevalent in manually designed architectures. Moreover, we employ a versatile search algorithm tailored to tackle the bi-level optimization challenge, ensuring efficient exploration of the search space. Empirical evaluations conducted on four benchmark datasets demonstrate that the proposed method significantly surpasses existing baselines in the task of multi-relational link prediction.}, archive = {J_NEUCOM}, author = {Junjie Wu and Haotong Du and Haowei Xu and Xianghua Li and Chao Gao and Zhen Wang}, doi = {10.1016/j.neucom.2025.130776}, journal = {Neurocomputing}, month = {10}, pages = {130776}, shortjournal = {Neurocomputing}, title = {Search to integrate multi-level heuristics with graph neural networks for multi-relational link prediction}, volume = {649}, year = {2025}, }
• (2025). A progressive sampling method for object detection performance surface based on gaussian process multi-kernel fusion. NEUCOM, 649, 130775. (www)

  @article{ , abstract = {With increased interest in robustness evaluation of deep learning models, performance assessments under single-dimensional perturbations have been extensively studied, resulting in the establishment of numerous benchmarks. However, the behavior of models under bi-dimensional perturbations remains underexplored. A key issue arises from the exponential growth in sampling requirements when modeling performance surfaces in two-dimensional perturbation spaces, resulting in significant computational overhead. To address this issue, we propose a progressive sampling method for object detection performance surfaces that uses multi-kernel Gaussian process fusion. Our method incorporates a genetic algorithm to optimize kernel composition, leveraging the superior surface fitting capabilities and uncertainty quantification of composite kernel Gaussian processes. A reinforcement learning strategy is used to generate an initial population with high diversity and broad coverage. In addition, a whale optimization algorithm is used to fine-tune the weights and parameters of individual kernels, thereby improving sampling efficiency. Experimental results show that the proposed method significantly improves the sampling efficiency of performance surfaces, effectively reducing the number of samples required. This provides a reliable and efficient solution for robustness evaluation of deep learning models under complex perturbation scenarios.}, archive = {J_NEUCOM}, author = {Pengcheng Wang and Huanyu Liu and Junbao Li}, doi = {10.1016/j.neucom.2025.130775}, journal = {Neurocomputing}, month = {10}, pages = {130775}, shortjournal = {Neurocomputing}, title = {A progressive sampling method for object detection performance surface based on gaussian process multi-kernel fusion}, volume = {649}, year = {2025}, }
• (2025). A review of few-shot image classification: Approaches, datasets and research trends. NEUCOM, 649, 130774. (www)

  @article{ , abstract = {Over the past decade, deep learning has made significant advancements in image classification. However, these models struggle with data scarcity and distribution shifts, commonly referred to as the few-shot image classification (FSIC) problem. FSIC aims to recognize novel classes using only a limited number of labeled samples, posing challenges for conventional deep learning models that rely on large datasets for optimal performance. This paper provides a comprehensive review of FSIC methodologies, categorizing them into five main approaches: meta-learning, transfer learning, data augmentation, attribute-related, and vision-language foundation model adaptation. Meta-learning approaches are further classified into metric-based, model-based, and optimization-based methods, while transfer learning approaches are divided into hybrid and non-hybrid methods. Vision-language foundation model adaptation approaches are grouped into few-shot parameter tuning, dynamic or unsupervised tuning, and training-free adaptation methods. Beyond general FSIC, this paper also explores specialized FSIC methods in fine-grained classification, cross-domain classification, and class-incremental learning. Additionally, it reviews commonly used few-shot image datasets and compares the performance of representative methods through experimental results. Practical applications of FSIC across various domains are also discussed, highlighting its potential to address real-world challenges. Finally, the research trends of FSIC are identified, offering insights into the state-of-the-art FSIC methods and guiding future advancements in this field.}, archive = {J_NEUCOM}, author = {Jit Yan Lim and Kian Ming Lim and Chin Poo Lee and Yong Xuan Tan}, doi = {10.1016/j.neucom.2025.130774}, journal = {Neurocomputing}, month = {10}, pages = {130774}, shortjournal = {Neurocomputing}, title = {A review of few-shot image classification: Approaches, datasets and research trends}, volume = {649}, year = {2025}, }
• (2025). MCGT: Multi-class graph model driven transformer for next POI recommendation. NEUCOM, 649, 130773. (www)

  @article{ , abstract = {Location-based social networks (LBSNs) play a critical role in the next point-of-interest (POI) recommendation by analyzing user movement patterns. However, it is difficult to accurately define user behavior, as it requires capturing spatial, temporal, and social dependencies together under data sparsity and dynamic change. Existing methods, such as recurrent neural network or global graph model, often ignore localized behavioral patterns and suffer from noise in irrelevant POI nodes. To address these challenges, we propose MCGT, a Multi-Classification Graph Model-Driven Transformer framework that combines community-driven graphs and hierarchical feature learning. MCGT partitions users into communities with similar spatial–temporal trajectories, constructs localized subgraphs to filter noise, and identifies fine-grained community-specific patterns. The framework employs stacked encoders and multi-head attention to model intra-community features and cross-community interactions, while integrating spatial semantics (POI sequences) and temporal semantics (check-in timestamps). Experiments on two real-world datasets (Foursquare-NYC and Foursquare-TKY) confirm MCGT’s superiority, which surpasses state-of-the-art method MCLP with 5.29% and 4.99% gains in ACC@1 and MRR (on Foursquare-TKY), respectively. These results highlight its capability to understand complex user behaviors for precise POI recommendations.}, archive = {J_NEUCOM}, author = {Xianjie Qiu and Ze Wang and Zixi Zang and Chao Yuan and Shimin Sun}, doi = {10.1016/j.neucom.2025.130773}, journal = {Neurocomputing}, month = {10}, pages = {130773}, shortjournal = {Neurocomputing}, title = {MCGT: Multi-class graph model driven transformer for next POI recommendation}, volume = {649}, year = {2025}, }
• (2025). Verbalization diversification for relation extraction with long-tailed entity augmentation. NEUCOM, 649, 130771. (www)

  @article{ , abstract = {In recent years, there has been growing interest in few-shot relation extraction on real-world datasets with long-tailed distribution. However, existing approaches often lack sufficient generalization abilities with limited annotations. In this paper, we propose an easy-and-effective architecture that gives sufficient consideration to relation verbalization diversification and tail generalization. Specifically, we focus on optimizing and diversifying verbalizations of relation labels to improve the performance of sentence-level relation extraction in low-resource settings. With augmented data, the architecture can benefit from the knowledge and patterns of common entities contained within the pre-trained model when facing long-tailed entities. The experimental results on TACRED, TACREV, Re-TACRED show that our model can be significantly and consistently comparable to existing state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Yiping Liang and Tianbao Jiang and Xin Yi and Yan Cai and Linlin Wang and Xiaoling Wang and Liang He}, doi = {10.1016/j.neucom.2025.130771}, journal = {Neurocomputing}, month = {10}, pages = {130771}, shortjournal = {Neurocomputing}, title = {Verbalization diversification for relation extraction with long-tailed entity augmentation}, volume = {649}, year = {2025}, }
• (2025). MV-guided deformable convolution network for compressed video action recognition with P-frames. NEUCOM, 649, 130770. (www)

  @article{ , abstract = {Large-scale deep models have driven substantial progress in action recognition, but their heavy computation and the use of full-resolution RGB frames raise latency and privacy concerns. Compressed-domain methods reduce overhead by operating on codec outputs (I-frames, P-frames) but still rely on privacy-sensitive I-frames and incur nontrivial decoding costs. To overcome these limitations, we propose a novel P-frame only framework that (1) employs deformable convolutions to exploit the spatial sparsity of residual maps in P-frames and (2) introduces a Motion Vector–guided Deformable Convolution Network (MV-DCN) that uses motion vectors to predict adaptive sampling offsets. To transfer semantic knowledge from RGB features without decoding I-frames, we further design a Motion-Appearance Mutual Learning (MA-ML) scheme for cross domain distillation. Extensive experiments demonstrate that our model achieves competitive accuracy and speed compared to raw domain and traditional compressed domain approaches, while effectively preserving privacy by utilizing only P-frames.}, archive = {J_NEUCOM}, author = {Yuting Mou and Ke Xu and Xinghao Jiang and Tanfeng Sun}, doi = {10.1016/j.neucom.2025.130770}, journal = {Neurocomputing}, month = {10}, pages = {130770}, shortjournal = {Neurocomputing}, title = {MV-guided deformable convolution network for compressed video action recognition with P-frames}, volume = {649}, year = {2025}, }
• (2025). APG-DPNet: A dual-path network with anatomical priors for perigastric veins segmentation and varicosity quantification. NEUCOM, 649, 130767. (www)

  @article{ , abstract = {The precise diagnosis of Pancreatic Portal Hypertension (PPH) hinges on the accurate anatomical localization and quantitative morphological assessment of perigastric veins. In order to address the deficiencies of conventional methods, which are prone to missegmentation and omission in complex vascular environments, this study proposes an Anatomical Prior Guided Dual-Path Network (APG-DPNet) to achieve improved segmentation of perigastric veins. The proposed network comprises two complementary pathways: the Anatomical Prior Path and the Vein Segmentation Path. The Anatomical Prior Path leverages gastric anatomical masks to provide spatial constraints, effectively mitigating interference from adjacent structures such as the portal vein. The Vein Segmentation Path adopts a cascaded U-Net architecture for pixel-level segmentation and integrates a scale-adaptive vein diameter estimation module for precise morphological quantification. Additionally, a Multi-Condition Constraint Module (MCCM) is designed to incorporate vein-specific characteristics, including morphological features, contrast enhancement, and diameter evaluation, to refine candidate regions and improve segmentation accuracy, particularly for the identification of varicose veins. Experimental results indicate that the proposed APG-DPNet outperforms state-of-the-art (SOTA) methods in perigastric vein segmentation. On internal, external validation, and public datasets, the model achieved Dice scores of 82.3%, 79.6%, and 80.6%, respectively. Moreover, the correlation between predicted vein diameters and manual measurements reached 0.90, validating the model’s accuracy and reliability in vein morphology quantification. APG-DPNet serves as a robust and dependable auxiliary tool for the automated quantitative diagnosis of PPH, offering strong generalization capability and substantial clinical applicability.}, archive = {J_NEUCOM}, author = {Kun Zhang and Chenggang Wu and Wenkai Wei and Fengxiu Yan and Lin Wang and Peijian Zhang and Liang Hua and Lei Cui}, doi = {10.1016/j.neucom.2025.130767}, journal = {Neurocomputing}, month = {10}, pages = {130767}, shortjournal = {Neurocomputing}, title = {APG-DPNet: A dual-path network with anatomical priors for perigastric veins segmentation and varicosity quantification}, volume = {649}, year = {2025}, }
• (2025). Node classification based graph classification with latent sample graph generation and dense graph optimization. NEUCOM, 649, 130766. (www)

  @article{ , abstract = {Recent studies have shown that the utilization of multi-scale graphs can achieve better results than single-scale graph structures in downstream tasks. However, due to the randomness of generating latent sample graphs from different data partitions, additional tuning costs are usually required for similar generated graphs, which requires more pre-training to adjust the overall model and makes it difficult to guarantee model stability. In this work, a new framework called Node Classification based Graph Classification (NbG) is proposed, which combines the advantage of node classification and graph classification. The NbG can construct latent sample graphs through multi-graph population fusion methods, enabling the exploration of graph structures across different granularity levels. To ensure the validity of the obtained latent sample graph, a novel method called Latent Sample Graph Generation (LSGG) is proposed for graph construction. Finally, a transformer-based graph node classification method is used to perform graph node classification on the obtained latent sample graph. To evaluate the superiority and generalizability of our model in graph classification, the extensive experiments is conducted on seven datasets in the fields of bioinformatics, small molecules, and social networks. The codes are publicly available at https://github.com/TWoodKayak/NbG .}, archive = {J_NEUCOM}, author = {Huayang Liu and Xiaoxin Guo and Xuanru Li and Longchen Su and Chenxi Liu and Guangqi Yang and Hongliang Dong}, doi = {10.1016/j.neucom.2025.130766}, journal = {Neurocomputing}, month = {10}, pages = {130766}, shortjournal = {Neurocomputing}, title = {Node classification based graph classification with latent sample graph generation and dense graph optimization}, volume = {649}, year = {2025}, }
• (2025). WaveletFT: Discrete wavelet transform for parameter-efficient fine-tuning. NEUCOM, 649, 130765. (www)

  @article{ , abstract = {Recently, Low-rank adaptation (LoRA) has achieved significant popularity for fine-tuning foundational models owing to its ability to substantially reduce the number of trainable parameters and avoid additional inference costs. This reduction is achieved by introducing low-rank matrices A and B to represent weight update,defined as Δ W = AB . Nonetheless, the accuracy gap frequently remains between LoRA and full fine-tuning (FT). Additionally, LoRA encounters difficulties concerning storage, particularly when extensive customization adaptations or larger base models are involved. In this work, we aim to resemble the learning capacity of FT and further reduce trainable parameters by leveraging the robust expressiveness of the wavelet transform (WT). We present a novel approach, named WaveletFT, which treats Δ W as a matrix within the spatial domain and focuses on learning only a small subset of its coefficients. By employing the trained spectral coefficients, we utilize the inverse discrete WT to reconstruct Δ W . Experimental results demonstrate that the proposed WaveletFT method offers comparable or superior performance with fewer parameters compared to LoRA across diverse tasks, such as natural language understanding, natural language generation, instruction tuning, image classification and text-to-image generation.}, archive = {J_NEUCOM}, author = {Can Hu and Jie Yang and Shien Song and Wentao Fan and Tao Xie}, doi = {10.1016/j.neucom.2025.130765}, journal = {Neurocomputing}, month = {10}, pages = {130765}, shortjournal = {Neurocomputing}, title = {WaveletFT: Discrete wavelet transform for parameter-efficient fine-tuning}, volume = {649}, year = {2025}, }
• (2025). Piecewise physics-constrained neural networks for solving second-order impulsive differential equations. NEUCOM, 649, 130764. (www)

  @article{ , abstract = {Due to the reliance of traditional numerical methods on grids, their performance significantly degrades when dealing with non-linear complex problems such as impulsive problems. In contrast, as mesh-free methods, physics-informed neural networks(PINNs) have more advantages in solving impulsive problems. This paper proposes a model named Piecewise-PCNNs. This model adopts multiple sub-networks architecture and hard enforcement method. It realizes the transformation from multi-objective optimization to single-objective optimization through prior information and improves the model performance by decoupling the solution space. To evaluate the performance of this model, this paper selects various types of second-order impulsive differential equations and conducts numerical experiments with other numerical methods. Experiments prove that under general circumstances, the Piecewise-PCNNs model has higher computational accuracy and efficiency, and this advantage can be extended to complex non-linear impulsive problems.}, archive = {J_NEUCOM}, author = {Junxi Chen and Liangcai Mei and Chuan Yang and Boyang Liu and Jiarong Han and Wulong Yao}, doi = {10.1016/j.neucom.2025.130764}, journal = {Neurocomputing}, month = {10}, pages = {130764}, shortjournal = {Neurocomputing}, title = {Piecewise physics-constrained neural networks for solving second-order impulsive differential equations}, volume = {649}, year = {2025}, }
• (2025). Domain generalization for person re-identification: A survey towards domain-agnostic person matching. NEUCOM, 649, 130763. (www)

  @article{ , abstract = {Person Re-identification (ReID) aims to retrieve images of the same individual captured across non-overlapping camera views, making it a critical component of intelligent surveillance systems. Traditional ReID methods assume that the training and test domains share similar characteristics and primarily focus on learning discriminative features within a given domain. However, they often fail to generalize to unseen domains due to domain shifts caused by variations in viewpoint, background, and lighting conditions. To address this issue, Domain-Adaptive ReID (DA-ReID) methods have been proposed. These approaches incorporate unlabeled target domain data during training and improve performance by aligning feature distributions between source and target domains. However, their reliance on access to target domain data limits their scalability and makes them less suitable for real-world deployments, where such data may not be available in advance. Domain-Generalizable ReID (DG-ReID) tackles a more realistic and challenging setting by aiming to learn domain-invariant features without relying on any target domain data. Recent methods have explored various strategies to enhance generalization across diverse environments, but the field remains relatively underexplored. In this paper, we present a comprehensive survey of DG-ReID. We first review the architectural components of DG-ReID including the overall setting, commonly used backbone networks and multi-source input configurations. Then, we categorize and analyze domain generalization modules that explicitly aim to learn domain-invariant and identity-discriminative representations. To examine the broader applicability of these techniques, we further conduct a case study on a related task that also involves distribution shifts. Finally, we discuss recent trends, open challenges, and promising directions for future research in DG-ReID. To the best of our knowledge, this is the first systematic survey dedicated to DG-ReID. A curated list of related resources and papers is also available at: https://github.com/PerceptualAI-Lab/Awesome-Domain-Generalizable-Person-Re-ID}, archive = {J_NEUCOM}, author = {Hyeonseo Lee and Juhyun Park and Jihyong Oh and Chanho Eom}, doi = {10.1016/j.neucom.2025.130763}, journal = {Neurocomputing}, month = {10}, pages = {130763}, shortjournal = {Neurocomputing}, title = {Domain generalization for person re-identification: A survey towards domain-agnostic person matching}, volume = {649}, year = {2025}, }
• (2025). Disinhibition maintains network performances in concept learning through regulating neural responses. NEUCOM, 649, 130760. (www)

  @article{ , abstract = {Concept learning is a role cortical function. During various cortical functions, disinhibition is a common phenomenon and underlying mechanism. However, the contribution of disinhibition in concept learning remains unclear. In this paper, we construct a two-layer disinhibitory recurrent neural spiking network to learn designed visual spatial concepts. During concept learning, the network employs a distance-based clustering method combining neural spikes and rates into identifications and receives rewards. Through reward-modulated STDP and updates of clustering sets, the trained network achieves acceptable performances in identifications. During identifications, the network simultaneously simulates phenomena of disinhibition, sparse neural responses and neural variability quenching. In disinhibition, VIP neural responses in both pre-stimulus and stimulus-presented durations interact with excitatory neural responses. Both locations of stimuli and VIP neural responding strengths modify neural interactions in disinhibition. Through neural interactions, VIP neurons regulate network responses and performances. Our simulations indicate that disinhibition serves as a regulating mechanism to maintain network performances in concept learning.}, archive = {J_NEUCOM}, author = {Weisi Liu and Ruihuan Ren}, doi = {10.1016/j.neucom.2025.130760}, journal = {Neurocomputing}, month = {10}, pages = {130760}, shortjournal = {Neurocomputing}, title = {Disinhibition maintains network performances in concept learning through regulating neural responses}, volume = {649}, year = {2025}, }
• (2025). TPSC:Text-guided parameter-efficient fine-tuning segmentation for object counting. NEUCOM, 649, 130759. (www)

  @article{ , abstract = {Existing class-agnostic object counting methods depend on closed datasets and extensive manual annotations, which are often unavailable in dynamic and open environments. In this paper, we propose a generalized neural network for text-guided zero-shot object counting based on segmentation. Our network employs the Parameter Efficient Fine-Tuning (PEFT) to explore the integration mechanisms of Contrastive Language-Image Pre-Training (CLIP) and Segment Anything Model (SAM), which extends the large model to object counting downstream tasks with minimal additional parameters. Firstly, semantic aligned features obtained from the pre-trained CLIP image and text encoder serve as prompts prior for SAM to reduce the need for manually labeled data. Then we design a Type-Text token and MaskCLIP module, both of which incorporate the semantic features. The Type-Text token focuses on text-related targets by interacting with SAM’s original tokens, and the MaskCLIP transfers the knowledge of CLIP to SAM to enhance mask quality. Finally, a lightweight Adapter module is embedded into SAM’s encoder to recognize scarce categories and count objects, which will improve counting accuracy and sensitivity. Experimental results show that our model achieves state-of-the-art performance on popular datasets, demonstrating its generalizability for deep learning segmentation counts. Code is available at https://github.com/lxinhui1109/TPSC.git .}, archive = {J_NEUCOM}, author = {Mingzhu Shi and Xinhui Lin and Yujiao Cai and Lei Liu and Wenxin Zhang and Yuhao Su}, doi = {10.1016/j.neucom.2025.130759}, journal = {Neurocomputing}, month = {10}, pages = {130759}, shortjournal = {Neurocomputing}, title = {TPSC:Text-guided parameter-efficient fine-tuning segmentation for object counting}, volume = {649}, year = {2025}, }
• (2025). MCIT: Multi-level cross-modal interactive transformer for RGBT tracking. NEUCOM, 649, 130758. (www)

  @article{ , abstract = {Visual tracking is a fundamental task in computer vision and has wide applications in various fields. By combining visible light (RGB) and thermal infrared (TIR), RGBT tracking can cope with challenges such as extreme illumination or severe weather. Due to the lack of global modeling capability, existing trackers have difficulty handling challenging scenarios such as target scale variation and occlusion. In addition, insufficient cross-modal fusion can lead to tracker failure in extreme illumination, motion blur, or other challenges. Therefore, we propose a novel RGBT tracker named Multi-level Cross-modal Interactive Transformer (MCIT), which leverages Vision Transformer (ViT) to capture global dependencies. In the MCIT, how to design cross-modal fusion module and prediction head for the ViT-extracted token features is an urgent problem. Specifically, we carefully design a Cross-modal Interaction (CMI) module and a Window-based Attention Head (WAH). The CMI modules are inserted into the original shared ViT to achieve cross-modal feature fusion at different levels. They utilize the multi-head attention mechanism, which is more suitable for capturing and integrating the semantic information between token features, thereby enhancing the diversity and complementarity of feature fusion. The WAH efficiently extracts local information of token features through self-attention within sliding windows, and aggregates global information using a hierarchical structure. It is more suitable for token feature representation, allowing for more accurate prediction of target location. Extensive experiments on four benchmarks show that our proposed MCIT tracker achieves state-of-the-art performance. The code is available at https://github.com/zjjqinyu/MCIT .}, archive = {J_NEUCOM}, author = {Yu Qin and Jianming Zhang and Shimeng Fan and Zikang Liu and Jin Wang}, doi = {10.1016/j.neucom.2025.130758}, journal = {Neurocomputing}, month = {10}, pages = {130758}, shortjournal = {Neurocomputing}, title = {MCIT: Multi-level cross-modal interactive transformer for RGBT tracking}, volume = {649}, year = {2025}, }
• (2025). DGI-GRU: Integrating dynamic graph convolution for enhanced spatio-temporal information extraction. NEUCOM, 649, 130757. (www)

  @article{ , abstract = {The ubiquity of spatio-temporal data in the real world presents significant challenges for predictive modeling due to the complex interplay between temporal trends and spatial correlations. This paper introduces the DGI-GRU model, a novel approach that enhances the Gated Recurrent Unit (GRU) architecture by integrating graph convolution with a dynamic adjacency matrix. The salient feature of the model is that it extracts dynamic graph information that leverages both the distance between nodes and time-varying observational data, effectively capturing evolving spatial relationships. Additionally, our improved GRU framework facilitates the simultaneous extraction of temporal and spatial features within a single layer, reducing parameter redundancy. Experimental results on real datasets demonstrate the superior performance of our model in spatio-temporal sequence prediction tasks, outperforming state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Xinru Liu and Weide Li and Da Zhou}, doi = {10.1016/j.neucom.2025.130757}, journal = {Neurocomputing}, month = {10}, pages = {130757}, shortjournal = {Neurocomputing}, title = {DGI-GRU: Integrating dynamic graph convolution for enhanced spatio-temporal information extraction}, volume = {649}, year = {2025}, }
• (2025). Grey-informed neural network for time-series forecasting. NEUCOM, 649, 130756. (www)

  @article{ , abstract = {The use of neural network models in a wide range of domains has shown remarkable success in solving complex problems. In spite of this, these models are often regarded as black boxes, in that they require substantial amounts of data to be trained and offer limited interpretability. As a result of data scarcity, it becomes increasingly difficult to construct effective neural network. Our study proposes a grey-informed neural network (GINN) framework in which the neural network output is constrained by a grey differential equation, thereby enhancing model transparency and accuracy when data are limited. Furthermore, we introduce a fractional-order extension, FGINN, which extends the grey-informed paradigm by introducing fractional dynamics to better capture memory properties. In empirical evaluations, the proposed GINN and FGINN reveal both critical underlying patterns in real-world datasets, as well as providing reliable forecasts despite constrained sample sizes, highlighting their potential for use as interpretable and robust solutions for small-data modeling scenarios.}, archive = {J_NEUCOM}, author = {Wanli Xie and Ruibin Zhao and Zhenguo Xu and Tingting Liang}, doi = {10.1016/j.neucom.2025.130756}, journal = {Neurocomputing}, month = {10}, pages = {130756}, shortjournal = {Neurocomputing}, title = {Grey-informed neural network for time-series forecasting}, volume = {649}, year = {2025}, }
• (2025). Small object detection network based on progressive enhanced multi-level feature fusion. NEUCOM, 649, 130755. (www)

  @article{ , abstract = {Small object detection in remote sensing imagery suffers from low-resolution and complex background interference. Feature fusion overcomes this by hierarchically combining contextual semantics and spatial details to amplify subtle discriminative patterns. However, existing fusion strategies face two major challenges. The first challenge lies in cross-layer interactions. These methods cannot dynamically establish connections between layers or adaptively select complementary features across different levels. The second challenge emerges as the network deepens. The spatial resolution of features gradually degrades, causing blurred boundaries of small objects. Additionally, the semantic discrepancies accumulate across layers, which further reduces detection reliability. To address these issues, we propose the Small Object Detection Network based on Progressive Enhanced Multi-level Feature Fusion (PEMFF) with three modules: (1) Re-parameterized Multi-information Fusion module (RepMF): Dynamic construction of feature pathways through learnable weight allocation enables semantic guidance from deep to shallow layers. It further enables the integration of details from shallow to deep layers, ensuring the efficient aggregation of multiple levels of features. This mechanism enables information to flow effectively between multiple layers. (2) Residual Multi-dimensional Attention module (RMA): Deploys cascaded spatial-channel attention with residual learning, progressively amplifying discriminative patterns in object regions while suppressing background interference. (3) Re-parameterized Information Injection module (RepIJ): Implements structural reparameterization to inject high-level semantic contexts into shallow layers via adaptive kernel transformation. It effectively resolves the semantic dilution problem in low-level features. Experimental results on the AI-TOD dataset demonstrate that our method achieves 58.4% AP 50 , surpassing YOLOv8 by 12.2%.}, archive = {J_NEUCOM}, author = {Yanshan Li and Fuxing Liu and Yusong Qin and Linhui Dai and Li Zhang and Weixin Xie}, doi = {10.1016/j.neucom.2025.130755}, journal = {Neurocomputing}, month = {10}, pages = {130755}, shortjournal = {Neurocomputing}, title = {Small object detection network based on progressive enhanced multi-level feature fusion}, volume = {649}, year = {2025}, }
• (2025). MLF-DOU: A metric learning framework with dual one-class units for network intrusion detection. NEUCOM, 649, 130754. (www)

  @article{ , abstract = {In this paper, a novel metric learning framework with dual one-class units (MLF-DOU) is proposed to address the challenges of low accuracy and poor generalization ability associated with existing models in network intrusion detection. Specifically, the merits of one-class units are leveraged, enabling compact feature representations of both normal and attack traffic to be sufficiently extracted. This extraction is beneficial for mitigating the overfitting phenomenon. On this basis, a metric learning method is introduced to further enhance the recognition ability of the model for traffic in different categories. The inter-class distance is increased, and the fine-grained representations of intra-class similarity are strengthened. By these means, both the detection performance and the generalization ability of the proposed MLF-DOU are significantly improved. Extensive experimental results are presented to demonstrate the effectiveness of MLF-DOU across three datasets, showing its superiority over other state-of-the-art methods in achieving accurate intrusion detection. The effectiveness of key components within MLF-DOU is validated, contributing to robust feature learning for each class. Moreover, the adaptability of the proposed framework is proven, as it can be integrated with various network architectures, demonstrating promising potential for real-world deployments.}, archive = {J_NEUCOM}, author = {Liang Chen and Han Li and Peishu Wu and Liwei Hu and Tengpeng Chen and Nianyin Zeng}, doi = {10.1016/j.neucom.2025.130754}, journal = {Neurocomputing}, month = {10}, pages = {130754}, shortjournal = {Neurocomputing}, title = {MLF-DOU: A metric learning framework with dual one-class units for network intrusion detection}, volume = {649}, year = {2025}, }
• (2025). Recommendation systems with LLM-based semantic embeddings and FAISS similarity search. NEUCOM, 649, 130753. (www)

  @article{ , abstract = {Content-based recommendation systems have gained significant attention for their ability to provide personalized suggestions by analyzing item descriptions. Leveraging the power of large language models (LLMs), this research introduces a novel recommendation approach that generates high-quality semantic embeddings to facilitate efficient similarity-based retrieval for Top-N recommendations. The proposed method capitalizes on the deep contextual understanding of LLMs to capture intricate semantic relationships within item content, thereby enhancing recommendation relevance. Furthermore, the system integrates FAISS (Facebook AI Similarity Search) to optimize similarity search, enabling faster and more scalable retrieval of relevant recommendations. To evaluate its effectiveness, the system is tested on four diverse real-world datasets: Yelp, Amazon Beauty, MovieLens, and LastFM, covering multiple domains. Performance is assessed using widely adopted evaluation metrics, including Normalized Discounted Cumulative Gain (NDCG), Precision, Recall, Hit Rate (HR), F1-Score and business-relevant evaluation measures. Extensive experimental results demonstrate that the proposed method, augmented with FAISS, consistently outperforms the existing state-of-the-art recommendation techniques. The code supporting this code is publicly available at: https://github.com/seemasafar/Reco-System-Using-LLM}, archive = {J_NEUCOM}, author = {Seema Safar and Babita Roslind Jose and Jimson Mathew and T. Santhanakrishnan}, doi = {10.1016/j.neucom.2025.130753}, journal = {Neurocomputing}, month = {10}, pages = {130753}, shortjournal = {Neurocomputing}, title = {Recommendation systems with LLM-based semantic embeddings and FAISS similarity search}, volume = {649}, year = {2025}, }
• (2025). MGCP: A multi-grained correlation based prediction network for multivariate time series. NEUCOM, 649, 130751. (www)

  @article{ , abstract = {Multivariate time series prediction is widely used in daily life, which poses significant challenges due to the complex correlations that exist at multi-grained levels. Unfortunately, the majority of current time series prediction models fail to simultaneously learn the correlations of multivariate time series at multi-grained levels, resulting in suboptimal performance. To address this, we propose a Multi-Grained Correlations-based Prediction (MGCP) Network, which simultaneously considers the correlations at three granularity levels to enhance prediction performance. Specifically, MGCP utilizes Adaptive Fourier Neural Operators and Graph Convolutional Networks to learn the global spatiotemporal correlations and inter-series correlations, enabling the extraction of potential features from multivariate time series at fine-grained and medium-grained levels. Additionally, MGCP employs adversarial training with an attention mechanism-based predictor and conditional discriminator to optimize prediction results at coarse-grained level, ensuring high fidelity between the generated forecast results and the actual data distribution. Finally, we compare MGCP with several state-of-the-art time series prediction algorithms on real-world benchmark datasets, and our results demonstrate the generality and effectiveness of the proposed model.}, archive = {J_NEUCOM}, author = {Zhicheng Chen and Xi Xiao and Ke Xu and Zhong Zhang and Yu Rong and Qing Li and Guojun Gan and Zhiqiang Xu and Peilin Zhao}, doi = {10.1016/j.neucom.2025.130751}, journal = {Neurocomputing}, month = {10}, pages = {130751}, shortjournal = {Neurocomputing}, title = {MGCP: A multi-grained correlation based prediction network for multivariate time series}, volume = {649}, year = {2025}, }
• (2025). SRPL-SFDA: Sam-guided reliable pseudo-labels for source-free domain adaptation in medical image segmentation. NEUCOM, 649, 130749. (www)

  @article{ , abstract = {Domain Adaptation (DA) is crucial for robust deployment of medical image segmentation models when applied to new clinical centers with significant domain shifts. Source-Free Domain Adaptation (SFDA) is appealing as it can deal with privacy concerns and access constraints on source-domain data during adaptation to target-domain data. However, SFDA faces challenges such as insufficient supervision in the target domain with unlabeled images. In this work, we propose a Segment Anything Model (SAM)-guided Reliable Pseudo-Labels method for SFDA (SRPL-SFDA) with three key components: (1) Test-Time Tri-branch Intensity Enhancement (T3IE) that not only improves quality of raw pseudo-labels in the target domain, but also leads to SAM-compatible inputs with three channels to better leverage SAM’s zero-shot inference ability for refining the pseudo-labels; (2) A reliable pseudo-label selection module that rejects low-quality pseudo-labels based on Consistency of Multiple SAM Outputs (CMSO) under input perturbations with T3IE; and (3) A reliability-aware training procedure in the unlabeled target domain where reliable pseudo-labels are used for supervision and unreliable parts are regularized by entropy minimization. Experiments conducted on two multi-domain medical image segmentation datasets for fetal brain and the prostate respectively demonstrate that: (1) SRPL-SFDA effectively enhances pseudo-label quality in the unlabeled target domain, and improves SFDA performance by leveraging the reliability-aware training; (2) SRPL-SFDA outperformed state-of-the-art SFDA methods, and its performance is close to that of supervised training in the target domain. The code of this work is available online: https://github.com/HiLab-git/SRPL-SFDA .}, archive = {J_NEUCOM}, author = {Xinya Liu and Jianghao Wu and Tao Lu and Shaoting Zhang and Guotai Wang}, doi = {10.1016/j.neucom.2025.130749}, journal = {Neurocomputing}, month = {10}, pages = {130749}, shortjournal = {Neurocomputing}, title = {SRPL-SFDA: Sam-guided reliable pseudo-labels for source-free domain adaptation in medical image segmentation}, volume = {649}, year = {2025}, }
• (2025). ODTL: An optimal deep transfer learning model for brain tumor classification. NEUCOM, 649, 130747. (www)

  @article{ , abstract = {Brain tumors are among the most life-threatening diseases worldwide, causing thousands of deaths each year due to the rapid growth of tumor cells. Consequently, developing automated methods for detecting brain tumors is vital to preserving countless lives worldwide. Deep transfer learning has proven to be an effective approach for detecting and classifying brain tumors, including tasks involving binary classification as well as multi-class classification. However, utilizing pretrained models often requires specific adaptations. Current adaptation methods focus on freezing or fine-tuning certain layers or the feature extraction component, without accounting for the contribution of individual kernels. In this work, we introduce a novel Optimal Deep Transfer Learning (ODTL) model for classifying brain tumors. Indeed, we propose to control the contribution of each source kernel using an adaptive optimization model, aiming to enhance the efficacy of transfer learning. The proposed model selects optimal source features that contribute in the target task. The derived optimization model is resolved utilizing Evolutionary Genetic Algorithms. The method was evaluated on both binary and multi-class brain tumor classification task using a medical MRI dataset. Several experiments confirm the efficacy of our model in identifying both frozen and fine-tuned kernels, thereby improving data classification. The proposed method surpasses traditional approaches and several significant state-of-the-art techniques in brain tumor classification.}, archive = {J_NEUCOM}, author = {Otmane Mallouk and Nour-Eddine Joudar and Mohamed Ettaouil}, doi = {10.1016/j.neucom.2025.130747}, journal = {Neurocomputing}, month = {10}, pages = {130747}, shortjournal = {Neurocomputing}, title = {ODTL: An optimal deep transfer learning model for brain tumor classification}, volume = {649}, year = {2025}, }
• (2025). ColorFormer: A novel colorization method based on a transformer. NEUCOM, 649, 130743. (www)

  @article{ , abstract = {This paper introduces a novel grayscale image colorization method named ColorFormer, which leverages a Transformer-based architecture enhanced with adversarial learning. Grayscale image colorization often suffers by problems such as inaccurate color placement, loss of fine details, poor color consistency across different regions, and the complexity of effectively capturing both local and global features. Conventional methods frequently produce images with artifacts, muted colors, and poor color consistency. ColorFormer addresses these challenges by introducing a novel architecture with a lightweight multi-head self-attention mechanism within the ColorFormer Block. The proposed design not only enhances the colorization process but also reduces the complexity of the self-attention mechanism. In addition, we employ a conditional Wasserstein generative adversarial network (CWGAN) framework to ensure improved color accuracy, stable training, and superior visual quality of the generated colorized images. To further enhance the visual quality of the colorized images, we incorporate perceptual loss and adversarial loss during the training phase. Experimental results demonstrate that ColorFormer significantly outperforms other state-of-the-art colorization techniques, producing more realistic and vibrant colorized images.}, archive = {J_NEUCOM}, author = {Hamza Shafiq and Truong Nguyen and Bumshik Lee}, doi = {10.1016/j.neucom.2025.130743}, journal = {Neurocomputing}, month = {10}, pages = {130743}, shortjournal = {Neurocomputing}, title = {ColorFormer: A novel colorization method based on a transformer}, volume = {649}, year = {2025}, }
• (2025). Enhanced multi-view graph convolutional networks for session-based recommendation. NEUCOM, 649, 130742. (www)

  @article{ , abstract = {Existing session-based recommendation works have demonstrated significant advantages of graph neural networks in improving recommendation prediction capabilities. However, these models still have certain limitations when it comes to deeper handling of recommendation problems. (1) Previous models are relatively weak in capturing the complex higher-order information of item transition patterns, which to some extent affects the accuracy of recommendations. (2) Existing models mainly rely on item ID information to construct user preference models, overlooking the importance of revealing user preferences from a categorical perspective, thus limiting the degree of personalized recommendations. (3) Existing models fail to effectively capture consistency preferences between sessions, thereby affecting the coherence of recommendations and user experience. To overcome these limitations, we propose an innovative enhanced multi-view graph convolutional network model. Firstly, we introduce a hypergraph on top of the local and global graphs to further model high-order item transition relationships, fully utilizing complex associative information between items. Additionally, we design a fusion network to integrate different levels of item representations, enriching the expression of recommendation information. Secondly, we innovatively propose a category transition graph aimed at accurately capturing the transfer of user category interests within sessions, providing a more precise basis for personalized recommendations. Finally, we fully consider the similarity between sessions by constructing a session graph to extract consistency preferences, enhancing the coherence and accuracy of recommendation results. Extensive experiments conducted on two real-world public datasets demonstrate the outstanding performance of our approach in achieving state-of-the-art performance, providing new insights and solutions for the development of session-based recommendation field.}, archive = {J_NEUCOM}, author = {Jin Zeng and Nan Wang and Jinbao Li}, doi = {10.1016/j.neucom.2025.130742}, journal = {Neurocomputing}, month = {10}, pages = {130742}, shortjournal = {Neurocomputing}, title = {Enhanced multi-view graph convolutional networks for session-based recommendation}, volume = {649}, year = {2025}, }
• (2025). Pseudoinverse-free zhang neurodynamics for temporally-variant nonlinear equation system solving applied to robot manipulator. NEUCOM, 649, 130735. (www)

  @article{ , abstract = {The solution of temporally-variant nonlinear equation system (TVNES) is regarded as a challenging dynamics problem with extensive applications across various fields, especially in control. In this paper, the TVNES problem is explored and solved from the perspective of control. Zhang neurodynamics (ZN) approach is used extensively for solving the TVNES problem because it is well equipped with time-derivative information. However, traditional ZN controller inevitably performs the pseudoinverse operation of temporally-variant Jacobian matrix, which has high computational complexity. Different from previous studies, a novel pseudoinverse-free ZN (PFZN) controller is proposed in this paper, which eliminates the pseudoinverse operation of temporally-variant Jacobian matrix, as a result reducing the computational complexity. Additionally, the convergence performance of the PFZN controller and the robustness of the disturbed PFZN controller are verified through theoretical analyses. Furthermore, numerical and comparative experiments show the validity of the theoretical analyses and the superiority of the PFZN controller over the gradient neurodynamics, gradient ZN, and varying-parameter ZN controllers. Finally, the path-following control problem of the UR5 robot manipulator is efficiently solved by the PFZN controller, which validates the effectiveness and applicability of the proposed PFZN controller.}, archive = {J_NEUCOM}, author = {Meichun Huang and Yunong Zhang and Shuai Li}, doi = {10.1016/j.neucom.2025.130735}, journal = {Neurocomputing}, month = {10}, pages = {130735}, shortjournal = {Neurocomputing}, title = {Pseudoinverse-free zhang neurodynamics for temporally-variant nonlinear equation system solving applied to robot manipulator}, volume = {649}, year = {2025}, }
• (2025). Adaptive zeroing neural dynamics-based cascaded control scheme for trajectory tracking of wheeled unmanned ground vehicles. NEUCOM, 649, 130734. (www)

  @article{ , abstract = {With fast convergence and high accuracy, zeroing neural dynamics (ZND) is widely used in control applications, particularly in robotic trajectory tracking. This paper focuses on the trajectory tracking of wheeled unmanned ground vehicles (WUGVs) and presents a novel adaptive zeroing neural dynamics-based cascaded control scheme (AZNDCCS). The scheme first develops an adaptive ZND (AZND) to overcome the limitations of traditional ZNDs with fixed parameters and to mitigate the unnecessary resource consumption associated with ZNDs employing varying parameters. The trajectory tracking is then decoupled into a two-step cascaded control process, which sequentially outputs the yaw rate and velocity. The AZNDCCS is ultimately achieved by integrating a single-layer AZND into the design of the yaw rate controller and a multi-layer AZND into the design of the velocity controller. Under the AZNDCCS, once the yaw angle error converges, the WUGVs reach the desired position under the velocity controller. This cascaded design further optimizes the control performance and improves the reliability of the system. As a validation, theorems about the stability and fixed-time convergence for trajectory tracking of WUGVs under the AZNDCCS are provided. Furthermore, based on the design process of AZNDCCS, two control schemes for comparison are presented: the fixed-parameter ZND-based cascaded control scheme and the varying-parameter ZND-based cascaded control scheme. Simulation results show that WUGVs under AZNDCCS achieve faster tracking speeds or lower resource consumption than the comparative schemes.}, archive = {J_NEUCOM}, author = {Lin Xiao and Wenyuan Huang and Qiuyue Zuo and Linju Li}, doi = {10.1016/j.neucom.2025.130734}, journal = {Neurocomputing}, month = {10}, pages = {130734}, shortjournal = {Neurocomputing}, title = {Adaptive zeroing neural dynamics-based cascaded control scheme for trajectory tracking of wheeled unmanned ground vehicles}, volume = {649}, year = {2025}, }
• (2025). Network dynamics mechanisms underlying the instigation and propagation of cortical spreading depression. NEUCOM, 649, 130733. (www)

  @article{ , abstract = {Cortical spreading depression (CSD) waves are widely recognized as the pathophysiological mechanism underlying migraine aura. Modeling the macroscopic phenomenological characteristics of CSD wave propagation is challenging due to the inability to capture biophysical features, while microscopic studies based on excitatory–inhibitory (E/I) neuron pairs struggle to link effectively with wave propagation behaviors. In order to couple the electrical activity of micro neurons with the macroscopic propagation behavior of the cortex, we adopt a network perspective and constructed a dual-layer ring network model. Within this unified framework, we identify four factors influencing CSD instigation and propagation: (i) the type and number of pathological neurons, (ii) the extracellular potassium concentration, (iii) the ratio of excitatory to inhibitory connections within the cortical network, and (iv) the architecture of network connectivity incorporating both short and long-range connections. Model results indicate counterintuitively that the number of initially pathological neurons does not significantly correlate with CSD propagation duration. The extracellular potassium concentration required for CSD instigation within the network is lower than that for single neurons, suggesting that coexisting cluster discharges alongside CSD may contribute to the comorbidity of epilepsy and migraine. An excessive imbalance in the E/I ratio can induce global re-entrant and retracting phenomena of CSD, whereas a higher proportion of long-range connections within the network can effectively reduce the probability of such occurrences. These findings suggest that designing intervention strategies that comprehensively consider these influential factors can effectively decrease the instigation probability of CSD or enhance the stability of brain networks during CSD propagation.}, archive = {J_NEUCOM}, author = {Chang He and Denggui Fan and Weiping Wang and Qingyun Wang and Gerold Baier}, doi = {10.1016/j.neucom.2025.130733}, journal = {Neurocomputing}, month = {10}, pages = {130733}, shortjournal = {Neurocomputing}, title = {Network dynamics mechanisms underlying the instigation and propagation of cortical spreading depression}, volume = {649}, year = {2025}, }
• (2025). MWENet: Multi-wavelet enhanced network for seasonal-trend analysis in long-term forecasting. NEUCOM, 649, 130730. (www)

  @article{ , abstract = {Deep neural networks have recently been applied to long-term time series forecasting (LTSF) tasks, aiming to capture dynamic temporal features from historical data for more accurate predictions. However, time series data inevitably contain noise, influencing prediction accuracy. Recent denoising studies suggest that high-frequency components in the frequency domain are the source of noise, and these studies use filters to remove them. While these filters effectively suppress noise, they also unintentionally discard valuable high-frequency information. To tackle this problem, We propose MWENet , a Multi-Wavelet Enhanced Network for seasonal-trend modeling in long-term forecasting. Specifically, we introduce an Enhanced Spectral Denoising module that combines Fourier Transform and Discrete Wavelet Transform (DWT) for joint denoising. This module achieves effective denoising through adaptive wavelet decomposition with multiple wavelet bases. Additionally, this work presents an Enhanced LSTM module, which integrates exponential gating units into LSTM to capture complex patterns and enhance LTSF capabilities. Extensive evaluations on seven real-world datasets for long-term forecasting tasks demonstrate that MWENet consistently achieves state-of-the-art performance, ranking first in 51 out of 70 evaluation cases and second in another 17. These results surpass a wide range of competitive baselines – including Transformer-based, MLP-based, RNN-based, and state-space models – highlighting the robustness and generalization capacity of our approach in capturing complex temporal dynamics over extended horizons.}, archive = {J_NEUCOM}, author = {Zhiwei Zhang and Qin Jiwei and Dezhi Sun and Zihao Zhang and Jing Ma and XueFeng Feng and Huiguo Zhang}, doi = {10.1016/j.neucom.2025.130730}, journal = {Neurocomputing}, month = {10}, pages = {130730}, shortjournal = {Neurocomputing}, title = {MWENet: Multi-wavelet enhanced network for seasonal-trend analysis in long-term forecasting}, volume = {649}, year = {2025}, }
• (2025). Infrared ship target tracking based on polarization enhanced features and regulations. NEUCOM, 649, 130725. (www)

  @article{ , abstract = {Visual tracking for infrared ship target is a fundamental task in machine vision, which marks the ship as the tracking and extracts feature descriptions of ship from the first infrared image frame, then tracks the ship in subsequent infrared frames with the updated target appearance models and tracking models. However, due to the existing deficiencies of correlation filter based visual trackers which included inadequate polarization context usage and spawning tracking noises within the tracking framework, this paper introduced a polarization enhanced features and regulations based infrared ship target tracker (PFRIT). The contributions of PFRIT include: 1) a polarization feature based ship mask extraction scheme is introduced to integrate the polarization context into the tracking framework via improved Mueller matrix; 2) a joint prediction framework based dual filter structure is proposed to tackle the inadequate assumptions within the traditional tracking framework; 3) a candidate based target prediction mechanism is provided to surmount the response aberrance effect and boost the tracking precision; 4) theoretical proofs for the joint tracking framework with dual filter construction and the candidate predictor are given. In addition, this paper also designated an infrared ship target tracking benchmark (ISTTB) with over 200 sequences and 80,000 frames. The extensive experiments under ISTTB evaluates the favorable performances of PFRIT and indicate that: 1) the PFRIT achieves comparable tracking results under variant and difficult infrared ship tracking scenarios against other 9 state-of-the-art trackers; 2) the PFRIT performs 38 FPS tracking speed on ISTTB benchmark.}, archive = {J_NEUCOM}, author = {Zhiyu Zhu and Denghao Yang and Huilin Ge and Xingyue Du and Xuedong Wu}, doi = {10.1016/j.neucom.2025.130725}, journal = {Neurocomputing}, month = {10}, pages = {130725}, shortjournal = {Neurocomputing}, title = {Infrared ship target tracking based on polarization enhanced features and regulations}, volume = {649}, year = {2025}, }
• (2025). DualNet: A dual-path network with adaptive compensation for multivariate time series forecasting. NEUCOM, 649, 130720. (www)

  @article{ , abstract = {Multivariate time series prediction remains a fundamental challenge across various domains due to its complex temporal dynamics and inherent dependencies among variables. This paper introduces DualNet, an innovative architectural paradigm based on decoupling principles, which achieves effective decomposition and complementary enhancement of prediction tasks through spatial decomposition. The architecture synergistically integrates global pattern extraction and adaptive local compensation mechanisms, establishing two independent yet complementary learning spaces through functional decoupling. Our framework introduces a dual-path learning strategy that decomposes the prediction process into complementary components: a primary trajectory estimator focused on capturing inherent temporal evolution patterns, and a dedicated compensation mechanism that performs fine-grained calibration through residual learning. This decoupled design not only reduces the learning complexity of individual modules but also enhances overall prediction performance through complementary effects. We developed a temperature-scaled adaptive weighting scheme that dynamically adjusts compensation intensity based on temporal context, enabling the model to achieve optimal balance between prediction stability and adaptive refinement. Through the dual mechanisms of spatial decoupling and complementary enhancement, the architecture achieves organic unification of global feature extraction and local fine-tuning, while incorporating novel normalization strategies and hierarchical feature transformation mechanisms to enhance the model’s representational capacity. Comprehensive experiments conducted on various benchmark datasets demonstrate that this decoupling-based dual-channel architecture significantly improves the model’s capability to capture complex temporal patterns while maintaining prediction stability. The code is available at https://github.com/ZS520L/DualNet .}, archive = {J_NEUCOM}, author = {Huashun Li and Weimin Wu and Wei Chen and Mei Zhang}, doi = {10.1016/j.neucom.2025.130720}, journal = {Neurocomputing}, month = {10}, pages = {130720}, shortjournal = {Neurocomputing}, title = {DualNet: A dual-path network with adaptive compensation for multivariate time series forecasting}, volume = {649}, year = {2025}, }
• (2025). Adaptive neural network visual control of closed-architecture robots with prescribed-time constraints. NEUCOM, 649, 130718. (www)

  @article{ , abstract = {Due to security or intellectual property protection concerns, most industrial robots are manufactured with closed architecture. This paper focuses on the adaptive visual control issue for robotic systems with closed architecture and prescribed performance requirements. The neural network (NN) is employed to approximate the robot torque loop controller, which is unmodified and completely unknown to the user. Then, an adaptive NN visual servoing loop controller is designed to dynamically compute the velocity commands for the torque control loop. Utilizing the Lyapunov stability analysis, it is proved that the proposed controller can guarantee the convergence of the visual servoing system, without modifying the manufacturer-embedded torque controller regardless of its configuration. Moreover, by combining a settling time regulator with a tracking error-dependent asymmetric barrier function (TEDABF), a prescribed-time constraints approach is developed, which exhibits the feature that the settling time and tracking precision can be predefined by the user and the general case of asymmetric error constraints can be handled. Experiments are conducted on an industrial manipulator to demonstrate the superior capabilities of the method.}, archive = {J_NEUCOM}, author = {Mingming Song and Changchun Hua and Yu Zhang and Keli Pang}, doi = {10.1016/j.neucom.2025.130718}, journal = {Neurocomputing}, month = {10}, pages = {130718}, shortjournal = {Neurocomputing}, title = {Adaptive neural network visual control of closed-architecture robots with prescribed-time constraints}, volume = {649}, year = {2025}, }
• (2025). Enhancing missing data imputation through combined bipartite graph and complete directed graph. NEUCOM, 649, 130717. (www)

  @article{ , abstract = {In this paper, we address a central challenge in tabular missing-data imputation: explicitly identifying and exploiting interdependencies among features to improve reconstruction quality. Current state-of-the-art methods mostly model similarity or interdependence between samples. However, our experiments on real-world tabular datasets show that, when samples are truly independent, building such observation-level graphs yields only marginal and dataset-specific performance gains, rather than consistent and generalizable benefits. We therefore introduce the Bipartite and Complete Directed Graph Neural Network (BCGNN). In BCGNN, observations and features are treated as two distinct node types, and each observed cell value is converted into an attributed edge connecting them. The bipartite component inductively learns node embeddings by fully leveraging the information encoded in these attributed edges, while the complete directed graph component explicitly describes and propagates intricate feature–feature dependencies. The combined graph furnishes a robust inductive framework for representation learning while explicitly parameterizing higher-order dependencies among features. Across diverse missing mechanisms, BCGNN outperforms leading imputation baselines, achieving an average 15% reduction in mean absolute error. Extensive experiments confirm that a deeper understanding of feature interdependence markedly enhances embedding quality. BCGNN also delivers superior performance on downstream label-prediction tasks with missing inputs and demonstrates robust generalization to unseen data.}, archive = {J_NEUCOM}, author = {Zhaoyang Zhang and Hongtu Zhu and Yingjie Zhang and Hai Shu and Ziqi Chen}, doi = {10.1016/j.neucom.2025.130717}, journal = {Neurocomputing}, month = {10}, pages = {130717}, shortjournal = {Neurocomputing}, title = {Enhancing missing data imputation through combined bipartite graph and complete directed graph}, volume = {649}, year = {2025}, }
• (2025). DICO: Distance-weighted contrast and instance correlation for salient object ranking. NEUCOM, 649, 130715. (www)

  @article{ , abstract = {Salient object ranking aims to infer the saliency levels of objects within an image and rank them accordingly. Existing methods mainly rely on attention mechanisms to model object-context relations. However, they tend to overlook two crucial aspects: the way spatial distance modulates contrast and the correlation between instances when assigning relative saliency. To address these issues, we propose a novel salient object ranking approach by modeling Distance-weighted COntrast and Instance COrrelation (DICO). Specifically, we propose the Distance-weighted Contrast (DCO) module, which utilizes Gaussian functions to simulate a dynamic attention distribution among regions. This distribution assigns higher weights to neighboring regions, capturing the inherent spatial relations in the scene. By integrating the dynamically generated attention distribution with feature-based contrast, the DCO module effectively enables more precise modeling of spatially-aware object-context and inter-object contrast. Furthermore, we propose an Instance Correlation (ICO) loss that takes into account both the inter-object correlation and individual object fitness. This dual consideration enables the model to more effectively learn the relative saliency of different objects. Specifically, the inter-object correlation helps to narrow the saliency gap between instance pairs that have close ranks, while individual object fitness aims to enhance the saliency scores of highly salient objects and reduce the saliency scores of less salient ones. Extensive experiments demonstrate that our method outperforms existing state-of-the-art approaches in terms of balancing computational complexity and performance.}, archive = {J_NEUCOM}, author = {Jinxia Zhang and Xinchao Zhu and Min Huang and Haikun Wei and Shixiong Fang and Kanjian Zhang}, doi = {10.1016/j.neucom.2025.130715}, journal = {Neurocomputing}, month = {10}, pages = {130715}, shortjournal = {Neurocomputing}, title = {DICO: Distance-weighted contrast and instance correlation for salient object ranking}, volume = {649}, year = {2025}, }
• (2025). Adversarial network for unsupervised infrared image colorization based on full-scale feature fusion and cosine contrastive learning. NEUCOM, 649, 130713. (www)

  @article{ , abstract = {Thermal infrared images, unaffected by lighting and haze, are widely used in security surveillance, autonomous vehicles, and nighttime traffic monitoring. However, their grayscale nature lacks color and texture details, limiting applications in image recognition and object detection. Converting infrared images to daytime color enhances visual perception and broadens their utility. Despite advancements in infrared image colorization, challenges such as texture distortion, detail blurring, and poor image quality persist. To address these issues, a novel unsupervised learning framework, termed Cosine Contrastive Learning Generative Adversarial Network (CCLGAN), is proposed. Firstly, the traditional UNet architecture is improved by introducing full-scale skip connections and deep supervision. Full-scale skip connections integrate low-level details with high-level semantic features, while deep supervision aids in learning hierarchical feature maps. Additionally, a parameter-free neuron-based 3D attention mechanism is incorporated into the Mamba module to capture long-range dependencies and enable effective feature selection and fusion. Secondly, a novel contrastive loss function is designed, incorporating cosine distance metrics into the traditional contrastive loss framework. By maximizing cosine decision margins and normalizing, intra-class variance is minimized, and inter-class variance is maximized, ensuring consistency between input infrared image patches and output color image patches. Finally, extensive comparative analysis on common datasets demonstrates that the proposed method outperforms existing state-of-the-art techniques in colorization performance. This research advances infrared image processing and enhances the visual quality of converted images. The code is available at https://github.com/LTTdouble/CCLGAN .}, archive = {J_NEUCOM}, author = {Tingting Liu and Yujue Cai and Guiping Chen and Hongguang Wei and Junqi Bai and Yuan Liu and Xiubao Sui and Qian Chen}, doi = {10.1016/j.neucom.2025.130713}, journal = {Neurocomputing}, month = {10}, pages = {130713}, shortjournal = {Neurocomputing}, title = {Adversarial network for unsupervised infrared image colorization based on full-scale feature fusion and cosine contrastive learning}, volume = {649}, year = {2025}, }
• (2025). A survey of RWKV. NEUCOM, 649, 130711. (www)

  @article{ , abstract = {The Receptance Weighted Key Value (RWKV) model offers a novel alternative to the Transformer architecture, merging the benefits of recurrent and attention-based systems. Unlike conventional Transformers, which depend heavily on self-attention, RWKV adeptly captures long-range dependencies with minimal computational demands. By utilizing a recurrent framework, RWKV addresses some computational inefficiencies found in Transformers, particularly in tasks with long sequences. RWKV has recently drawn considerable attention for its robust performance across multiple domains. Despite its growing popularity, no systematic review of the RWKV model exists. This paper seeks to fill this gap as the first comprehensive review of the RWKV architecture, its core principles, and its varied applications, such as natural language generation, natural language understanding, and computer vision. We assess how RWKV compares to traditional Transformer models, highlighting its capability to manage long sequences efficiently and lower computational costs. Furthermore, we explore the challenges RWKV encounters and propose potential directions for future research and advancement. We consistently maintain the related open-source materials at: https://github.com/MLGroupJLU/RWKV-Survey .}, archive = {J_NEUCOM}, author = {Zhiyuan Li and Tingyu Xia and Yi Chang and Yuan Wu}, doi = {10.1016/j.neucom.2025.130711}, journal = {Neurocomputing}, month = {10}, pages = {130711}, shortjournal = {Neurocomputing}, title = {A survey of RWKV}, volume = {649}, year = {2025}, }
• (2025). Simple yet effective: An explicit query-based relation learner for human–object-interaction detection. NEUCOM, 649, 130709. (www)

  @article{ , abstract = {Human-object-interaction (HOI) detection is a task that requires the model to detect human-object pairs and recognize the interactions between them. Recent advances in Detection Transformers have had a great impact on the HOI detection research field. However, Transformer-based HOI detection methods only exploit randomized queries for decoding HOI triplets. For the HOI detection model using visual and other modalities of features as explicit queries, the computation complexity of the model can be significantly increased compared to randomized implicit queries. In this paper, a simplified approach for constructing explicit queries with only visual features of instances combined with a set prediction training strategy, used in Transformer-based HOI detection models, is proposed. This paper also proposes a novel method to extract and reduce the dimension of instance visual features, which can benefit the learning process of HOIs. The model proposed, in this paper, achieves state-of-the-art performance on the HICO-Det benchmark and shows a competitive training efficiency compared to other HOI detection methods.}, archive = {J_NEUCOM}, author = {Tianlun Luo and Qiao Yuan and Boxuan Zhu and Steven Guan and Rui Yang and Jeremy S. Smith and Eng Gee Lim}, doi = {10.1016/j.neucom.2025.130709}, journal = {Neurocomputing}, month = {10}, pages = {130709}, shortjournal = {Neurocomputing}, title = {Simple yet effective: An explicit query-based relation learner for human–object-interaction detection}, volume = {649}, year = {2025}, }
• (2025). Multi-granularity knowledge embedding for multi-label herbal medicine recognition. NEUCOM, 649, 130707. (www)

  @article{ , abstract = {Accurate identification of herbs in medical images is crucial for clinical diagnosis and treatment. Since a single image usually contains multiple herbs, multi-label learning methods are crucial. However, Chinese medicine recognition faces unique challenges, especially the occlusion problem of herbs and the high similarity in the appearance of certain herbs. This paper proposes a multi-granularity knowledge embedding framework that enhances multi-label recognition by exploring the correlation between visual features of herbal images and prescription knowledge. The framework extracts multi-granularity visual features from the images and incorporates the co-occurrence relationships of herbs within prescriptions as prompts in the recognition process. The potential associations of herbs within prescriptions in prompts are treated as prior knowledge providing crucial contextual cues that further guide the recognition of visual features. By integrating prompt information with visual features, the model is able to perform multi-label prediction more accurately, which significantly improves the accuracy and robustness of traditional Chinese herbal medicine recognition. The experimental results demonstrate that the proposed method outperforms the state-of-the-art methods in terms of recognition accuracy and exhibits strong scalability, which can be generalized to other multi-label learning tasks.}, archive = {J_NEUCOM}, author = {Qi Shen and Canguang Ruan and Liu Yang}, doi = {10.1016/j.neucom.2025.130707}, journal = {Neurocomputing}, month = {10}, pages = {130707}, shortjournal = {Neurocomputing}, title = {Multi-granularity knowledge embedding for multi-label herbal medicine recognition}, volume = {649}, year = {2025}, }
• (2025). Alleviating subgraph-induced oversmoothing in link prediction via coarse graining. NEUCOM, 649, 130666. (www)

  @article{ , abstract = {State-of-the-art link prediction methods often rely on extracting an enclosing subgraph for each target link and subsequently encoding these subgraphs into a Graph Neural Network (GNN). However, a fundamental challenge for extracting subgraphs is that statistically, high-degree nodes appear more frequently in subgraphs for different target links, yet the number of these high-degree nodes is small. Therefore, enclosing subgraphs for different target links tend to be quite similar, resulting in similar receptive fields for the GNN. Traditional methods, such as residual blocks or multiple hops, cannot solve this issue. We propose a simple coarse-graining-based subgraph representation for GNN. For each local subgraph, it merges nodes that are most strongly correlated in the correlation matrix, which ensures the GNN obtains diverse and informative receptive fields. This effectively mitigates the issue of over-smoothing. Additionally, it reduces the GNN’s training time by decreasing subgraph size, thereby enhancing the scalability of GNN for link prediction on large-scale graphs. Our experiments on widely used benchmark datasets highlight the adverse impact of high-degree nodes on LP tasks. The experiments also demonstrate the effectiveness of our proposed method in addressing this challenge.}, archive = {J_NEUCOM}, author = {Xudong Zhu and Dong Hao and Ziqin Gao and Liming Pan}, doi = {10.1016/j.neucom.2025.130666}, journal = {Neurocomputing}, month = {10}, pages = {130666}, shortjournal = {Neurocomputing}, title = {Alleviating subgraph-induced oversmoothing in link prediction via coarse graining}, volume = {649}, year = {2025}, }
• (2025). LDM-KGC: A low-dimensional knowledge graph completion model based on multi-head attention mechanism. NEUCOM, 649, 130665. (www)

  @article{ , abstract = {Existing Transformer-based knowledge graph completion methods often rely on high-dimensional embeddings to achieve competitive performance, which to some extent limits their scalability on large-scale knowledge graphs. To address this challenge, the LDM-KGC model based on the multi-head attention mechanism is proposed. By combining QKV-layer and Update-layer, LDM-KGC can not only learn rich information but also reduce information loss during training, thereby achieving superior embedding representations in low-dimensional spaces. Specifically, QKV-layer utilizes the multi-head attention mechanism to effectively capture interactions between entities and relations, while Update-layer further refines the resulting embeddings. Experimental results on the FB15k-237 and WN18RR datasets demonstrate that LDM-KGC outperforms 14 baseline models, significantly improving mean reciprocal rank (MRR) by 12.4 percentage points and 24.4 percentage points over the worst baseline, respectively. Notably, LDM-KGC achieves MRR of 36.5%, Hits@1 of 27.1%, Hits@3 of 40.2%, and Hits@10 of 55.2% on the FB15k-237 dataset. Furthermore, LDM-KGC reaches a Hits@10 score of 65.2% on the NELL-995 dataset. These results underscore the effectiveness of LDM-KGC in generating low-dimensional embeddings, thereby offering a scalable solution for large-scale knowledge graph completion.}, archive = {J_NEUCOM}, author = {Chaoqun Zhang and Bingjie Qiu and Weidong Tang and Bicheng Liang and Danyang Cui and Haisheng Luo and Qiming Chen}, doi = {10.1016/j.neucom.2025.130665}, journal = {Neurocomputing}, month = {10}, pages = {130665}, shortjournal = {Neurocomputing}, title = {LDM-KGC: A low-dimensional knowledge graph completion model based on multi-head attention mechanism}, volume = {649}, year = {2025}, }
• (2025). Adaptive decoupled strategy for robust and efficient low-rank matrix decomposition. NEUCOM, 649, 130656. (www)

  @article{ , abstract = {Low-rank matrix decomposition with missing values is vital in computer vision and pattern recognition, yet it presents significant challenges. This problem can be viewed as a separable nonlinear optimization, but traditional methods often fail to account for the coupling between parameters and the impact of solution properties on visual reconstruction. We observe that such separable nonlinear problems frequently encounters narrow ravines filled with sharp minima. Classic alternating optimization methods, the Wiberg algorithm and its variants tend to linger in these regions, converging to sharp minima, thereby slowing convergence and degrading reconstruction quality. This promotes us to introduce the Adaptive Decoupled Variable Projection algorithm (ADVP), which can adaptively handle the coupling of parameters, significantly accelerate the convergence rate, and dynamically adjust the parameter search subspace, helping algorithms avoid these ravines towards flatter local minima. These flat minima exhibit robustness against missing data, noise, and outliers, enhancing the quality of visual reconstruction. Extensive experiments on synthetic and real datasets have validated the efficiency of ADVP and its superior performance in visual reconstruction.}, archive = {J_NEUCOM}, author = {Min Gan and Peng Xue and Fan Zhang and Xiang-Xiang Su and Xin Lin and Guang-Yong Chen}, doi = {10.1016/j.neucom.2025.130656}, journal = {Neurocomputing}, month = {10}, pages = {130656}, shortjournal = {Neurocomputing}, title = {Adaptive decoupled strategy for robust and efficient low-rank matrix decomposition}, volume = {649}, year = {2025}, }
• (2025). Local spatial self-attention based deep network for meteorological data downscaling. NEUCOM, 649, 130653. (www)

  @article{ , abstract = {High-resolution meteorological data are essential for simulation and decision-making in weather-sensitive industries such as agriculture and forestry. However, existing meteorological products typically have low spatial resolution (coarser than 0.1°), making it difficult to capture the fine-grained spatial distribution of meteorological variables. Most existing deep learning-based downscaling methods treat the task as an image super-resolution problem, overlooking key characteristics of meteorological data, such as multi-scale local spatial correlation, local–global spatial dependency, and the complex relationship between terrain and meteorological fields, thus limiting modeling accuracy. To address this issue, this paper proposes a deep neural network based on local spatial self-attention, LSSANet, for the spatial downscaling of meteorological data. Specifically, the Local Spatial Self-Attention Module (LSAM) is proposed to capture local–global spatial correlations of meteorological fields. The Multi-scale Dynamic Aggregation Module (MDAM) is introduced to handle multi-scale local spatial dependencies. Furthermore, an elevation embedding layer and a two-stage training strategy are developed to integrate the relationship between terrain and the meteorological field. Experimental results show that LSSANet achieves superior performance compared to traditional and state-of-the-art methods. In the 4 × downscaling task, LSSANet reduces MAE by 5.1%–75.8%; in the 8 × task, by 4.3%–59.7%; and in the 16 × task, by 1.9%–53.4%. Engineering application experiments further demonstrate that the proposed method can generate high-resolution future meteorological forecasts based on the GFS product. These results indicate that LSSANet can accurately reconstruct or predict high-resolution meteorological fields in specific regions, providing valuable support for planning and decision-making in meteorology-sensitive industries.}, archive = {J_NEUCOM}, author = {Sheng Gao and Lianlei Lin and Zongwei Zhang and Junkai Wang and Hanqing Zhao and Hangyi Yu}, doi = {10.1016/j.neucom.2025.130653}, journal = {Neurocomputing}, month = {10}, pages = {130653}, shortjournal = {Neurocomputing}, title = {Local spatial self-attention based deep network for meteorological data downscaling}, volume = {649}, year = {2025}, }
• (2025). Causally-guided graph mamba for detecting socially abnormal vehicle trajectories. NEUCOM, 649, 130649. (www)

  @article{ , abstract = {With the rapid development of autonomous driving and internet of things (IoT) technologies, vehicle trajectory anomaly detection becomes one of the important tasks in road monitoring to help traffic data collection and intelligent traffic management. Most methods use simple vehicle interaction with static information to accomplish this task. In contrast, long sequences of trajectory spatio-temporal information can better describe the state of vehicle traveling. In this study, we introduce a trajectory anomaly detection pipeline for spatio-temporal graph modeling to address the challenges in this task. First, we propose a causally guided dynamic graph representation (CGDG) to efficiently model key interactions in complex trajectories. This bootstrapping and decoupling mechanism lays the foundation for our subsequent process. Subsequently, we parse the dynamic evolution process in the spatiotemporal graph by using spatio-tempora feature fusion Mamba (STFMamba) as an encoder and decoder. The anomaly results are generated by reconstruction probabilities. To address the challenge of scarce and expensive datasets in trajectory anomaly detection, we propose the TRAREAL benchmark dataset supplemented with various anomalous event scenarios for experiments. Our method performs well in the evaluation of natural trajectory benchmark datasets. The source codes are available at https://github.com/yongfengB/DATMamba .}, archive = {J_NEUCOM}, author = {Yongfeng Bu and Haoxiang Liang and Huansheng Song and Shijie Sun and Zhaoyang Zhang}, doi = {10.1016/j.neucom.2025.130649}, journal = {Neurocomputing}, month = {10}, pages = {130649}, shortjournal = {Neurocomputing}, title = {Causally-guided graph mamba for detecting socially abnormal vehicle trajectories}, volume = {649}, year = {2025}, }
• (2025). Low redundancy cell-based neural architecture search for large convolutional neural networks. NEUCOM, 649, 130644. (www)

  @article{ , abstract = {The cell-based search space is one of the main paradigms in Neural Architecture Search (NAS). However, the current research on this search space tends to optimize on small-size models, and the performance improvement of NAS might be stuck in a bottleneck. This situation has led to a growing performance gap between NAS and hand-designed models in recent years. In this paper, we focus on how to effectively expand the cell-based search space and proposes Low redundancy Cell-based Neural Architecture Search for Large Convolutional neural networks ( L C 2 N A S ), a gradient-based NAS method to search large-scale convolutional models with better performance based on low redundant cell search space. Specifically, a cell-based search space with low redundancy and large kernel is designed. Then train and sample a super network under computational constraints. Finally the network structure is optimized by gradient-based search. Experimental results show that the performance of the proposed search method is comparable to the popular hand-designed models in recent years at different scales. Moreover, LC-NASNet-B achieves an 83.7% classification accuracy on the ImageNet-1k dataset with 86.2M parameters, surpassing previous NAS methods and comparable to the most prominent hand-designed models.}, archive = {J_NEUCOM}, author = {Libin Hou and Linyuan Wang and Senbao Hou and Tianyuan Liu and Shuxiao Ma and Jian Chen and Bin Yan}, doi = {10.1016/j.neucom.2025.130644}, journal = {Neurocomputing}, month = {10}, pages = {130644}, shortjournal = {Neurocomputing}, title = {Low redundancy cell-based neural architecture search for large convolutional neural networks}, volume = {649}, year = {2025}, }
• (2025). Cross-modal interaction and multi-view misalignment network for multimodal sarcasm detection. NEUCOM, 649, 130640. (www)

  @article{ , abstract = {Sarcasm often relies on semantic or emotional misalignment between text and images to convey deeper meaning. With the increasing variety of modalities on social media, multimodal sarcasm detection (MSD) has become a challenging task. However, existing multimodal methods still have limitations in heterogeneous modality fusion, text-image sentiment modeling, and modality misalignment detection, making it difficult to fully utilize multimodal information for cross-modal interaction. Therefore, this study proposes the Cross-modal Interaction and Multi-view Misalignment Network for Multimodal Sarcasm Detection (CIMMN) to enhance modality fusion capability and measure cross-modal inconsistencies. CIMMN employs a cross-modal co-attention mechanism with intermediary features to extract global correlation features between text and images, facilitating balanced modality interaction while reducing the computational complexity of large-scale processing. Additionally, a specially designed prompt and large language models (LLMs) are used to convert dispersed visual information into high-quality structured image captions, providing a uniform processing format for heterogeneous modalities. A graph attention network is further applied to extract deep sentiment features from text and images, addressing the shortcomings of existing research in image sentiment extraction. At the same time, a distance measurement method based on optimal transport theory is adopted to compute semantic deviations and emotional conflicts between modalities from three different perspectives, allowing for a comprehensive assessment of modality misalignment features. Finally, by integrating global correlation features, deep sentiment features, and misalignment features, CIMMN captures comprehensive multimodal sarcasm cues, enabling accurate sarcasm detection. Experimental comparisons on three public datasets demonstrate that CIMMN outperforms existing state-of-the-art methods, verifying its effectiveness in multimodal sarcasm detection tasks.}, archive = {J_NEUCOM}, author = {Bengong Yu and Chenyue Li and Zhonghao Xi and Haoyu Wang and Shuping Zhao}, doi = {10.1016/j.neucom.2025.130640}, journal = {Neurocomputing}, month = {10}, pages = {130640}, shortjournal = {Neurocomputing}, title = {Cross-modal interaction and multi-view misalignment network for multimodal sarcasm detection}, volume = {649}, year = {2025}, }
• (2025). Imperceptible diffusion modification for facial privacy protection. NEUCOM, 649, 130614. (www)

  @article{ , abstract = {The remarkable success of Face Recognition (FR) systems has raised significant privacy concerns, particularly the potential for unauthorized tracking of individuals on social media. To mitigate these risks, various adversarial attack methods have been proposed to protect individuals from unauthorized identification by FR systems. In this work, we propose a novel A dversarial F ace generation method via D iffusion M odification (AFDM), designed for imperceptible and transferable attacks against FR systems. The AFDM operates in the diffusion latent space of the face image, leveraging DDIM inversion and the DDIM process to achieve precise modifications. To ensure high visual fidelity, we propose a self-attention structure loss and a perceptual loss, effectively constraining the extent of diffusion-based alterations. Additionally, to enhance transferability, we introduce a face attribute loss that aligns the attribute embeddings of the protected face with those of the target face. Extensive experiments demonstrate that the proposed AFDM achieves state-of-the-art performance in attacking various FR models, with an average Protect Success Rate (PSR) of 93.32% on the CeleA-HQ and LADN datasets, while maintaining a more natural-looking protected face image with an FID score of 22.96. Through our systematic analysis, we justify the efficacy and generality of the proposed method.}, archive = {J_NEUCOM}, author = {Jia Li and Junhao Dong and Jianhuang Lai and Xiaohua Xie}, doi = {10.1016/j.neucom.2025.130614}, journal = {Neurocomputing}, month = {10}, pages = {130614}, shortjournal = {Neurocomputing}, title = {Imperceptible diffusion modification for facial privacy protection}, volume = {649}, year = {2025}, }
• (2025). Why divisive normalization works in image segmentation?. NEUCOM, 649, 130569. (www)

  @article{ , abstract = {Scene variability and data diversity are major challenges for image segmentation. Previous work suggested that a biologically motivated computation, the so-called Divisive Normalization, could be useful to deal with image variability, but its effects have not been systematically studied over different data sources and environmental factors. Explanation of the reasons why Divisive Normalization layers lead to enhanced capabilities in segmentation networks is still an open issue. In this work, we respond to the questions of when Divisive Normalization layers are useful for image segmentation, and why. First, we propose systematic domain adaptation experiments based on dissecting the problem according to physically meaningful dimensions to delineate the limits of the gain. Then, we provide quantitative explanations of the special capabilities of Divisive Normalization layers. Results show that neural networks augmented with Divisive Normalization get better segmentation results in a wide range of shifted scenarios and their performance remains more stable with regard to the considered environmental factors and the nature of the image sources. This behavior is understood in two ways: (1) by quantifying the invariance of the responses that incorporate Divisive Normalization, and (2) by illustrating the adaptive nonlinearity of the different layers that depend on the local activity.}, archive = {J_NEUCOM}, author = {Pablo Hernández-Cámara and Jorge Vila-Tomás and Paula Dauden-Oliver and Nuria Alabau-Bosque and Valero Laparra and Jesús Malo}, doi = {10.1016/j.neucom.2025.130569}, journal = {Neurocomputing}, month = {10}, pages = {130569}, shortjournal = {Neurocomputing}, title = {Why divisive normalization works in image segmentation?}, volume = {649}, year = {2025}, }
• (2025). Attributed graph clustering with multi-scale weight-based pairwise coarsening and contrastive learning. NEUCOM, 648, 130796. (www)

  @article{ , abstract = {This study introduces the Multi-Scale Weight-Based Pairwise Coarsening and Contrastive Learning (MPCCL) model, a novel approach for attributed graph clustering that effectively bridges critical gaps in existing methods, including long-range dependency, feature collapse, and information loss. Traditional methods often struggle to capture high-order graph features due to their reliance on low-order attribute information, while contrastive learning techniques face limitations in feature diversity by overemphasizing local neighborhood structures. Similarly, conventional graph coarsening methods, though reducing graph scale, frequently lose fine-grained structural details. MPCCL addresses these challenges through an innovative multi-scale coarsening strategy, which progressively condenses the graph while prioritizing the merging of key edges based on global node similarity to preserve essential structural information. It further introduces a one-to-many contrastive learning paradigm, integrating node embeddings with augmented graph views and cluster centroids to enhance feature diversity, while mitigating feature masking issues caused by the accumulation of high-frequency node weights during multi-scale coarsening. By incorporating a graph reconstruction loss and KL divergence into its self-supervised learning framework, MPCCL ensures cross-scale consistency of node representations. Experimental evaluations reveal that MPCCL achieves a significant improvement in clustering performance, including a remarkable 15.24 % increase in NMI on the ACM dataset and notable robust gains on smaller-scale datasets such as Citeseer, Cora and DBLP. In the large-scale Reuters dataset, it significantly improved by 17.84 %, further validating its advantage in enhancing clustering performance and robustness. These results highlight MPCCL’s potential for application in diverse graph clustering tasks, ranging from social network analysis to bioinformatics and knowledge graph-based data mining. The source code for this study is available at https://github.com/YF-W/MPCCL}, archive = {J_NEUCOM}, author = {Binxiong Li and Yuefei Wang and Binyu Zhao and Heyang Gao and Benhan Yang and Quanzhou Luo and Xue Li and Xu Xiang and Yujie Liu and Huijie Tang}, doi = {10.1016/j.neucom.2025.130796}, journal = {Neurocomputing}, month = {10}, pages = {130796}, shortjournal = {Neurocomputing}, title = {Attributed graph clustering with multi-scale weight-based pairwise coarsening and contrastive learning}, volume = {648}, year = {2025}, }
• (2025). MSHRT-net: Multi-scale hierarchical residual transfer network for image manipulation detection and localization. NEUCOM, 648, 130788. (www)

  @article{ , abstract = {The proliferation of malicious image tampering has triggered a trust crisis in the authenticity of visual content, posing potential risks. However, existing methods have limitations when dealing with complex tampering. With the rapid development of forgery techniques and the increasing stealthiness of tampering methods, these methods are gradually becoming ineffective, struggling to effectively detect and accurately locate the tampered areas in images. To address this issue, we have developed the Multi-Scale Hierarchical Residual Transfer Network (MSHRT-Net), which focuses on edge texture and multi-scale information extraction for efficient image tampering detection and localization. Specifically, the Adaptive Gabor Texture Extractor (AGTE) employs a dual-stream-like structure with edge texture and spatial features extracted in parallel. To enhance the expressiveness of the extracted features, we presented the Multi-Scale Hierarchical Residual Module (MSHRM) as the encoder-decoder layer of the backbone network, which captured global and local information via three parallel branches at distinct scales. Subsequently, the Detail-Preserving Skip Module (DPSM), constructed with skip connections, further improves the network’s feature-capturing capability. Additionally, to address inconsistencies between features at different scales, we designed a Dual-Dimensional Attention Module (DAM), which filtered critical information from coarse feature maps while suppressing irrelevant content. Finally, to address tampering-type imbalance in training data, we proposed a class of loss functions that improved the model’s ability to detect and localize various types of tampering. Extensive experimental validation on multiple datasets demonstrates that our model surpasses previous methods, particularly on the DSO dataset with realistic scenarios (pixel-level AUC and IoU: 0.993, 0.966).}, archive = {J_NEUCOM}, author = {Xiao Yang and Xiuli Chai and Zhihua Gan and Lvchen Cao and Yushu Zhang}, doi = {10.1016/j.neucom.2025.130788}, journal = {Neurocomputing}, month = {10}, pages = {130788}, shortjournal = {Neurocomputing}, title = {MSHRT-net: Multi-scale hierarchical residual transfer network for image manipulation detection and localization}, volume = {648}, year = {2025}, }
• (2025). Multimodal prototypical networks with co-metric fusion for few-shot hyperspectral image classification. NEUCOM, 648, 130782. (www)

  @article{ , abstract = {In the field of Hyperspectral image (HSI) classification, prototype-based network methods have achieved significant research progress. These methods utilize pixel-level information from images to construct central prototypes for each class, providing effective solutions for few-shot learning. However, traditional prototype networks have some inherent flaws; they primarily rely on a single image modality and fail to fully leverage the potential complementarity between different modalities, using only a single modality to generate class prototypes, which limits the model's performance in representing class prototypes and enhancing discriminative capabilities. And subtle inter-class differences are also a challenging task in cross-domain scenarios. To overcome these challenges, this study proposes an innovative Multimodal Prototypical Networks with Co-metric Fusion (MPCF). By integrating prototype information from both image and text modalities, MPCF significantly enhances the performance of few-shot learning. The method not only captures the spectral and spatial features of images to construct image prototypes but also extracts textual features from category descriptions to generate text prototypes. Furthermore, by integrating contrastive learning strategies with the Co-metric fusion mechanism, the method effectively harnesses the information from different modalities. This integration allows for the capture of category information across multiple dimensions, significantly boosting the model's discriminative power among various classes and enhancing its capacity to address few-shot learning scenarios. Experiments conducted on several public benchmark HSI datasets (Indian Pines-84.06 %, Houston-80.41 %, Salinas-92.63 %) demonstrate that MPCF exhibits excellent performance under few-shot and cross-domain conditions, achieving higher classification accuracy and stronger robustness compared to state-of-the-art methods. The related code will be made publicly available at the following URL: https://github.com/AIYAU/MPCF .}, archive = {J_NEUCOM}, author = {Yuhang Li and Jinrong He and Hanchi Liu and Yurong Zhang and Zhaokui Li}, doi = {10.1016/j.neucom.2025.130782}, journal = {Neurocomputing}, month = {10}, pages = {130782}, shortjournal = {Neurocomputing}, title = {Multimodal prototypical networks with co-metric fusion for few-shot hyperspectral image classification}, volume = {648}, year = {2025}, }
• (2025). Digital-twin-based AGV cluster dynamic scheduling for solar cell production workshop using deep reinforcement learning. NEUCOM, 648, 130772. (www)

  @article{ , abstract = {In recent years, the demand for renewable energy sources, notably solar energy, has rapidly increased. As the most essential photovoltaic module, solar cells with high cleanliness and fragility rely on automated guided vehicles (AGVs) for transportation between various processes. However, the solar cell production workshop with massive AGVs has the characteristics of high dynamics, complexity, and uncertainty, which makes the traditional AGV scheduling methods unable to meet the dynamic scheduling requirements. Therefore, this paper proposes a digital-twin-based (DT-based) AGV cluster dynamic scheduling method using deep reinforcement learning (DRL) . Firstly, a DT-based AGV cluster dynamic scheduling framework is constructed, ensuring operational synergy among DT, decision-making model formulation, and real-world application. Secondly, an AGV cluster dynamic scheduling mathematical model that minimizes the average waiting time is established. Thirdly, the problem of AGV cluster dynamic scheduling is transformed into a Markov Decision Process (MDP) with detailed descriptions. Moreover, an improved soft actor-critic (ISAC) DRL algorithm, adding the Softmax function to the actor network and introducing a multi-stage sample selection strategy, is implemented to resolve the established MDP model. Finally, the six cases derived from real-world solar cell production workshops are studied, and the results demonstrate the effectiveness of the proposed methodology.}, archive = {J_NEUCOM}, author = {Zhuo Zhou and Liyun Xu and Yiyang Chen and Liqiang Liao and Zhun Xu}, doi = {10.1016/j.neucom.2025.130772}, journal = {Neurocomputing}, month = {10}, pages = {130772}, shortjournal = {Neurocomputing}, title = {Digital-twin-based AGV cluster dynamic scheduling for solar cell production workshop using deep reinforcement learning}, volume = {648}, year = {2025}, }
• (2025). Emerging trends and strategic opportunities in tiny machine learning: A comprehensive thematic analysis. NEUCOM, 648, 130746. (www)

  @article{ , abstract = {This study comprehensively reviews 779 Scopus-indexed documents, critically analyzing the trends, challenges, and opportunities in Tiny Machine Learning (TinyML). Through thematic analysis, 21 key themes are identified, covering areas such as edge computing, deep learning, IoT integration, microcontroller efficiency, and energy utilization. Unlike previous reviews focusing on specific domains or applications, this study adopts a text mining-based thematic approach to identify cross-sector research patterns and uncover underexplored areas. This positions the review as a broad yet deep mapping of the TinyML landscape. By synthesizing insights from the literature, this research highlights strategic opportunities, future directions, and technological advancements necessary to expand the application of TinyML in resource-constrained environments, neural networks, and real-time systems. The review also identifies key challenges such as balancing accuracy and energy efficiency in low-power devices, optimizing on-device learning, and ensuring data privacy without cloud dependency. In doing so, it outlines actionable directions for future research, including scalable deployment in large-scale IoT systems and application expansion into areas such as UAV security and smart cities. The findings are crucial for advancing AI applications in low-power, embedded systems and contribute to the growing body of knowledge on TinyML.}, archive = {J_NEUCOM}, author = {Juan D. Velasquez and Lorena Cadavid and Carlos J. Franco}, doi = {10.1016/j.neucom.2025.130746}, journal = {Neurocomputing}, month = {10}, pages = {130746}, shortjournal = {Neurocomputing}, title = {Emerging trends and strategic opportunities in tiny machine learning: A comprehensive thematic analysis}, volume = {648}, year = {2025}, }
• (2025). Non-separation method-based finite-time synchronization of discrete-time fractional-order complex-valued neural networks with time-varying delays. NEUCOM, 648, 130732. (www)

  @article{ , abstract = {In this paper, the finite-time synchronization (FTS) of discrete-time fractional-order complex-valued neural networks (FOCVNNs) with time-varying delays is studied using a non-separation method. In the literature, fractional-order Gronwall-type inequalities are typically used to analyze the FTS of fractional-order delayed systems, with the norm being estimated by an increasing function. However, for stable delay systems where the solution norm tends to zero, such estimations are not optimal. To improve the estimate for stable systems, a nabla Caputo difference inequality is first rigorously derived. Next, a delayed complex-valued adaptive nonlinear controller is designed, and a verifiable FTS criterion based on the established inequality is proposed, ensuring that the error norm is estimated by a decreasing function. Finally, numerical simulations are performed to validate the effectiveness of the proposed theoretical results.}, archive = {J_NEUCOM}, author = {Weiheng He and Feifei Du and Shilong Zhang}, doi = {10.1016/j.neucom.2025.130732}, journal = {Neurocomputing}, month = {10}, pages = {130732}, shortjournal = {Neurocomputing}, title = {Non-separation method-based finite-time synchronization of discrete-time fractional-order complex-valued neural networks with time-varying delays}, volume = {648}, year = {2025}, }
• (2025). Oscillations enhance time-series prediction in reservoir computing with feedback. NEUCOM, 648, 130728. (www)

  @article{ , abstract = {Reservoir computing, a machine learning framework used for modeling the brain, can predict temporal data with little observations and minimal computational resources. However, it is difficult to accurately reproduce the long-term target time series because the reservoir system becomes unstable. This predictive capability is required for a wide variety of time-series processing, including predictions of motor timing and chaotic dynamical systems. This study proposes oscillation-driven reservoir computing (ODRC) with feedback, where oscillatory signals are fed into a reservoir network to stabilize the network activity and induce complex reservoir dynamics. The ODRC can reproduce long-term target time series more accurately than conventional reservoir computing methods in a motor timing and chaotic time-series prediction tasks. Furthermore, it generates a time series similar to the target in the unexperienced period, that is, it can learn the abstract generative rules from limited observations. Given these significant improvements made by the simple and computationally inexpensive implementation, the ODRC would serve as a practical model of various time series data. Moreover, we will discuss biological implications of the ODRC, considering it as a model of neural oscillations and their cerebellar processors.}, archive = {J_NEUCOM}, author = {Yuji Kawai and Takashi Morita and Jihoon Park and Minoru Asada}, doi = {10.1016/j.neucom.2025.130728}, journal = {Neurocomputing}, month = {10}, pages = {130728}, shortjournal = {Neurocomputing}, title = {Oscillations enhance time-series prediction in reservoir computing with feedback}, volume = {648}, year = {2025}, }
• (2025). Spike-driven incepformer: A hierarchical spiking transformer with inception-inspired feature learning. NEUCOM, 648, 130727. (www)

  @article{ , abstract = {Designing spike-based attention mechanisms and Transformer architectures has gradually become a hot topic in the field of Spiking Neural Networks (SNNs). Spiking Transformers typically rely on floating-point and integer computations to achieve performance improvements. However, models that maintain Spike-Driven characteristics, while exhibiting lower energy consumption, often suffer from suboptimal performance. This paper proposes an innovative solution to address this trade-off. Firstly, we introduce feature convolution, expanding the receptive field of attention learning through multi-scale feature connections. Secondly, we design a Spike-Driven Feature Attention (SDFA) mechanism, which significantly reduces computational complexity and enhances performance by utilizing feature matrix operations. Thirdly, we integrate the Inception structure into the Spike-Driven Transformer, replacing traditional MLP layers. Finally, we incorporate diverse branch convolutions to mitigate information loss caused by neurons in the final layer. Experimental results demonstrate that the Spike-Driven Incepformer achieves excellent performance while balancing parameter count and computational cost. On the ImageNet-1k dataset, it attains an accuracy of 80.41%, representing the state-of-the-art for Spike-Driven SNNs. These findings provide new insights for designing low-energy, high-performance spiking neural networks and promote the application of spiking Transformers in broader artificial intelligence domains. Code will be available at https://github.com/2ephyrus/SDIncepformer .}, archive = {J_NEUCOM}, author = {Wei Meng and Zeyu Huang and Quan Liu and Mincheng Cai and Kun Chen and Li Ma}, doi = {10.1016/j.neucom.2025.130727}, journal = {Neurocomputing}, month = {10}, pages = {130727}, shortjournal = {Neurocomputing}, title = {Spike-driven incepformer: A hierarchical spiking transformer with inception-inspired feature learning}, volume = {648}, year = {2025}, }
• (2025). Fixed/preassigned-time bipartite synchronization for delayed fractional-order multilayer multi-link signed networks via adaptive aperiodically semi-intermittent control. NEUCOM, 648, 130724. (www)

  @article{ , abstract = {In this paper, the bipartite synchronization (BS) problem of fractional-order multilayer multi-link signed networks (FOMMKSNs) with internal time-varying delay, intra-layer and inter-layer coupling time-varying delays is studied in fixed-time or preassigned-time. Firstly, a mathematical model is established by integrating the fractional-order dynamics of nodes, the edges of multilayer and multi-link networks, and the positive and negative weights, making it more diverse and practical. Secondly, an expanded fixed-time aperiodically intermittent strategy theory is proposed, building on the existing theory, along with an improved estimation of the settling time (ST). Then, utilizing the signed graph theory and Lyapunov method, several sufficient conditions for fixed-time BS (FXTBS) of FOMMKSNs are derived through an adaptive aperiodically semi-intermittent control scheme. Furthermore, based on the FXTBS results, a new controller is designed to study the BS of FOMMKSNs within a specified time, where the ST is independent of any initial values and parameters of the network and the controller. Finally, the effectiveness of the obtained results is verified by two simulation examples.}, archive = {J_NEUCOM}, author = {Qiu Peng and Xiaotang Zhang and Manchun Tan}, doi = {10.1016/j.neucom.2025.130724}, journal = {Neurocomputing}, month = {10}, pages = {130724}, shortjournal = {Neurocomputing}, title = {Fixed/preassigned-time bipartite synchronization for delayed fractional-order multilayer multi-link signed networks via adaptive aperiodically semi-intermittent control}, volume = {648}, year = {2025}, }
• (2025). GSRDR-GAN: Global search result diversification ranking approach based on multi-head self-attention and GAN. NEUCOM, 648, 130723. (www)

  @article{ , abstract = {Search result diversification ranking aims to generate rankings that comprehensively cover multiple subtopics, but existing methods often struggle to balance ranking diversity with relevance and face challenges in modeling document interactions and dealing with limited high-quality training data. While GAN have proven highly successful in fields like computer vision, their application to search result diversification has been limited due to the discrete nature of ranking items and the complex interactions among documents. To address these challenges, we propose GSRDR-GAN, a novel approach that integrates multi-head self-attention with GAN. Our method consists of four key components designed to address the limitations of traditional approaches: the Selected Document State Retriever, the Subtopic Encoder with Multi-head Self-Attention, the Subtopic Decoder with Multi-head Self-Attention, and the Relevance Predictor. First, a self-attention-based feature extraction module is employed to enhance document representations, enabling the model to capture both global and local context effectively. Second, a GAN framework is introduced to improve generalization by generating diverse rankings, mitigating limited high-quality training data. Third, a carefully designed reward function optimizes the trade-off between ranking diversity and relevance, allowing the model to adaptively prioritize these competing objectives during training. Notably, the method improves the generator’s stability and the diversity of search results by reducing training variance, even without pre-trained models. Extensive experiments on the TREC Web Track dataset demonstrate that the proposed GSRDR-GAN method significantly enhances result diversity, achieving relative improvements of 1.7% in α -nDCG, 3.0% in ERR-IA, 3.3% in NRBP, and 0.9% in S-rec over strong baseline methods. Ablation studies and comparative analyses of different reward computation methods further validate the effectiveness of the proposed approach.}, archive = {J_NEUCOM}, author = {Weidong Liu and Jinzhong Li and Shengbo Chen}, doi = {10.1016/j.neucom.2025.130723}, journal = {Neurocomputing}, month = {10}, pages = {130723}, shortjournal = {Neurocomputing}, title = {GSRDR-GAN: Global search result diversification ranking approach based on multi-head self-attention and GAN}, volume = {648}, year = {2025}, }
• (2025). Statistical modeling of univariate multimodal data. NEUCOM, 648, 130719. (www)

  @article{ , abstract = {Unimodality constitutes a key property indicating grouping behavior of the data around a single mode of its density. We propose a method that partitions univariate data into unimodal subsets through recursive splitting around valley points of the data density. For valley point detection, we introduce properties of critical points on the convex hull of the empirical cumulative density function (ecdf) plot that provide indications on the existence of density valleys. Next, we apply a unimodal data modeling approach that provides a statistical model for each obtained unimodal subset in the form of a Uniform Mixture Model (UMM). Consequently, a hierarchical statistical model of the initial dataset is obtained in the form of a mixture of UMMs, named as the Unimodal Mixture Model (UDMM). The proposed method is non-parametric, hyperparameter-free, automatically estimates the number of unimodal subsets and provides accurate statistical models as indicated by experimental results on clustering and density estimation tasks.}, archive = {J_NEUCOM}, author = {Paraskevi Chasani and Aristidis Likas}, doi = {10.1016/j.neucom.2025.130719}, journal = {Neurocomputing}, month = {10}, pages = {130719}, shortjournal = {Neurocomputing}, title = {Statistical modeling of univariate multimodal data}, volume = {648}, year = {2025}, }
• (2025). Heterogeneous multi-agent reinforcement learning for zero-shot scalable collaboration. NEUCOM, 648, 130716. (www)

  @article{ , abstract = {The emergence of multi-agent reinforcement learning (MARL) is significantly transforming various fields like autonomous vehicle networks. However, real-world multi-agent systems typically contain multiple roles, and the scale of these systems dynamically fluctuates. Consequently, in order to achieve zero-shot scalable collaboration, it is essential that strategies for different roles can be updated flexibly according to the scales, which is still a challenge for current MARL frameworks. To address this, we propose a novel MARL framework named S calable and H eterogeneous P roximal P olicy O ptimization (SHPPO) , integrating heterogeneity into parameter-shared PPO-based MARL networks. We first leverage a latent network to learn strategy patterns for each agent adaptively. Second, we introduce a heterogeneous layer to be inserted into decision-making networks, whose parameters are specifically generated by the learned latent variables. Our approach is scalable as all the parameters are shared except for the heterogeneous layer, and gains both inter-individual and temporal heterogeneity, allowing SHPPO to adapt effectively to varying scales. SHPPO exhibits superior performance in classic MARL environments like Starcraft Multi-Agent Challenge (SMAC) and Google Research Football (GRF), showcasing enhanced zero-shot scalability, and offering insights into the learned latent variables’ impact on team performance by visualization.}, archive = {J_NEUCOM}, author = {Xudong Guo and Daming Shi and Junjie Yu and Wenhui Fan}, doi = {10.1016/j.neucom.2025.130716}, journal = {Neurocomputing}, month = {10}, pages = {130716}, shortjournal = {Neurocomputing}, title = {Heterogeneous multi-agent reinforcement learning for zero-shot scalable collaboration}, volume = {648}, year = {2025}, }
• (2025). SC-BSN: Shifted convolutions based blind-spot network for self-supervised image denoising. NEUCOM, 648, 130714. (www)

  @article{ , abstract = {Self-supervised image denoising methods have garnered significant attention recently due to their ability to train solely on noisy images without requiring paired clean-noisy data. However, real-world noise is often spatially correlated, leading to poor performance in self-supervised algorithms that assume pixel-wise independent noise. To address this limitation, we design multi-branch directional shifted operations to create blind spots in different regions, which effectively disrupt noise correlation. Further, the Shifted Convolutions Blind-Spot Network (SC-BSN) is proposed for self-supervised denoising. This network leverages three distinct blind-spot branches with varying shifted distances to effectively balance noise correlation suppression and the preservation of local spatial structures. Finally, we develop the Complementary Random-Replacing Refinement (CR3) to complement denoising results instead of relying on the iterative averaging of R3. The new post-processing technique efficiently retains the details of denoised images. Experimental results demonstrate that SC-BSN outperforms existing self-supervised denoising methods across multiple datasets, achieving superior performance in both visual quality and quantitative metrics.}, archive = {J_NEUCOM}, author = {Guo Yang and Chengyun Song and Minglong Xue and Jian Yu}, doi = {10.1016/j.neucom.2025.130714}, journal = {Neurocomputing}, month = {10}, pages = {130714}, shortjournal = {Neurocomputing}, title = {SC-BSN: Shifted convolutions based blind-spot network for self-supervised image denoising}, volume = {648}, year = {2025}, }
• (2025). LVSA: Lightweight and verifiable secure aggregation for federated learning. NEUCOM, 648, 130712. (www)

  @article{ , abstract = {Federated learning (FL) is a decentralized machine learning paradigm that facilitates collaborative training of global models through the exchange of local gradients while maintaining the confidentiality of raw data. However, recent studies have identified gradient leakage attacks and server-forged aggregation results as significant threats to user data privacy. This issue is especially pronounced in large-scale mobile devices (e.g., tablets, smartphones, and smartwatches), which store highly sensitive user data, making the protection of such data critical. In addition, it is essential to consider the limitations of mobile devices, such as potential power outages, disconnections, and their limited computational and communication resources. To address these challenges, LVSA, a lightweight and verifiable secure aggregation scheme is proposed. LVSA employs a non-interactive masking scheme to protect gradient privacy and allows any user to drop out at any stage. Moreover, a lightweight verification method based on the inner product is introduced, which eliminates complex computations and is more suitable for devices with limited computational resources. Security analysis shows that LVSA not only protects users’ original gradients from being leaked, but also verifies the correctness of the aggregation results. Experimental analysis shows that when the gradient dimension reaches 10 6 , the computation time in LVSA is two orders of magnitude faster than the most advanced existing schemes. In addition, the communication overhead for users is reduced by more than eight times compared to other schemes offering the same functionality.}, archive = {J_NEUCOM}, author = {Gongli Li and Zhe Zhang and Ruiying Du}, doi = {10.1016/j.neucom.2025.130712}, journal = {Neurocomputing}, month = {10}, pages = {130712}, shortjournal = {Neurocomputing}, title = {LVSA: Lightweight and verifiable secure aggregation for federated learning}, volume = {648}, year = {2025}, }
• (2025). Quantum-inspired multimodal fusion with lindblad master equation for sentiment analysis. NEUCOM, 648, 130710. (www)

  @article{ , abstract = {In multimodal sentiment analysis, the primary challenge lies in effectively modeling the complicated interactions among different data modalities. A promising approach is leveraging quantum concepts like superposition and entanglement to enhance the feature representation ability. However, existing quantum-inspired models neglect the intricate nonlinear dynamics inside their multimodal components. Drawing inspiration from the Lindbladian concept in quantum mechanics, we proposes quantum-inspired neural network with the Lindblad Master Equation (LME) and complex-valued LSTM. The proposed model treats each modality as an individual quantum system and superposes them into a mixed quantum system. The trainable LME process models the interaction of this multimodal system with its semantic environment, thereby enhancing the representation of complex interactions among modalities. The efficacy of the proposed model, along with its key components, are validated through extensive experiments on the MVSA and Memotion datasets. The performance are complemented by a comparative analysis that benchmarks the model against state-of-the-art methods, including traditional methods, large language models and quantum-insipred methods. Furthermore, the interpretability of the model is enhanced by quantifying the entanglement entropy of modality combinations using the von-Neumann Entanglement entropy.}, archive = {J_NEUCOM}, author = {Kehuan Yan and Peichao Lai and Yang Yang and Yi Ren and Tuyatsetseg Badarch and Yiwei Chen and Xianghan Zheng}, doi = {10.1016/j.neucom.2025.130710}, journal = {Neurocomputing}, month = {10}, pages = {130710}, shortjournal = {Neurocomputing}, title = {Quantum-inspired multimodal fusion with lindblad master equation for sentiment analysis}, volume = {648}, year = {2025}, }
• (2025). CrossFingerprint: A multi-level fusion network for 3D fingerprint verification. NEUCOM, 648, 130706. (www)

  @article{ , abstract = {Contactless 3D fingerprint gains significant attention in recent years by considering hygiene and template safety. Many previous studies on 3D fingerprint verification relied heavily on minutiae-based approaches, necessitating intricate extracting procedures. To alleviate these problems, this paper presents a two-stage multi-fusion network for fingerprint verification called CrossFingerprint. In the pretraining phase, it achieves a 2D-3D correspondence of fingerprints by maximizing the agreement between point clouds and their corresponding rendered 2D images in the invariant space. In the fine-tuning phase, two weight-shared encoders are designed to directly extract features from 3D fingerprints. Within this framework, the fusion of 2D images and 3D point clouds occurs at both the feature and score levels, resulting in a further improvement in matching accuracy. The experiments on a public database and an in-house database verify that our approach achieves 1.21% in Equal Error Rate (EER) and significantly speeds up feature extraction and matching compared with state-of-the-art approaches based on minutiae.}, archive = {J_NEUCOM}, author = {Feng Liu and Yixin Huang and Yin Li and Qiuheng Wang}, doi = {10.1016/j.neucom.2025.130706}, journal = {Neurocomputing}, month = {10}, pages = {130706}, shortjournal = {Neurocomputing}, title = {CrossFingerprint: A multi-level fusion network for 3D fingerprint verification}, volume = {648}, year = {2025}, }
• (2025). Dataset-adaptive and bias-constrained brain age estimation using pyramid squeeze and excitation transformer. NEUCOM, 648, 130705. (www)

  @article{ , abstract = {Modeling the biological changes of the human brain is crucial for identifying brain-related diseases and health monitoring. The brain age predicted from MRI data is one useful biomarker for quantifying the maturation and ageing process of human brain. However, the acquisition and preprocessing of MRI data can introduce significant variations between datasets, making it essential to develop models with higher accuracy and robustness for cross-dataset evaluation. To achieve this goal, our paper combines the strengths of CNNs and transformers, proposing the Pyramid Squeeze and Excitation Transformer (PSET) as a novel approach for brain age estimation. In the PSET framework, 3D inception blocks function as an advanced CNN module to capture localized features while the self-attention mechanism is integrated with a squeeze-and-excitation module to extract global features across disparate patches. In particular, a dataset-adaptive and bias-constrained (DABC) model training strategy is proposed to improve the robustness for cross-dataset situations and reduce the bias by introducing self-supervised pre-training, meta-learning and novel loss functions. Experiment results on the dataset of 15,437 healthy brain T1-MRIs (MAE=2.342), demonstrated that the proposed method outperforms both classic visual models and existing brain age estimation models, in the aspect of accuracy, generality and unbiasedness. Additionally, through visualization analysis, we identified the key brain regions that play significant roles in brain age estimation, including the occipital lobe. We compared the brain age gap between patients with diseases and healthy control groups, demonstrating the phenomenon of abnormal aging in conditions such as Alzheimer’s disease and mild cognitive impairment.}, archive = {J_NEUCOM}, author = {Yixiao Hu and Haolin Wang and Jiaxiang Cao and Baobin Li}, doi = {10.1016/j.neucom.2025.130705}, journal = {Neurocomputing}, month = {10}, pages = {130705}, shortjournal = {Neurocomputing}, title = {Dataset-adaptive and bias-constrained brain age estimation using pyramid squeeze and excitation transformer}, volume = {648}, year = {2025}, }
• (2025). MLGINet: An efficient network for short-term load forecasting based on multi-scale learning and local–global interactive learning. NEUCOM, 648, 130704. (www)

  @article{ , abstract = {For the power grid to operate in a safe, stable, and dependable manner, accurate short-term load forecasting (STLF) is crucial. As smart grid develops, numerous deep learning models have been proposed for STLF. Some studies devote to decompose the input signal to obtain different signal components by transforming the signal into other domains to mitigate noise and singular value interference. However, existing approaches struggle to perceive subtle signal fluctuations and to integrate short-range and long-range dependencies to ensure the robustness and accuracy of the model. To address these issues, we introduce multi-scale learning to mine multi-granularity semantic knowledge and local–global interactive learning to fuse instructive feature information, thereby improving discriminative representation and optimizing the model’s decision boundary. Specifically, a novel multi-scale sampling (MSS) block including sliding window sampling reorganization and interval sampling reorganization is embedded in backbone to extract discrete signal components at different scales by adjusting sampling field and sampling interval. Accordingly, the load sequence’s dynamic change and periodic trend can aid in the learning of model. Afterwards, the multi-scale interactive attention including intra-signal attention and inter-signal attention is designed to facilitate multi-granularity semantic knowledge interaction within and between multi-scale signal features and filter out valuable local and global information. The designed strategy enhances the model’s ability to resist forgetting and improves its sensitivity to multi-scale perturbations. Experimental results on public datasets collected from Australia, Morocco and China demonstrate that our method outperforms other nine state-of-the-art (SOTA) models by at least on average 35.50% in MAE, MSE, RMSE and MAPE. The model provides a potential idea for load forecasting methods.}, archive = {J_NEUCOM}, author = {Xinzhi Ding and Dayi Li and Geshuai Luo and Wei Huang and Yiping Chen and Qingyong Zhang and Quan Zhou}, doi = {10.1016/j.neucom.2025.130704}, journal = {Neurocomputing}, month = {10}, pages = {130704}, shortjournal = {Neurocomputing}, title = {MLGINet: An efficient network for short-term load forecasting based on multi-scale learning and local–global interactive learning}, volume = {648}, year = {2025}, }
• (2025). Open set domain adaptation via unknown construction and dynamic threshold estimation. NEUCOM, 648, 130668. (www)

  @article{ , abstract = {Open set domain adaptation (OSDA) focuses on adapting a model from the source domain to the target domain when their class distributions differ. The goal is to accurately recognize unknown classes while correctly classifying known classes. Existing research has indicated that adversarial networks can be efficient for unknown class recognition, yet threshold setting remains a challenge. We address this challenge by proposing an OSDA method that uses unknown construction and dynamic threshold estimation (UCDTE), which consists of three stages: unknown construction, dynamic threshold estimation, and distribution alignment. In the first stage, known as unknown construction, pseudo-unknown samples are constructed through feature fusion to learn information regarding the unknown class. In the second stage, dynamic threshold estimation, an unknown discriminator is constructed to further explore different semantic information in the unknown classes, and a dynamic threshold is generated for each target sample by combining it with the domain discriminator. Finally, in the distribution alignment stage, the dynamic threshold adversarial network aligns known samples between the source and target domains while reducing the intra-class gap of unknown samples in the target domain. Experiments conducted on three datasets have demonstrated the robustness and effectiveness of our approach in adapting models across different domains.}, archive = {J_NEUCOM}, author = {Yong Zhang and Qi Zhang and Wenzhe Liu}, doi = {10.1016/j.neucom.2025.130668}, journal = {Neurocomputing}, month = {10}, pages = {130668}, shortjournal = {Neurocomputing}, title = {Open set domain adaptation via unknown construction and dynamic threshold estimation}, volume = {648}, year = {2025}, }
• (2025). Variable admittance control via reinforcement learning: Enhancing UAV interactions across diverse platforms. NEUCOM, 648, 130667. (www)

  @article{ , abstract = {A compliant control model based on Reinforcement Learning (RL) is proposed to allow UAVs (Unmanned Aerial Vehicles) to interact with the environment more effectively and autonomously execute force control tasks. The model learns an optimal admittance adjustment policy for interaction and simultaneously optimizes energy consumption and trajectory tracking of the UAV state. This facilitates stable manipulation of UAVs in unknown environments with interaction forces. Furthermore, the model ensures safe, compliant, and flexible interaction while safeguarding the UAV’s external structures from damage. To assess the model performance, we validated the approach in a simulation environment using a UAV. The model was also tested across different UAV types and various low-level control parameters, demonstrating superior performance in all scenarios. Additionally, we applied this methodology to two distinct UAV types used in real-world applications. Empirical evidence shows that our proposed methods consistently achieve superior results. We also applied similar methodologies to verify 6D interaction in a simulation of a fully actuated platform consisting of three UAVs. Using a high-level training strategy, we evaluated the platform’s ability to slide along a bevel and achieve optimal results in our comparative experiments.}, archive = {J_NEUCOM}, author = {Yuting Feng and Tao Yang and Kaidi Wang and Jiali Sun and Yushu Yu}, doi = {10.1016/j.neucom.2025.130667}, journal = {Neurocomputing}, month = {10}, pages = {130667}, shortjournal = {Neurocomputing}, title = {Variable admittance control via reinforcement learning: Enhancing UAV interactions across diverse platforms}, volume = {648}, year = {2025}, }
• (2025). Learning referee evaluation and assessing action quality from coarse to fine in diving sport. NEUCOM, 648, 130664. (www)

  @article{ , abstract = {Intelligently assessing the quality of athletic performances in sports scenarios remains a fascinating challenge in computer vision. However, unraveling the subtle distinctions between two similar actions in videos and mapping those video representations to quality scores remain significant obstacles. To address these challenges, this work redefines the paradigm of quality score estimation from traditional relative quality score prediction to relative referee score prediction. To make this shift, a cross-feature fusion module rooted in Transformer-based video representation is introduced, to improve pairwise video feature learning in the realm of action quality assessment. Then, a novel contrastive action parsing decoder module generates mid-level representations to effectively connect visual features with detailed quality scores. Both modules utilize cross-attention mechanisms; the former refines the pairwise video features to represent the differences between video pairs, while the latter updates the input queries corresponding to each referee’s evaluation. Finally, to achieve precise quality score estimation, we introduce a meticulous coarse-to-fine decision process, integrating a score classifier and offset regressor. After validation on challenging diving datasets, including MTL-AQA, FineDiving, and TASD-2, the experimental results show that the proposed approach demonstrates effectiveness and feasibility when compared with state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Hong-Ming Qiu and Hong-Bo Zhang and Qing Lei and Jing-Hua Liu and Ji-Xiang Du}, doi = {10.1016/j.neucom.2025.130664}, journal = {Neurocomputing}, month = {10}, pages = {130664}, shortjournal = {Neurocomputing}, title = {Learning referee evaluation and assessing action quality from coarse to fine in diving sport}, volume = {648}, year = {2025}, }
• (2025). FRAKI: K-space deep learning with offline data-universal and online scan-specific priors. NEUCOM, 648, 130663. (www)

  @article{ , abstract = {Sampling a limited number of phase-encoding lines followed by estimating missing lines is an efficient method for shortening scan time of MRI. GeneRalized Autocalibarating Partial Parallel Acquisition (GRAPPA) is such a classical method and is widely used in clinical MRI. As a non-linear method, Robust Artificial-neural-networks for K-space Interpolation (RAKI) is a break-through of GRAPPA in the sense of much higher estimation accuracy. However, RAKI takes much longer estimation time because it requires online training a network for each receiving coil. To overcome the low-efficiency problem, we propose a fast version of RAKI (called fRAKI). fRAKI is roughly 26 times faster and can obtain much higher estimation accuracy compared with RAKI. The high efficiency of fRAKI is due to two properties: (1) A single network is shared to estimate missing lines of all the coils. (2) The online training of fRAKI can converge after a smaller number of iterations. Fast convergency is obtained by using a pre-trained model for initializing learnable parameters. High accuracy benefits from that the pre-train model contains data-universal prior and is also used as a sub-network of fRAKI so that the online training subnetwork can focus on learning scan-specific prior without the risk of overfitting the scan-specific data. Experimental results on the NYU fastMRI knee and brain datasets demonstrate the efficiency and accuracy of the proposed fRAKI.}, archive = {J_NEUCOM}, author = {Jing Pan and Xiaohan Liu and Yiming Liu and Xuebin Sun and Yanwei Pang}, doi = {10.1016/j.neucom.2025.130663}, journal = {Neurocomputing}, month = {10}, pages = {130663}, shortjournal = {Neurocomputing}, title = {FRAKI: K-space deep learning with offline data-universal and online scan-specific priors}, volume = {648}, year = {2025}, }
• (2025). NamPnP: Noise-aware mechanism within plug-and-play framework for image enhancement. NEUCOM, 648, 130662. (www)

  @article{ , abstract = {Low-light Image Enhancement (LIE) strives to improve contrast and restore details for images captured in dark conditions. Most of the previous LIE algorithms were developed based on the Retinex theory, which decomposes the observed image into illumination and reflectance components for pertinent processing. However, most of such methods that address the noise issue of the reflectance component regard the noise as Gaussian noise, which limits the applicability to diverse noise conditions. In this paper, we employ an appropriate noise degradation model in the designed Noise-Aware network to achieve the suppression of various noises in real-world scenarios. Specifically, the designed network leverages the powerful modeling capabilities of the Transformer to better integrate with the proposed degradation model, effectively eliminating noise with unknown distributions in real-world scenarios. Subsequently, it is plugged into the Retinex-based framework to achieve better enhancement performance. Additionally, the proposed method incorporates an edge-guided adaptive weight matrix and an iterative process to regularize the illumination component, resulting in more natural decomposition and further promoting the harmonious integration of illumination and reflectance components. Extensive evaluations on public datasets reveal that the proposed method outperforms existing techniques both qualitatively and quantitatively, demonstrating superior performance in noise removal under dark conditions while preserving finer texture and structural details.}, archive = {J_NEUCOM}, author = {Chenping Zhao and Yan Wang and Guohong Gao and Xixi Jia and Lijun Xu and Jianping Wang and Xiaofang Li}, doi = {10.1016/j.neucom.2025.130662}, journal = {Neurocomputing}, month = {10}, pages = {130662}, shortjournal = {Neurocomputing}, title = {NamPnP: Noise-aware mechanism within plug-and-play framework for image enhancement}, volume = {648}, year = {2025}, }
• (2025). Spatial-frequency dual-domain Kolmogorov–Arnold networks for multimodal medical image fusion. NEUCOM, 648, 130661. (www)

  @article{ , abstract = {Multimodal Medical Image Fusion (MMIF) can significantly enhance the efficiency and accuracy of clinical diagnosis and treatment by integrating medical images from different modalities into a single image with rich information. Recent advancements in Kolmogorov–Arnold Networks (KAN) have demonstrated significant potential in nonlinear fitting, owing to their ability to decompose complex multivariate functions into simpler univariate functions while maintaining high accuracy and interpretability. While most existing methods focus on developing increasingly complex architectures, addressing MMIF from a frequency analysis perspective and leveraging both spatial and frequency domains for interpretable and effective cross-modal fusion through KAN remains an underexplored frontier in prior research. To address this gap, we introduce Spatial-Frequency Dual-domain KAN (SFDKAN), a novel framework for MMIF. Initially, we apply a Hierarchical Wavelet Decomposition strategy to decompose the input modality into different frequency bands and introduce the powerful nonlinear mapping capability of KAN into the sub-bands of varying frequencies. This approach refines unimodal feature extraction and enhances the retention of high-frequency details and structural integrity. Next, we design a Spatial-Frequency Integration KAN (SFIKAN), leveraging complementary information from both spatial and frequency domains to facilitate effective cross-modality feature interaction and fusion. The Spatial KAN effectively focuses on critical regions in the fusion result, while ignoring irrelevant areas and suppressing redundant information. Meanwhile, the Frequency KAN overcomes the local limitations of the spatial domain, effectively handling long-range dependencies and enhancing global feature representation, thereby enabling more efficient cross-modality feature fusion. Extensive experiments on CI-MRI, PET-MRI, and SPECT-MRI datasets demonstrate the superiority of our method over state-of-the-art (SOTA) medical image fusion algorithms in both quantitative metrics and visual quality. The code will be available at https://github.com/xiejiaaax/SFDKAN .}, archive = {J_NEUCOM}, author = {Lewu Lin and Jiaxin Xie and Yingying Wang and Jialing Huang and Rongjin Zhuang and Xiaotong Tu and Xinghao Ding and Na Shen and Qing Lu}, doi = {10.1016/j.neucom.2025.130661}, journal = {Neurocomputing}, month = {10}, pages = {130661}, shortjournal = {Neurocomputing}, title = {Spatial-frequency dual-domain Kolmogorov–Arnold networks for multimodal medical image fusion}, volume = {648}, year = {2025}, }
• (2025). RoSiLC-RS:A robust similar legal case recommender system empowered by large language model and step-back prompting. NEUCOM, 648, 130660. (www)

  @article{ , abstract = {Legal case recommendation systems face significant challenges in the era of Large Language Models (LLMs). While LLMs offer unprecedented opportunities for understanding legal texts, they also introduce risks through AI-generated false legal content. Our survey reveals concerning gaps in public awareness: 41% of respondents incorrectly believe AI never generates false legal information, while only 6% understand potential legal liabilities. To address these issues, we propose RoSiLC-RS, a Robust Similar Legal Case Recommender System that guides LLMs to understand legal concepts at a higher abstraction level. Our system employs four key components: (1) abstraction processing to extract core legal elements, (2) semantic matching to identify similar case features, (3) LLM-powered explanation generation to provide detailed recommendation rationales, enhancing system explainability, and (4) a specialized detection module to identify and filter AI-generated false content. Comprehensive experiments on real-world legal datasets demonstrate that our method significantly outperforms traditional retrieval approaches in precision, relevance, explainability, and resistance to AI-generated content interference. This research provides both technological solutions and insights for the safe application of LLMs in legal domains.}, archive = {J_NEUCOM}, author = {Guohang Zeng and George Tian and Guangquan Zhang and Jie Lu}, doi = {10.1016/j.neucom.2025.130660}, journal = {Neurocomputing}, month = {10}, pages = {130660}, shortjournal = {Neurocomputing}, title = {RoSiLC-RS:A robust similar legal case recommender system empowered by large language model and step-back prompting}, volume = {648}, year = {2025}, }
• (2025). Interest transfer graph convolutional networks for multi-behavior recommendation. NEUCOM, 648, 130659. (www)

  @article{ , abstract = {Multi-behavior recommendation is an effective approach to address the problem of data sparsity. Graph-based multi-behavior recommendation is one of the most promising branches. Most research on multi-behavior recommendation uses Graph Convolutional Networks (GCNs) to model user features, as GCNs can capture high-order relationships and global features between nodes. However, the existing approaches suffer from multi-level user interests and noisy interactions. To address this issue, we propose a novel Interest Transfer Graph Convolutional Networks (ITGCN) for multi-behavior recommendation. Specifically, to model multi-level user interests, we designed a multi-level GCN by removing multi-layer aggregation operations to capture high-order relationships between nodes. In addition, to address the issue of noisy interactions, we propose a multi-behavior interest transfer method. This approach uses similarity-based comparison to reduce the impact of noisy interactions. It makes both target and auxiliary behaviors more robust to noise. At the same time, it transfers interests from auxiliary behaviors into the semantic space of the target behavior. Experiments on four datasets demonstrated the effectiveness of ITGCN.}, archive = {J_NEUCOM}, author = {Minjie Fan and Yongquan Fan and Yajun Du and Xianyong Li}, doi = {10.1016/j.neucom.2025.130659}, journal = {Neurocomputing}, month = {10}, pages = {130659}, shortjournal = {Neurocomputing}, title = {Interest transfer graph convolutional networks for multi-behavior recommendation}, volume = {648}, year = {2025}, }
• (2025). LitANFIS: Literal-aware adaptive neuro-fuzzy inference system to learn conjunctive normal form. NEUCOM, 648, 130658. (www)

  @article{ , abstract = {In this paper, a novel neuro-fuzzy inference system able to learn fuzzy rules based on using both positive and negative predicates (literals) is proposed. We propose a novel fuzzy neuron named Fuzzy Positive–Negative Neuron to decide whether a predicate or its negated form should be considered in a fuzzy rule. Consequently, to exclude a lingual value, there is no need for many fuzzy rules to cover the whole universe of discourse, except that value. Instead, the proposed method can consider its negated lingual value. Moreover, the proposed unit can relax the effect of a variable in forming a fuzzy rule that leads to unstructured fuzzy rules. Since the proposed method can consider literals (positive and negative predicates), it can calculate the Peirce’s arrow which is functionally complete and also forms Conjunctive Normal Form (CNF). Moreover, considering literals improves the interpretability of neuro-fuzzy systems by decreasing the number of fuzzy rules along with making the inference closer to the human’s. Considering the presence of negated predicates, to initialize the rules’ parameters we propose a learning scheme based on minimizing the classification error along with the reconstruction one instead of applying usual clustering approaches. Moreover, dropout regularization is also applied during the training process to extract independent fuzzy rules. The performance and number of fuzzy rules of the proposed method have been compared with state-of-the-art studies, and based on these comparisons, it has the best performance along with the most parsimonious structure.}, archive = {J_NEUCOM}, author = {Mohammad Mahdi Parchamijalal and Armin Salimi-Badr}, doi = {10.1016/j.neucom.2025.130658}, journal = {Neurocomputing}, month = {10}, pages = {130658}, shortjournal = {Neurocomputing}, title = {LitANFIS: Literal-aware adaptive neuro-fuzzy inference system to learn conjunctive normal form}, volume = {648}, year = {2025}, }
• (2025). Robust neuro-adaptive bipartite formation control for nonlinear multi-agent systems subject to markovian switching topologies and additive noise. NEUCOM, 648, 130655. (www)

  @article{ , abstract = {This article explores the problem of robust neuro-adaptive bipartite time-varying formation (BTVF) control for nonlinear multi-agent systems (MASs) under Markovian randomly switching topologies and additive noise. Both leaderless and leader-following cases are considered, with the communication topology modeled as a continuous-time Markov chain. Additionally, the impact of communication noise in realistic scenarios is addressed. To mitigate disturbances and nonlinearities, an adaptive control strategy incorporating neural network (NN) approximation is employed. With the aid of the infinitesimal generator and the indicator function, the expected leaderless and leader-following BTVF can be achieved. Towards the end, the proposed theoretical method is confirmed through two numerical examples.}, archive = {J_NEUCOM}, author = {Xin Li and Jie Wu and Xisheng Zhan and Lingli Cheng and Qingsheng Yang}, doi = {10.1016/j.neucom.2025.130655}, journal = {Neurocomputing}, month = {10}, pages = {130655}, shortjournal = {Neurocomputing}, title = {Robust neuro-adaptive bipartite formation control for nonlinear multi-agent systems subject to markovian switching topologies and additive noise}, volume = {648}, year = {2025}, }
• (2025). Contrastive learning of adaptive social information fusion for recommender systems. NEUCOM, 648, 130654. (www)

  @article{ , abstract = {To address the challenge of sparse user–item interaction data, social recommendation models based on graph neural networks (GNNs) provide a promising solution due to their ability to capture critical structural information, such as connectivity patterns and topological structures of graph-structured data. Recently, contrastive learning models that leverage richer graph structure information to deal with the sparsity problem have gradually attracted the attention of researchers. However, during the graph augmentation phase, random structure perturbations may destroy the original sparse interaction graph structure and introduce noise, resulting in greater variability in the embedding learned by the same node. Besides, the rich social relationships among users can assist in making preference decisions, but the existing work lacks effective exploitation of them. To solve the above issues, in this work, we propose an adaptive information fusion contrastive learning model (AFCL) for social recommendation, which adaptively integrates the information from social graphs and interaction graphs to balance semantic invariance and view hardness (the distinguishability between views). To be specific, we develop a contrastive learning framework for graph neural networks. The framework utilizes a trainable graph generation model and the original graph to generate contrastive views. We also propose an adaptive information fusion approach which integrates information from social graphs into the two contrastive views. As a result, the sparsity of the interaction graph is reduced and the similarity of specific nodes in the two views is enhanced to generate finer-grained embedding representations. Extensive experiments are conducted on five public datasets. The results demonstrate that the proposed AFCL model outperforms the state-of-the-art methods. Our source code is available at https://github.com/MwLcd/AFCL .}, archive = {J_NEUCOM}, author = {Yafang Li and Chenda Li and Caiyan Jia and Baokai Zu and Hongyuan Wang}, doi = {10.1016/j.neucom.2025.130654}, journal = {Neurocomputing}, month = {10}, pages = {130654}, shortjournal = {Neurocomputing}, title = {Contrastive learning of adaptive social information fusion for recommender systems}, volume = {648}, year = {2025}, }
• (2025). FATL: Frozen-feature augmentation transfer learning for few-shot long-tailed sonar image classification. NEUCOM, 648, 130652. (www)

  @article{ , abstract = {Current deep learning-based approaches for sonar image classification have demonstrated effectiveness in scenarios with abundant and balanced samples. However, real-world applications often face challenges such as few-shot learning and long-tailed distribution due to the inherent limitations of underwater sonar image acquisition. In this work, a Frozen-feature Augmentation Transfer Learning (FATL) framework is proposed for few-shot long-tailed sonar image classification. The feature extractor and classifier are independently frozen during training to improve generalization capability. A frozen-feature augmentation module with contrastive learning is introduced to enrich the representational capacity of the feature space in both value and channel dimensions. In addition, a balanced sampling strategy that divides the training set into multiple balanced subsets is employed to mitigate sample bias during classifier training. Extensive experiments conducted on the three public sonar image datasets (KLSG, FLSMDD, and NKSID) indicate that the proposed FATL framework achieves superior performance and outperforms existing approaches in the few-shot long-tailed sonar image classification task.}, archive = {J_NEUCOM}, author = {Zhongyu Bai and Hongli Xu and Qichuan Ding and Xiangyue Zhang}, doi = {10.1016/j.neucom.2025.130652}, journal = {Neurocomputing}, month = {10}, pages = {130652}, shortjournal = {Neurocomputing}, title = {FATL: Frozen-feature augmentation transfer learning for few-shot long-tailed sonar image classification}, volume = {648}, year = {2025}, }
• (2025). A dynamic hybrid expert framework with encoder–decoder interaction for robust image enhancement in train environment perception. NEUCOM, 648, 130650. (www)

  @article{ , abstract = {Train environment perception technology is one of the critical factors in ensuring safe train operations, particularly in challenging conditions such as foggy, rainy weather, and poorly lit environments like tunnels. The clarity of images directly influences the accuracy of obstacle detection and decision-making processes during train operation. However, existing image restoration methods are typically tailored to single scenarios, making them inadequate for the diverse and complex environmental variations encountered during train operations. Most of these methods lack specificity, rendering them ineffective in handling complex textures and fine details, resulting in suboptimal image quality under adverse conditions, often plagued by blurriness and noise interference. To address these challenges, we propose a dynamic hybrid expert image restoration framework specifically designed for train environment perception. This framework integrates multiple expert modules and a dynamic weight generation mechanism, enabling flexible adaptation to various environmental characteristics. Specifically, the framework comprises multiple expert modules, each focusing on distinct feature extraction tasks, thereby enhancing image clarity and detail restoration in challenging conditions such as foggy weather, low-light situations, and tunnels. The system dynamically generates weights based on the input image characteristics, allowing for the seamless integration of features extracted by each expert, which significantly improves image clarity and detail restoration. Additionally, the interaction between encoder–decoder attention mechanisms enhances the fusion of global and local information, ensuring robust image restoration in complex environments. Experimental results demonstrate that our method performs exceptionally well across various train operating conditions, particularly in foggy image enhancement and low-light image restoration in tunnels. Compared to existing methods, our approach achieves superior restoration quality and efficiency. Our method significantly enhances the image processing capabilities of train environment perception systems, providing a robust safeguard for safe train operations.}, archive = {J_NEUCOM}, author = {Xin Liu and Hongping Wang and Linsen Song and Yiwen Zhang and Xiaoxu Zhang and Chunbo Liu and Xiao Shang and Jingru Liu and Yuanting Yang and Xinming Zhang}, doi = {10.1016/j.neucom.2025.130650}, journal = {Neurocomputing}, month = {10}, pages = {130650}, shortjournal = {Neurocomputing}, title = {A dynamic hybrid expert framework with encoder–decoder interaction for robust image enhancement in train environment perception}, volume = {648}, year = {2025}, }
• (2025). Similar concept matters: Prototype analogy for few-shot classification. NEUCOM, 648, 130648. (www)

  @article{ , abstract = {Few-shot classification aims to produce a classifier to recognize novel classes not seen during training with few labeled samples. The unseen classes and scarce samples make few-shot classification truly challenging. Transfer learning has been demonstrated to be an efficient paradigm for few-shot classification problem in recent literature. However, most methods based on transfer learning only utilize the parameters of the pre-trained model, ignoring the base features themselves which can be seen as the semantic concepts learned by the pre-trained model. In this paper, our main innovation lies in highlighting the importance of the learned semantic concepts of base classes. We propose a simple yet effective approach for few-shot classification to explore the pre-trained semantic features of base classes. Our approach innovatively employs prototype analogy inside and outside the few-shot classification task, to perform clustering and to select base features respectively, in an alternate and iterative way. We further design the best arrangement for these two steps. The initial centroids for clustering are constantly optimized by more and more accurate base features which are selected by the clustered novel prototypes from the previous iteration. When the iteration converges, the best semantic base features are selected to complete the prototypes of novel classes. Extensive experiments on four standard datasets and two deep backbones are conducted to demonstrate the effectiveness of our proposed prototype analogy method. Notably, our method requires neither sophisticated transductive algorithm nor additional learnable parameters besides the pre-trained model yet achieving comparable or even state-of-the-art performance on the miniImageNet, tieredImageNet and CIFAR-FS datasets.}, archive = {J_NEUCOM}, author = {Tong Ouyang and Bo Ma and Hao Xu}, doi = {10.1016/j.neucom.2025.130648}, journal = {Neurocomputing}, month = {10}, pages = {130648}, shortjournal = {Neurocomputing}, title = {Similar concept matters: Prototype analogy for few-shot classification}, volume = {648}, year = {2025}, }
• (2025). FastPAD: A fast self-supervised image anomaly detection method based on patch aggregation and discrimination. NEUCOM, 648, 130647. (www)

  @article{ , abstract = {Self-supervised learning has been gradually applied to image anomaly detection tasks because it can alleviate the imbalance of positive and negative samples. However, most of the current self-supervised-based methods still have problems such as synthetic abnormal samples not authentic enough and slow inference speed. Therefore, we proposed a fast self-supervised image anomaly detection method based on patch aggregation and discrimination (FastPAD). Different from the previous CNN-based methods, FastPAD does not use convolution to process the extracted pre-trained features and introduced the idea of image patch to aggregate feature maps, which reduces the overall calculation of the network and ensures that the network has faster inference speed. Meanwhile, an abnormal sample synthesis correction module (ASCM) was proposed to improve the authenticity of the synthetic abnormal samples and reduce the false anomaly knowledge learned by the network. In addition, a fast attention feature fusion module (FAFFM) that can complement the deep and shallow feature information and a more reasonable patch-level adaptive weighted fusion strategy (ADWF) for score maps were proposed to further improve the detection accuracy. On MVTecAD and VisA datasets, the detection accuracy of our proposed method achieved 98.5% and 98.8% average AUC-ROC at pixel-level respectively, 93.9% and 93.5% average AUC-PRO respectively, the inference speed of our proposed method achieved 113 FPS, which has reached the current advanced level and provided a new idea for image anomaly detection.}, archive = {J_NEUCOM}, author = {Yihang Gao and Zhiyan Han and Jian Wang}, doi = {10.1016/j.neucom.2025.130647}, journal = {Neurocomputing}, month = {10}, pages = {130647}, shortjournal = {Neurocomputing}, title = {FastPAD: A fast self-supervised image anomaly detection method based on patch aggregation and discrimination}, volume = {648}, year = {2025}, }
• (2025). Learning hierarchical image feature for efficient image rectification. NEUCOM, 648, 130646. (www)

  @article{ , abstract = {Image stitching methods often use single-homography or multi-homography estimation for alignment, resulting in images with undesirable irregular boundaries. To address this, cropping and image inpainting are the common operations but discard image regions or introduce content that differs from reality. Recently, deep learning-based methods improve the content fidelity of the rectified images, while suffering from distortion, artifacts, and discontinuous deformations between adjacent image regions. In this work, we propose an efficient network based on the transformer (Rectformer) for image rectification. Specifically, we propose the Global and Local Features (GLF) module, which consists of the Hybrid Self-Attention module and Dynamic Convolution module to capture hierarchical image features. We further introduce two auxiliary losses for better image rectification, bidirectional contextual (BC) loss and deformation consistency (DC) loss. The bidirectional contextual loss encourages the model to preserve image local structure information. The loss of deformation consistency improves the network’s geometric recovery and generalization capabilities through a self-supervised learning strategy. Finally, extensive experiments demonstrate that our method outperforms the existing state-of-the-art methods for rotation correction and rectangling.}, archive = {J_NEUCOM}, author = {Nanjun Yuan and Fan Yang and Yuefeng Zhang and Luxia Ai and Wenbing Tao}, doi = {10.1016/j.neucom.2025.130646}, journal = {Neurocomputing}, month = {10}, pages = {130646}, shortjournal = {Neurocomputing}, title = {Learning hierarchical image feature for efficient image rectification}, volume = {648}, year = {2025}, }
• (2025). From honeybee to helicopter: A low-cost dual bionic model for flight control under hazardous situations. NEUCOM, 648, 130645. (www)

  @article{ , abstract = {Ensuring flight safety is the first priority in aviation operations. As the low-altitude transport sector continues to evolve, concerns regarding the safety of helicopters operating at lower altitudes have become increasingly pronounced. In contrast to high-altitude flights, low-altitude operations often encounter various topographical features, such as hills, which heighten the risk of hazardous incidents. To mitigate the effects of pilot spatial disorientation on flight safety, this study introduces an advanced dual bionic vision model, referred to as AVDM2. A simulation experiment was conducted utilizing the Unity platform to evaluate this model. AVDM2 is designed to take over control of the helicopter’s maneuvers in instances of pilot spatial disorientation, thereby maintaining a safe altitude above the ground. In comparison to AVDM and other bioinspired vision-based algorithms, AVDM2 exhibits enhanced efficacy in maintaining a safe distance between the helicopter and the ground. This advantage is particularly salient in high-speed operational scenarios, specifically at velocities of approximately 70 m/s to 80 m/s, as well as during prolonged missions that involve extensive routes. Furthermore, we administered a survey to assess public perceptions regarding the reliability of bionic vision models and deep learning methodologies. The results indicate that bionic vision models characterized by greater interpretability are more likely to engender public trust compared to the ’black-box’ nature of deep learning, despite the latter potentially offering superior performance.}, archive = {J_NEUCOM}, author = {Shiwen Pan and Haodong Zhang and Mingyang Gao and Fengshuo Yan and Hong Cheng and Kun Guo and Gary J.W. Xu and Poly Z.H. Sun and Xiaoru Wanyan and Edmond Q. Wu}, doi = {10.1016/j.neucom.2025.130645}, journal = {Neurocomputing}, month = {10}, pages = {130645}, shortjournal = {Neurocomputing}, title = {From honeybee to helicopter: A low-cost dual bionic model for flight control under hazardous situations}, volume = {648}, year = {2025}, }
• (2025). Neural network-based distributed adaptive fault-tolerant containment control. NEUCOM, 648, 130643. (www)

  @article{ , abstract = {In practical applications, multi-agent systems (MASs) often face challenges arising from incomplete knowledge of system dynamics, and agent actuators may suffer from faults such as partial failures or biased inputs. This paper investigates the fault-tolerant containment control problem for nonlinear MASs subject to actuator faults and proposes a neural network-based control approach. The system model is assumed to involve unknown nonlinearities, and the follower agents may experience actuator faults. Neural networks are employed to approximate the unknown nonlinear dynamics, and adaptive parameters are introduced and updated online based on the system evolution. An adaptive distributed fault-tolerant control protocol is developed by integrating neural network approximations, adaptive parameter adjustments, and relative state errors between neighboring agents. By dynamically tuning the control effort through the adaptive parameters, the proposed protocol effectively compensates for system nonlinearities and ensures the achievement of the containment control objective, even in the presence of actuator faults. Simulation results are presented to demonstrate the effectiveness of the proposed control strategy.}, archive = {J_NEUCOM}, author = {Ziying Fang and Xiaojian Yi and Tao Xu and Xiaoguang Wang}, doi = {10.1016/j.neucom.2025.130643}, journal = {Neurocomputing}, month = {10}, pages = {130643}, shortjournal = {Neurocomputing}, title = {Neural network-based distributed adaptive fault-tolerant containment control}, volume = {648}, year = {2025}, }
• (2025). S2R-CMI: A robust deep reinforcement learning method based on counterfactual estimation and state importance evaluation under additive noise disturbance. NEUCOM, 648, 130642. (www)

  @article{ , abstract = {The development of Deep Reinforcement Learning (DRL) overcomes the limitations of traditional reinforcement learning in discrete spaces, extending its applications to scenarios in continuous spaces. However, disturbance commonly encountered in real-world environments poses significant threats to the performance of DRL algorithms, potentially leading to erroneous decision-making by agents and severe consequences. To address this issue, this paper proposes a Robust Deep Reinforcement Learning (RDRL) method named S2R-CMI, which aims to mitigate the impact of additive noise in the state space on DRL performance without requiring prior knowledge of the noise disturbance. Specifically, a state-based and reward-based conditional mutual information mechanism is designed to dynamically capture state importance and estimate its contribution to rewards. To address the lack of counterfactual data during training, a counterfactual label estimation method is proposed to approximate the counterfactual reward distribution while avoiding local optima during network training. State importance is then evaluated to quantify the impact of disturbance on states. Finally, we validate the proposed method in five scenarios: Cartpole, LunarLander-Discrete, LunarLander-Continuous, Build-Marine, and Half-Cheetah, under state disturbance. Experimental results demonstrate that S2R-CMI significantly enhances the robustness of DRL algorithms. Furthermore, we conduct experiments in some scenarios without state disturbance, and the results indicate that the method also achieves strong performance, further verifying its superiority and generalization capabilities.}, archive = {J_NEUCOM}, author = {Zhenyuan Chen and Zhi Zheng and Wenjun Huang and Xiaomin Lin}, doi = {10.1016/j.neucom.2025.130642}, journal = {Neurocomputing}, month = {10}, pages = {130642}, shortjournal = {Neurocomputing}, title = {S2R-CMI: A robust deep reinforcement learning method based on counterfactual estimation and state importance evaluation under additive noise disturbance}, volume = {648}, year = {2025}, }
• (2025). FS-GNN: Improving fairness in graph neural networks via joint sparsification. NEUCOM, 648, 130641. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) have emerged as powerful tools for analyzing graph-structured data, but their widespread adoption in critical applications is hindered by inherent biases related to sensitive attributes such as gender and race. While existing debiasing approaches typically focus on either modifying input graphs or incorporating fairness constraints into model objectives, we propose Fair Sparse GNN (FS-GNN), a novel framework that simultaneously enhances fairness and efficiency through joint sparsification of both input graphs and model architectures. Our approach iteratively identifies and removes less informative edges from input graphs while pruning redundant weights from the GNN model, guided by carefully designed fairness-aware objective functions. Through extensive experiments on real-world datasets, we demonstrate that FS-GNN achieves superior fairness metrics (reducing Statistical Parity from 7.94 to 0.6) while maintaining competitive prediction accuracy compared to state-of-the-art methods. Additionally, our theoretical analysis reveals distinct fairness implications of graph versus architecture sparsification, providing insights for future fairness-aware GNN designs. The proposed method not only advances fairness in GNNs but also offers substantial computational benefits through reduced model complexity, with FLOPs reductions ranging from 24% to 67%.}, archive = {J_NEUCOM}, author = {Jiaxu Zhao and Tianjin Huang and Shiwei Liu and Jie Yin and Yulong Pei and Meng Fang and Mykola Pechenizkiy}, doi = {10.1016/j.neucom.2025.130641}, journal = {Neurocomputing}, month = {10}, pages = {130641}, shortjournal = {Neurocomputing}, title = {FS-GNN: Improving fairness in graph neural networks via joint sparsification}, volume = {648}, year = {2025}, }
• (2025). Beyond game environments: Evolutionary algorithms with parameter space noise for task-oriented dialogue policy exploration. NEUCOM, 648, 130639. (www)

  @article{ , abstract = {Reinforcement learning (RL) has achieved significant success in task-oriented dialogue (TOD) policy learning. Nevertheless, training dialogue policy through RL faces a critical challenge: insufficient exploration, which leads to the policy getting trapped in local optima. Evolutionary algorithms (EAs) enhance exploration breadth by maintaining and selecting diverse individuals, and they often add parameter space noise among different individuals to simulate mutation, thereby increasing exploration depth. This approach has proven to be an effective method for enhancing RL exploration and has shown promising results in game domains. However, previous research has not analyzed its effectiveness in TOD dialogue policy. Given the substantial differences between gaming contexts and TOD dialogue policy, this paper explores and validates the efficacy of EAs in TOD dialogue policy, investigating the effects of different evolutionary cycles and various noise strategies across different dialogue tasks to determine which combination of evolutionary cycle and noise strategy is most suitable for TOD dialogue policy. Additionally, we propose an adaptive noise evolution method that dynamically adjusts noise scales to improve exploration efficiency. Experiments on the MultiWOZ dataset demonstrate significant performance improvements, achieving state-of-the-art results in both on-policy and off-policy settings.}, archive = {J_NEUCOM}, author = {Qingxin Xiao and Yangyang Zhao and Lingwei Dang and Yun Hao and Le Che and Qingyao Wu}, doi = {10.1016/j.neucom.2025.130639}, journal = {Neurocomputing}, month = {10}, pages = {130639}, shortjournal = {Neurocomputing}, title = {Beyond game environments: Evolutionary algorithms with parameter space noise for task-oriented dialogue policy exploration}, volume = {648}, year = {2025}, }
• (2025). Dynamic task balancing for joint information extraction. NEUCOM, 648, 130638. (www)

  @article{ , abstract = {Joint Information Extraction (IE) aims for joint extraction of various semantic structures such as entities and relations. Most recent joint IE works use static weighting methods by combining task losses with predefined and fixed weights. In this paper, we identify the limitations of the static weighting methods with empirical analysis. We then study the feasibility of applying several dynamic weighting methods for the joint IE problem and evaluate the methods on three benchmark IE datasets in terms of their performance. We find that existing dynamic weighting methods can achieve reasonably good results in a single run, demonstrating their effectiveness and advantages over the static weighting methods. Further, we propose a hybrid dynamic weighting method, Adaptive Weighting for Joint IE (AWIE), based on gradient dynamic task weighting. Experimental results show that our proposed method obtains competitive performance results across datasets cost-effectively with task preference accommodation.}, archive = {J_NEUCOM}, author = {Anran Hao and Shuo Sun and Jian Su and Siu Cheung Hui and Anh Tuan Luu}, doi = {10.1016/j.neucom.2025.130638}, journal = {Neurocomputing}, month = {10}, pages = {130638}, shortjournal = {Neurocomputing}, title = {Dynamic task balancing for joint information extraction}, volume = {648}, year = {2025}, }
• (2025). Graph neural architecture for dynamic hybrid flowshop: Addressing stochastic events and uncertain processing sequences. NEUCOM, 648, 130636. (www)

  @article{ , abstract = {Dynamic hybrid flow shop scheduling (DHFSP) presents substantial difficulties in attaining optimal production efficiency. These challenges arise particularly when addressing indeterminate job sequences and stochastic events. Current dynamic scheduling methodologies employing deep reinforcement learning (DRL) predominantly target static-scale problems while exhibiting limited scalability to evolving complexities in industrial production systems. To enable responsive shop-floor decisions, this work introduces a hierarchical interaction attention-driven multi-graph network (HIAMG) framework engineered for DHFSP with uncertain job sequences and stochastic events. The proposed architecture implements graph-structured scenario encoding to address neural network dimension constraints. It is coupled with a novel graph embedding paradigm that hierarchically models job-machine interdependencies. Through explicit incorporation of stochastic events in scheduling simulations, our approach achieves enhanced operational fidelity reflecting real manufacturing settings. Integrating attention-based mechanisms further empowers the model to dynamically prioritize critical scheduling parameters and self-tune to environmental fluctuations. Comprehensive simulations across diverse operational regimes demonstrate that HIAMG surpasses conventional scheduling benchmarks in both performance metrics and configuration adaptability.}, archive = {J_NEUCOM}, author = {Weixiang Xu and Xiaochuan Luo ( Co-ordinator ) and Yejian Zhao and Yulin Zhang}, doi = {10.1016/j.neucom.2025.130636}, journal = {Neurocomputing}, month = {10}, pages = {130636}, shortjournal = {Neurocomputing}, title = {Graph neural architecture for dynamic hybrid flowshop: Addressing stochastic events and uncertain processing sequences}, volume = {648}, year = {2025}, }
• (2025). EDSDF: An efficient deep supervised distillation framework for medical image segmentation. NEUCOM, 648, 130635. (www)

  @article{ , abstract = {Significant progress has been made in medical image segmentation, but it still faces persistent challenges: inefficient multi-scale feature fusion that neglects fine-grained local details; attention to irrelevant regions causing computational waste; and boundary ambiguity exacerbated by upsampling artifacts. We, therefore, propose the EDSDF to address these issues through three key innovations: (1) Multi-Scale Grouped Attention Gate (MSGAG), which leverages multi-scale gating mechanisms to suppress redundant features and enhance critical feature representation; A hierarchical feature fusion architecture comprising: Efficient Multi-Scale Fusion Block (EMSFB), lightweight multi-scale fusion using depthwise separable convolutions; Efficient Frequency Boundary Extraction Block (EFBEB), which utilizes frequency domain information to improve boundary prediction. (2) Knowledge-Aligned Multi-Scale Distillation (KAMSD): Multi-stage alignment of encoder–decoder features between teacher-student models. (3) Universal lightweight architecture: Compatible with both CNN and Transformer encoders achieving 88% parameter reduction(compared to MADGNet). When tested on four publicly available medical image segmentation datasets, the proposed EDSDF achieves state-of-the-art performance, with a mean DSC of 84.65%. Furthermore, ablation studies confirm the contributions of each component, with KAMSD playing a pivotal role in enhancing segmentation performance for the lightweight student model. Consequently, the compatibility of the proposed framework with diverse encoders enables deployment across clinical scenarios without sacrificing accuracy. Our EDSDF code is available in https://github.com/lzn-ai/EDSDF .}, archive = {J_NEUCOM}, author = {Xuefang Chen and Zhining Li and Suigu Tang and Jiayao Li and Huazhu Liu and Jian Nong}, doi = {10.1016/j.neucom.2025.130635}, journal = {Neurocomputing}, month = {10}, pages = {130635}, shortjournal = {Neurocomputing}, title = {EDSDF: An efficient deep supervised distillation framework for medical image segmentation}, volume = {648}, year = {2025}, }
• (2025). Neural adaptive delay differential equations. NEUCOM, 648, 130634. (www)

  @article{ , abstract = {Continuous-depth neural networks, such as neural ordinary differential equations (NODEs), have garnered significant interest in recent years owing to their ability to bridge deep neural networks with dynamical systems. This study introduced a new type of continuous-depth neural network called neural adaptive delay differential equations (NADDEs). Unlike recently proposed neural delay differential equations (NDDEs) that require a fixed delay, NADDEs utilize a learnable, adaptive delay. Specifically, NADDEs reformulate the learning process as a delay-free optimal control problem and leverage the calculus of variations to derive their learning algorithms. This approach enables the model to autonomously identify suitable delays for given tasks, thereby establishing more flexible temporal dependencies to optimize the utilization of historical representations. The proposed NADDEs can reconstruct dynamical systems with time-delay effects by learning true delays from data, a capability beyond both NODEs and NDDEs, and achieve superior performance on concentric and image-classification datasets, including MNIST, CIFAR-10, and SVHN.}, archive = {J_NEUCOM}, author = {Chao Zhou and Qieshi Zhang and Jun Cheng}, doi = {10.1016/j.neucom.2025.130634}, journal = {Neurocomputing}, month = {10}, pages = {130634}, shortjournal = {Neurocomputing}, title = {Neural adaptive delay differential equations}, volume = {648}, year = {2025}, }
• (2025). Knowledge-based hyper-parameter adaptation of multi-stage differential evolution by deep reinforcement learning. NEUCOM, 648, 130633. (www)

  @article{ , abstract = {Differential evolution (DE) is a prominent algorithm in evolutionary computation, with adaptive control mechanisms for its operators and parameters being a critical research focus due to their impact on performance. Existing studies often rely on trial-and-error methods or deep reinforcement learning (DRL) for per-generation adaptive control, yet they inadequately explore adaptive hyper-parameter tuning across different stages of the evolution process. To address this limitation, this paper presents a knowledge-based framework named DRL-HP-* for multi-stage DE hyper-parameter adaptation using DRL. The framework divides the algorithm’s search procedure into multiple equal stages, where a DRL agent determines hyper-parameters in each stage based on five types of states that characterize the evolutionary process. A novel reward function is designed to comprehensively train the agent across all training functions, integrating the performance of the backbone algorithm. This approach results in the development of three new algorithms (DRL-HP-jSO, DRL-HP-LSHADE-RSP, and DRL-HP-EjSO). Experimental evaluations on the CEC’18 benchmark suite demonstrate that the proposed algorithms outperform eight state-of-the-art methods, demonstrating superior optimization performance. Further extensive experiments validate the effectiveness of the designed reward function and the framework’s scalability and robustness, highlighting its contribution to enabling stage-wise adaptive hyper-parameter control.}, archive = {J_NEUCOM}, author = {Mingzhang Han and Mingjie Fan and Xinchao Zhao and Lingjuan Ye}, doi = {10.1016/j.neucom.2025.130633}, journal = {Neurocomputing}, month = {10}, pages = {130633}, shortjournal = {Neurocomputing}, title = {Knowledge-based hyper-parameter adaptation of multi-stage differential evolution by deep reinforcement learning}, volume = {648}, year = {2025}, }
• (2025). MTR-MSE: Motion-text retrieval method based on motion semantics expansion. NEUCOM, 648, 130632. (www)

  @article{ , abstract = {The motion-text cross-retrieval task aims to bridge the motion and text spaces, enabling mutual retrieval between motion and language. However, existing methods suffer from limited feature extraction due to both insufficient data and inadequate feature extraction techniques, which restrict retrieval accuracy and semantic richness. To address this, we propose a Motion-Text Retrieval Method Based on Motion Semantics Expansion (MTR-MSE). We design specialized motion and text encoders to create a comprehensive shared feature space. Furthermore, recognizing the limitations of overly simplistic textual descriptions in existing datasets, we enhance motion semantics using large language models to generate more detailed and varied descriptions, thereby improving motion understanding. Experimental results demonstrate that our method achieves state-of-the-art performance, validating its effectiveness in addressing the challenges of cross-modal motion-text retrieval.}, archive = {J_NEUCOM}, author = {Daoliang Xu and Tianyou Zheng and Yang Zhang and Xiaodong Yang and Weiwei Fu}, doi = {10.1016/j.neucom.2025.130632}, journal = {Neurocomputing}, month = {10}, pages = {130632}, shortjournal = {Neurocomputing}, title = {MTR-MSE: Motion-text retrieval method based on motion semantics expansion}, volume = {648}, year = {2025}, }
• (2025). Tuning multi-objective multi-population evolutionary models for high-dimensional problems: The case of the migration process. NEUCOM, 648, 130631. (www)

  @article{ , abstract = {Multi-objective multi-population evolutionary procedures have become one of the most outstanding metaheuristics for solving problems characterized by the curse of dimensionality . A critical aspect of these models is the migration process, defined as the exchange of individuals between subpopulations every few iterations or generations, which has typically been adjusted according to a set of guidelines proposed more than 20 years ago, when the capacity to deal with problems was significantly less than it is today. However, the constant increase in computational power has made it possible to tackle today’s complex real-world problems of great interest more plausibly, but with larger populations than before. Against this background, this paper aims to study whether these classical recommendations are still valid today, when both the magnitude of the problems and the size of the population have increased considerably, considering how this adjustment affects the performance of the procedure. In addition, the increase in the population size, coupled with the fact that multi-objective optimization is being addressed, has led to the development of a novel elitist probabilistic migration strategy that considers only the Pareto front. The results show some interesting and unexpected conclusions, in which other issues, such as the number of subpopulations or their size, should be considered when fitting multi-population models. Furthermore, some of the previously mentioned classical recommendations may not be well-suited for high-dimensional problems.}, archive = {J_NEUCOM}, author = {Juan Carlos Gómez-López and Manuel Rodríguez-Álvarez and Daniel Castillo-Secilla and Jesús González}, doi = {10.1016/j.neucom.2025.130631}, journal = {Neurocomputing}, month = {10}, pages = {130631}, shortjournal = {Neurocomputing}, title = {Tuning multi-objective multi-population evolutionary models for high-dimensional problems: The case of the migration process}, volume = {648}, year = {2025}, }
• (2025). M4Caster: Multi-source, multi-spatial, multi-temporal modeling for precipitation nowcasting. NEUCOM, 648, 130621. (www)

  @article{ , abstract = {Precipitation nowcasting, especially accurate rainstorm warnings, is crucial for mitigating meteorological risks and enhancing public safety. Owing to the inherently chaotic evolution of precipitation systems, addressing the problem poses significant challenges. The challenges include the lack of comprehensive understanding of chaotic precipitation systems in mainstream single-source approaches and the limitations of pure convolutional methods in capturing global information due to restricted receptive fields, as well as the instability in training associated with ConvRNN-based models. To solve issues, we propose a novel Multi-source, Multi-spatial, and Multi-temporal feature extraction and fusion framework (M4Caster), that utilizes the observations of satellite and the ground-based radar data over the previous 0–1 h to predict radar echo sequences for the forthcoming 0–1 h. In particular, the meteorological satellite, by mining information on cloud cluster movements, is capable of capturing the features of convective initiation (CI). The M4Caster further integrates a Multi-spatial and Multi-temporal Aggregator (MMA) to adaptly extract and refine spatiotemporal features, where multi-path Cross-scale Perception Refinement (CPR) facilitates perception and communication across multiple pathway. Moreover, a bidirectional bridging technique aligns the representations of the two data sources, which fully leverages the advantages inherent in multi-source data. Experiments on a meteorological dataset in the Yangtze River Delta (YRD) region show that our M4Caster outperforms the ingenious PredRNN++ and Transformer-based methods, delivering better nowcasting performance, especially in rainstorm prediction and the prediction of convective initiation events.}, archive = {J_NEUCOM}, author = {Dan Niu and Chunlei Shi and Tianbao Zhang and Hongbin Wang and Zengliang Zang and Mingbo Jiang and Jun Yang}, doi = {10.1016/j.neucom.2025.130621}, journal = {Neurocomputing}, month = {10}, pages = {130621}, shortjournal = {Neurocomputing}, title = {M4Caster: Multi-source, multi-spatial, multi-temporal modeling for precipitation nowcasting}, volume = {648}, year = {2025}, }
• (2025). Mathematical modeling for precise disease diagnosis: Distance and similarity measures in complex neutrosophic soft structures. NEUCOM, 648, 130620. (www)

  @article{ , abstract = {One of the primary objectives of technological advancement is to provide rapid and effective medical treatment. Achieving this requires a fast and accurate diagnostic process. Since there are many various diseases that share common symptoms with other diseases, due to which accurate diagnosis becomes difficult. Therefore, it is necessary to develop a method that systematically analyzes the symptom states to diagnose diseases accurately in patients. To achieve this, a framework which is capable of handling uncertainty and vagueness is essential, because symptom data often contain elements of fuzziness, vagueness, and uncertainty. To address these challenges, this article proposed the complex single-valued neutrosophic soft set along with its distance measure. After this a similarity measure based on the proposed distance measure of complex single-valued neutrosophic soft is also presented. In the application section, an algorithmic approach based on both distance and similarity measures of the complex single-valued neutrosophic soft set is introduced. This mathematical model is then applied to a real-life disease diagnosis problem and its results are verified through laboratory tests. The proposed model demonstrates an enhancement of approximately 24.6% for Patient 1, 26.3% for Patient 2 and 28.1% for Patient 3, these percentages reflect improvement in performance compared to the competing method, which demonstrate the effectiveness of the proposed approach in handling uncertainty and enhancing decision quality. Additionally, a comparative analysis is conducted between the proposed model and existing approaches, demonstrating its effectiveness in improving diagnostic accuracy. Finally, the paper concludes with a summary of key findings and potential future research directions.}, archive = {J_NEUCOM}, author = {Yeliang Xiao and Ali Asghar and Tasadduq Niaz and Rana Safdar Ali}, doi = {10.1016/j.neucom.2025.130620}, journal = {Neurocomputing}, month = {10}, pages = {130620}, shortjournal = {Neurocomputing}, title = {Mathematical modeling for precise disease diagnosis: Distance and similarity measures in complex neutrosophic soft structures}, volume = {648}, year = {2025}, }
• (2025). Data vulnerability index for the “crowding problem” in nonlinear dimensionality reduction. NEUCOM, 648, 130619. (www)

  @article{ , abstract = {We propose a data vulnerability index measuring the intensity and harmfulness level of the “crowding problem” in nonlinear dimensionality reduction. The index is useful in supporting nonlinear dimensionality reduction by increasing its robustness to this problem. The index informs about the necessity of using the methods secured from the problem and justifies their employment. The vulnerability index provides auxiliary preliminary information that is helpful in conducting and guiding further dimensionality reduction and data visualization. The introduced index is formulated on the basis of the k -Nearest Neighbors ( k -NN) graph of the data. The graph allows for estimating the intrinsic dimensionality of the low-dimensional manifold embedded in the input high-dimensional linear Euclidean space, which is required during our index computation. The experiments on thirteen real datasets confirm the usefulness of our index in nonlinear dimensionality reduction and its ability to detect the “crowding problem” and determine its gravity. The index values ranged from 2 to 26 corresponding to an increase in superiority of the methods using the t -distribution over those not using it. Moreover, we conducted additional experiments on tuning the neighborhood width parameter in Neighborhood Preserving Projections (NPPs). For most datasets, an improvement was achieved based on Adjusted Mutual Information (AMI) and silhouette values. The highest increase in AMI was obtained for t -NeRV (0.9410 vs. 0.8169) and in silhouette for t -SNE (0.8124 vs. 0.6992).}, archive = {J_NEUCOM}, author = {Dominik Olszewski}, doi = {10.1016/j.neucom.2025.130619}, journal = {Neurocomputing}, month = {10}, pages = {130619}, shortjournal = {Neurocomputing}, title = {Data vulnerability index for the “crowding problem” in nonlinear dimensionality reduction}, volume = {648}, year = {2025}, }
• (2025). CoST: Comprehensive structural and temporal learning of social propagation for fake news detection. NEUCOM, 648, 130618. (www)

  @article{ , abstract = {The widespread dissemination of fake news on social media platforms presents significant threats to individual privacy and societal stability. Traditional content-based fake news detection methods are vulnerable to sophisticated adversarial manipulations, while existing propagation-based approaches often fail to fully capture the complex structural and temporal dynamics of news diffusion. To address these limitations, this paper proposes CoST, a Comprehensive Structural and Temporal learning of social propagation for fake news detection that jointly models propagation structural patterns and multi-grained temporal dynamics. Specifically, for structural patterns, as the existing Graph Convolution Networks (GCN) based methods are inadequate to embed news’ propagation graphs that typically have hub structures and deep propagation paths, we propose a bi-directional Graph Attention LSTM module to capture the social hub and deep propagation patterns of news’ propagation graphs. Besides structural patterns, news propagation may also have complicated and diverse temporal patterns. To model the multi-grained temporal dynamics of propagation, we adopt a temporal-aware attention mechanism and a Transformer encoder-based self-attention mechanism to learn the local temporal interval and global propagation sequence features, respectively. Experimental results on several real-world datasets demonstrate the superiority of CoST over various state-of-the-arts, especially in the early detection of fake news.}, archive = {J_NEUCOM}, author = {Zechen Guo and Peng Wu and Xiaoliang Liu and Li Pan}, doi = {10.1016/j.neucom.2025.130618}, journal = {Neurocomputing}, month = {10}, pages = {130618}, shortjournal = {Neurocomputing}, title = {CoST: Comprehensive structural and temporal learning of social propagation for fake news detection}, volume = {648}, year = {2025}, }
• (2025). Knowledge guided controllable diffusion for enhanced autonomous driving scenarios generation. NEUCOM, 648, 130616. (www)

  @article{ , abstract = {Considering the shortcomings of the natural driving datasets composed by typical scenarios in the verification and evaluation of autonomous driving, how to improve the diversity, realism and controllability to satisfy the complex driving scenarios need is an urgent problem to be solved. Therefore, a knowledge-guided controllable diffusion (KGCD) scenarios generation framework is proposed, which allows for the generation behaviors that conform to expected attributes through a differentiable cost designed based on prior driving knowledge during the inference stage. Specifically, factorized attention is utilized for capturing and fusion of the spatio-temporal interaction features for the scene embeddings, based on which the accurate scene-level interactive joint motion prediction is achieved. In the inference stage, diversified guidance is designed based on prior knowledge and embedded into a generation framework to guide the generation of controllable and realistic driving scenarios, filling the gap in the existing scenario library. In response to the slow inference speed of typical diffusion model, a learnable motion pattern estimator (MPE) module is proposed to improve inference speed while ensuring generation quality. Based on the experiments with different datasets, KGCD proposed in this paper can satisfy the requirements of diverse scene generation. Compared with the best performance obtained by other baseline, the realism, controllability and stability metrics have been improved on average by 3.57%, 7.66% and 7.71% respectively evaluated by nuScenes dataset.}, archive = {J_NEUCOM}, author = {Ce Shan and Lulu Guo and Hong Chen}, doi = {10.1016/j.neucom.2025.130616}, journal = {Neurocomputing}, month = {10}, pages = {130616}, shortjournal = {Neurocomputing}, title = {Knowledge guided controllable diffusion for enhanced autonomous driving scenarios generation}, volume = {648}, year = {2025}, }
• (2025). Clustering with dynamic bipartite graph learning. NEUCOM, 648, 130615. (www)

  @article{ , abstract = {Anchor-based bipartite graph clustering algorithms greatly enhance data analysis by accelerating computations without sacrificing performance. However, these methods face two limitations: reliance on fixed bipartite graphs and lack of interaction between the bipartite graph and anchor labels, which reduces adaptability to dynamic data and clustering performance. To overcome these, a novel method, called Clustering with Dynamic Bipartite Graph Learning (DBGL) is proposed. The DBGL objective function is composed of two parts: construct the bipartite graph and learn the anchor labels. The bipartite graph and anchor labels are updated iteratively, resulting in a dynamic bipartite graph that evolves throughout the clustering process. Our dynamic technique is more flexible and performs better on large, complicated datasets than fixed bipartite graph methods. Additionally, we introduce label transmission into DBGL, enabling autonomous interaction between bipartite graph and anchor labels, which mutually guide each other toward optimal clustering results. We design an alternating optimization algorithm where the bipartite graph is updated using an Iteratively Re-Weighted (IRW) algorithm, while the anchor labels are optimized through an improved Coordinate Descent (CD) algorithm. Comprehensive experiments on benchmark datasets demonstrate that DBGL outperforms leading methods in clustering accuracy, efficiency, and robustness, with theoretical analysis confirming its convergence and stability.}, archive = {J_NEUCOM}, author = {Yun Liang and Wang Gao and Qimin Liang and Cankun Zhong and Feiping Nie}, doi = {10.1016/j.neucom.2025.130615}, journal = {Neurocomputing}, month = {10}, pages = {130615}, shortjournal = {Neurocomputing}, title = {Clustering with dynamic bipartite graph learning}, volume = {648}, year = {2025}, }
• (2025). TSAD: Temporal–spatial association differences-based unsupervised anomaly detection for multivariate time-series. NEUCOM, 648, 130611. (www)

  @article{ , abstract = {Modern industrial control systems are vast and intricate, requiring the monitoring of data from numerous interconnected sensors and actuators for precise intrusion and anomaly detection. While unsupervised time series anomaly detection methods based on deep learning effectively capture complex nonlinear contextual dependencies, the anomaly metrics employed by current methods lack contextual anomaly information, thereby hindering the distinction between anomalies and normalies. Addressing this issue, a Temporal–Spatial Association Differences-based Anomaly Detection model (TSAD) is proposed. This model introduces temporal association difference learning, capturing the temporal association distribution of normal sequences while considering temporal association loss to calculate temporal association differences. Additionally, it incorporates spatial association difference learning, capturing the spatial association distribution of normal sequences while considering spatial association loss to calculate spatial association differences. By focusing on extracting temporal–spatial association patterns from multivariate time-series data under normal operating conditions, the model aggregates reconstruction errors and temporal–spatial association differences during testing to detect anomalies using a novel anomaly metric. Experimental results on four real-world datasets (SWaT, WADI, PSM, and MSL) demonstrate the state-of-the-art performance of the approach.}, archive = {J_NEUCOM}, author = {Hanbing Zhu and Nan Xiao and Hefei Ling and Zongyi Li and Yuxuan Shi and Chuang Zhao and Hongxu Ji and Ping Li and Hui Liu}, doi = {10.1016/j.neucom.2025.130611}, journal = {Neurocomputing}, month = {10}, pages = {130611}, shortjournal = {Neurocomputing}, title = {TSAD: Temporal–spatial association differences-based unsupervised anomaly detection for multivariate time-series}, volume = {648}, year = {2025}, }
• (2025). Low-rank tensor recovery via jointing the non-convex regularization and deep prior. NEUCOM, 648, 130610. (www)

  @article{ , abstract = {This paper addresses the low-rank tensor completion (LRTC) and tensor robust principal component analysis (TRPCA), which have broad applications in the recovery of real-world multi-dimensional data. To enhance recovery performance, we propose novel non-convex tensor recovery models for both LRTC and TRPCA by combining low-rank priors with data-driven deep priors. Specifically, we use the tensor ℓ r p pseudo-norm to effectively capture the low-rank structure of the tensor, providing a more accurate approximation of its rank. In addition, a convolutional neural network (CNN) denoiser is incorporated to learn deep prior information, further improving recovery accuracy. We also develop efficient iterative algorithms for solving the proposed models based on the alternating direction method of multipliers (ADMM). Experimental results show that the proposed methods outperform state-of-the-art techniques in terms of recovery accuracy for both LRTC and TRPCA.}, archive = {J_NEUCOM}, author = {Qing Liu and Huanmin Ge and Xinhua Su}, doi = {10.1016/j.neucom.2025.130610}, journal = {Neurocomputing}, month = {10}, pages = {130610}, shortjournal = {Neurocomputing}, title = {Low-rank tensor recovery via jointing the non-convex regularization and deep prior}, volume = {648}, year = {2025}, }
• (2025). Var-nmODE: Model with l2-stability based on nmODE for defending against adversarial attacks. NEUCOM, 648, 130605. (www)

  @article{ , abstract = {Deep neural networks (DNN) have demonstrated remarkable performance in various applications. However, their performance is significantly influenced by a wide range of perturbations, particularly adversarial perturbations, especially adversarial perturbations, which are difficult to recognize by the naked eye but cause the network to produce incorrect classifications. Some studies have shown that ordinary differential equation (ODE) networks are inherently more robust to adversarial perturbations than general deep networks. nmODE (Neural Memory Ordinary Differential Equation) is a recently proposed artificial neural network model, which has strong nonlinearity. Despite its potential, nmODE still faces challenges in adversarial defense. In this paper, we propose a variant model of neural memory ordinary differential equations (var-nmODE) to defend against adversarial attacks. Based on the theoretical foundation, var-nmODE has L 2 stable mapping, which corresponds to authentication defense against L 2 adversarial perturbations. Further, we conduct adversarial training on the proposed model and show that var-nmODE has better performance through experiments than nmODE. In addition, through adversarial training, the performance of var-nmODE is significantly improved, which indicates that our proposed model can resist adversarial disturbance. It is worth mentioning that var-nmODE provides inherent and certified stability, making it a valuable addition to deep learning defense research.}, archive = {J_NEUCOM}, author = {Qing Yang and Xiaobing Hu and Jiali Yu and Qixun Sun and Lan Shu and Zhang Yi and Yong Liao}, doi = {10.1016/j.neucom.2025.130605}, journal = {Neurocomputing}, month = {10}, pages = {130605}, shortjournal = {Neurocomputing}, title = {Var-nmODE: Model with l2-stability based on nmODE for defending against adversarial attacks}, volume = {648}, year = {2025}, }
• (2025). Event-triggered control for exponential synchronization of reaction–diffusion fractional-order clifford-valued delayed neural networks and its application to image encryption. NEUCOM, 648, 130604. (www)

  @article{ , abstract = {This paper investigates the α -exponential synchronization problem for reaction–diffusion fractional-order Clifford-valued delayed neural networks (RDFOCLVDNNs) using an event-triggered control (ETC) strategy. First, a general form of neural networks (NNs), namely RDFOCLVDNNs is considered, which provides deeper insights into the dynamics of Clifford-valued neural networks (CLVNNs). To address the challenges posed by the non-commutative nature of Clifford algebra, the RDFOCLVDNNs are decomposed into multi-dimensional real-valued neural networks (RVNNs), which avoid the complexities of Clifford number multiplication and also simplify the analysis. Then, by constructing a suitable Lyapunov–Krasovskii functional (LKF) and applying appropriate inequalities new robust conditions are derived to guarantee the α -exponential synchronization of RDFOCLVDNNs under the proposed ETC strategy. To validate the synchronization criteria, a numerical example is presented along with graphical analysis. Furthermore, proposed theoretical framework is utilized to develop an effective color image encryption algorithm for secure image transmission. Finally, the effectiveness and security of the proposed encryption scheme are verified through simulations and various performance analyses.}, archive = {J_NEUCOM}, author = {R. Sriraman and N. Manoj and P. Agarwal and J. Vigo-Aguiar and Shilpi Jain}, doi = {10.1016/j.neucom.2025.130604}, journal = {Neurocomputing}, month = {10}, pages = {130604}, shortjournal = {Neurocomputing}, title = {Event-triggered control for exponential synchronization of reaction–diffusion fractional-order clifford-valued delayed neural networks and its application to image encryption}, volume = {648}, year = {2025}, }
• (2025). Dual-channel hypergraph networks in the time–frequency domain for learning advanced spatiotemporal dependencies in multivariate time series. NEUCOM, 648, 130600. (www)

  @article{ , abstract = {Multivariate time series play a crucial role in applications such as human activity recognition and fault diagnosis, where capturing spatiotemporal dependencies is paramount. While recent methods have leveraged graph neural networks (GNNs) or hybrid models that integrate convolutional neural networks (CNNs) and long short-term memory (LSTM) networks to learn these dependencies, two significant challenges remain: First, previous research has neglected the dynamic dependencies and overall correlations among multivariate time series, which are crucial for comprehensive data utilization. Second, when processing both temporal and spatial dimensional information, how to effectively integrate higher-order information while avoiding complexity issues in model design and implementation remains a challenge. To address these challenges, we propose a Dual-channel Temporal–Spatial Hypergraph Network (DTSHN) classification framework based on the time–frequency domain. Specifically, our DTSHN is a time–frequency domain dual-channel interactive learning model that enhances the utilization and diversity of time series data. To capture higher-order relationships among signals, we design a dynamic hypergraph structure learning module and construct temporal and spatial hypergraph structures in the time–frequency domain. Additionally, we propose a temporal–spatial hypergraph network module to learn temporal correlations and spatial dependencies within time series. Furthermore, to address the feature consistency issue in the time–frequency domain under the dual-channel paradigm, we introduce a time–frequency contrastive learning strategy to enable the model to learn effective feature representations across both domains. The experimental results show that our proposed model achieves 96.77%, 94.33% and 97.57% optimal accuracy in HAR, FM and PS datasets, respectively. In addition, we have reduced the FLOPs by 19.8% compared to other optimal models.}, archive = {J_NEUCOM}, author = {Ziwei Chen and Jianjian Jiang and Xiangmin Luo and Fangyuan Lei and Xiaochen Yuan and Jin Zhan}, doi = {10.1016/j.neucom.2025.130600}, journal = {Neurocomputing}, month = {10}, pages = {130600}, shortjournal = {Neurocomputing}, title = {Dual-channel hypergraph networks in the time–frequency domain for learning advanced spatiotemporal dependencies in multivariate time series}, volume = {648}, year = {2025}, }
• (2025). End-to-end one-stream object tracking based on uncertainty regression. NEUCOM, 648, 130599. (www)

  @article{ , abstract = {Most popular tracking methods leverage visual information for tracking through deep neural network architectures. However, issues such as occlusion and disappearance can arise during the target’s movement, leading to inaccuracies or failures in bounding box prediction. Therefore, it is crucial to model the uncertainty of the bounding box which not only improves the localization accuracy but also provides a confidence score to assess the validity of the localization. To address these issues, we propose an end-to-end one-stream object tracking with uncertainty prediction network (OOTU) based on Vision Transformer. The localization head predicts the box coordinates and simultaneously estimates the uncertainty of the predicted bounding box coordinates. Meanwhile, we introduce an innovative regression loss function that concurrently learns both bounding box regression and localization variance. This approach significantly enhances the localization precision and reflects the level of uncertainty of the tracking, all while incurring minimal additional computational overhead. Extensive experiments on public datasets like TrackingNet demonstrate the effectiveness of our approach, achieving an AUC of 83.5% on TrackingNet, outperforming the baseline of 83.1%. The experimental results show that the proposed end-to-end one-stream object tracking method with uncertainty prediction can effectively improve the localization accuracy of tracking bounding box.}, archive = {J_NEUCOM}, author = {Youcheng Tang and Ang Su and Zhang Li and Zi Wang}, doi = {10.1016/j.neucom.2025.130599}, journal = {Neurocomputing}, month = {10}, pages = {130599}, shortjournal = {Neurocomputing}, title = {End-to-end one-stream object tracking based on uncertainty regression}, volume = {648}, year = {2025}, }
• (2025). Few-shot object counting with frequency attention and multi-perception head. NEUCOM, 648, 130598. (www)

  @article{ , abstract = {Few-Shot Object Counting (FSC) is a critical technique in computer vision, which focuses on estimating the number of exemplar objects in target tasks. This technique is highly versatile and applicable in diverse domains, such as crowd monitoring, traffic management, and wildlife tracking. The primary challenge in FSC is achieving robust feature matching despite the gap between the diversity of targets and the scarcity of exemplars. In this research, we propose the Few-shot Object Counting Network with Frequency Attention and Multi-Perception Head (FFMP), which aims to enhance the limited examples by identifying additional instances within query images. The FFMP framework comprises three core components: Frequency Domain Feature Fusion (FDF), Self-Adaptive Feature Enhancement (SFE), and Multi-Perception Head (MP). The FDF component fuses features from both spatial and frequency domains to generate more precise similarity maps. The SFE component identifies and focuses on recurrent target features within query images, enriching the initial set of examples and providing a detailed understanding of the target category. Additionally, the MP component integrates counting and detection tasks, thereby improving overall performance. Extensive experiments on the FSC-147 dataset and various class-specific counting datasets demonstrate that FFMP achieves competitive counting performance compared to state-of-the-art methods. Code is available at https://github.com/dsl161/FFMP .}, archive = {J_NEUCOM}, author = {Gaoxin Ma and Xingquan Zhu and Zhen Tian and Yangdong Ye and Zhenfeng Zhu}, doi = {10.1016/j.neucom.2025.130598}, journal = {Neurocomputing}, month = {10}, pages = {130598}, shortjournal = {Neurocomputing}, title = {Few-shot object counting with frequency attention and multi-perception head}, volume = {648}, year = {2025}, }
• (2025). Beyond KV caching: Shared attention for efficient LLMs. NEUCOM, 648, 130587. (www)

  @article{ , abstract = {The rapid scaling of Large Language Models (LLMs) necessitates advancements in computational and memory efficiency during inference. While methods like Multi-Query Attention (MQA), Grouped-Query Attention (GQA), and Cross-Layer Attention (CLA) reduce Key–Value (KV) cache size by sharing K/V pairs, strategies to further reduce computational load include reusing computed attention weights across layers, an idea explored previously (e.g., LazyFormer (Ying et al., 2021)). This paper provides an extensive empirical investigation into the phenomenon of attention weight isotropy—high similarity in attention distributions across layers—within diverse modern LLMs (7B-72B scale). We demonstrate how this isotropy develops during pretraining, offering a fundamental insight into LLM attention dynamics. Leveraging these findings, we systematically evaluate and validate a cross-layer weight sharing technique, termed Shared Attention (SA). SA selectively reuses computed attention weights in layer spans identified as isotropic through our analysis. Our experiments across multiple benchmarks show that strategically applied SA maintains comparable performance to baseline models, particularly in later layers where isotropy is pronounced, while significantly reducing computational FLOPs and key cache requirements associated with attention calculation. This work provides principled guidance for optimizing attention mechanisms based on empirically observed layer dynamics in contemporary LLMs. Code and resources are available at https://github.com/metacarbon/shareAtt .}, archive = {J_NEUCOM}, author = {Liao Bingli and Danilo Vasconcellos Vargas}, doi = {10.1016/j.neucom.2025.130587}, journal = {Neurocomputing}, month = {10}, pages = {130587}, shortjournal = {Neurocomputing}, title = {Beyond KV caching: Shared attention for efficient LLMs}, volume = {648}, year = {2025}, }
• (2025). Collision-free motion-constrained path planning for multiple unmanned delivery vehicles based on heuristic deep reinforcement learning. NEUCOM, 648, 130586. (www)

  @article{ , abstract = {This paper proposes a collision-free motion-constrained path planning method for multiple Unmanned Delivery Vehicles (UDVs) based on heuristic deep reinforcement learning. The method focuses on addressing the vehicles’ collision avoidance, motion constraints and mission timeliness in complex environments. First, a priority-based collaborative path planning strategy is introduced to avoid spatiotemporal resource conflicts among UDVs. By pre-planning delivery paths using the A ∗ algorithm, each vehicle’s priority is determined upon mission urgency. Second, hierarchical action sets are designed based on priority, with turning constraints incorporated by restricting the possible action directions. Then, collision avoidance between UDVs and adherence to turning constraints are guaranteed during the action selection process. The action selection process follows a hierarchical strategy, where UDVs with varying priorities use distinct selection strategies. Finally, an improved DQN algorithm, integrated with the heuristic A* algorithm, is proposed to improve learning efficiency. The A* algorithm provides priori knowledge to initialize the DQN and guide the training process to prevent collisions, so as to reduce invalid exploration and avoid deadlocks. Experimental results demonstrate that the proposed method significantly enhances the overall task completion efficiency and safety of UDV groups, showing great potential for practical applications.}, archive = {J_NEUCOM}, author = {Honggui Han and Yabin Zhang and Yanting Huang}, doi = {10.1016/j.neucom.2025.130586}, journal = {Neurocomputing}, month = {10}, pages = {130586}, shortjournal = {Neurocomputing}, title = {Collision-free motion-constrained path planning for multiple unmanned delivery vehicles based on heuristic deep reinforcement learning}, volume = {648}, year = {2025}, }
• (2025). Pinning boundary sampled-data synchronization of coupled reaction–diffusion neural networks. NEUCOM, 648, 130585. (www)

  @article{ , abstract = {In this paper, the pinning boundary sampled-data (SD) synchronization problem of coupled reaction–diffusion neural networks (CRDNNs) with SD-based couplings is investigated under boundary and distributed SD measurements. First, a pinning boundary SD controller is designed to effectively address the synchronization problem of CRDNNs, utilizing boundary and distributed SD measurements and requiring only a small number of sensors and actuators. Then, linear matrix inequalities provide adequate conditions that ensure exponential stability of the synchronization error system by using the designed pinning boundary SD controller, derived through inequality techniques and Lyapunov functionals. Finally, we present a numerical example to showcase the practical applicability and effectiveness of the proposed pinning boundary SD synchronization method.}, archive = {J_NEUCOM}, author = {Zi-Peng Wang and Bo-Ming Chen and Junfei Qiao and Biao Luo and Huai-Ning Wu and Tingwen Huang and Guangwei Chen}, doi = {10.1016/j.neucom.2025.130585}, journal = {Neurocomputing}, month = {10}, pages = {130585}, shortjournal = {Neurocomputing}, title = {Pinning boundary sampled-data synchronization of coupled reaction–diffusion neural networks}, volume = {648}, year = {2025}, }
• (2025). Fixed-time passivity of multi-weighted coupled memristive Cohen–Grossberg neural networks. NEUCOM, 648, 130581. (www)

  @article{ , abstract = {This article presents a type of multi-weighted coupled memristive Cohen–Grossberg neural networks. Firstly, by utilizing some inequality techniques and constructing a suitable Lyapunov function, a sufficient condition is established to ensure the fixed-time passivity (FXTP) in such networks. Secondly, based on adaptive state feedback control strategy, the FXTP, fixed-time input strict passivity, and fixed-time output strict passivity of the proposed networks are investigated. Further, two fixed-time synchronization criteria for the fixed-time passive multi-weighted coupled memristive Cohen–Grossberg neural networks are derived. Finally, a numerical example is proposed to demonstrate the validity of the theoretical results.}, archive = {J_NEUCOM}, author = {Hong-An Tang and Jin-Wei Li and Xiaofang Hu and Shukai Duan and Lidan Wang}, doi = {10.1016/j.neucom.2025.130581}, journal = {Neurocomputing}, month = {10}, pages = {130581}, shortjournal = {Neurocomputing}, title = {Fixed-time passivity of multi-weighted coupled memristive Cohen–Grossberg neural networks}, volume = {648}, year = {2025}, }
• (2025). DSAIS-PINN: Dynamic seeds allocation importance sampling for physics-informed neural networks. NEUCOM, 648, 130578. (www)

  @article{ , abstract = {Physics-Informed Neural Networks (PINNs) have gained widespread recognition as a capable tool for solving partial differential equations (PDEs), achieving notable success across various applications. Conventional PINNs, on the other hand, encounter significant challenges, including slow convergence rates and substantial computational costs. Recent advancements have incorporated importance sampling methods into the training process of PINNs, leading to considerable improvements in convergence speed and overall efficiency. Despite these advances, current importance sampling techniques typically calculate precise sampling probabilities for seed points and interpolate for non-seed points, which may lead to inaccuracies in high-loss regions. This study introduces the Dynamic Seed Allocation Importance Sampling (DSAIS) method, which dynamically adjusts the number of seed points in regions, i.e., Voronoi cells, with varying loss levels, prioritizing high-loss areas. This refinement improves interpolation accuracy for non-seed points. The proposed method’s effectiveness is confirmed through its application to three benchmark issues: the Schrödinger equation, the 1D Burgers equation, the Korteweg–de Vries equation, and the 2D heat equation. Performance is assessed using standard evaluation metrics, including maximum error, mean absolute error, and root mean square error. The outcomes demonstrate that DSAIS achieves state-of-the-art performance. DSAIS shows up to 63%, 30%, 30% smaller root mean square error in Schrodinger equation, Burgers equation and KDV equation, respectively.}, archive = {J_NEUCOM}, author = {Wentao Feng and Ming Hu and Rui Huang and Chenwei Tang and Shudong Huang and Jiancheng Lv}, doi = {10.1016/j.neucom.2025.130578}, journal = {Neurocomputing}, month = {10}, pages = {130578}, shortjournal = {Neurocomputing}, title = {DSAIS-PINN: Dynamic seeds allocation importance sampling for physics-informed neural networks}, volume = {648}, year = {2025}, }
• (2025). Towards stereoscopic vision: Attention-guided gaze estimation with EEG in 3D space. NEUCOM, 648, 130577. (www)

  @article{ , abstract = {Since traditional gaze-tracking methods rely on line-of-sight estimation, spatial attention modeling from neural activity offers an alternative perspective to gaze estimation. This paper presents a proof-of-concept study on attention-guided gaze estimation with Electroencephalography (EEG), investigating whether brain signals can be leveraged to estimate attentional focus within a controlled 3D environment. We first conducted a preliminary survey to gather public opinions, revealing a generally positive attitude towards EEG-driven gaze tracking. Building on this insight, we collected an EEG dataset in VR, where participants engaged with stimuli presented at predefined spatial locations. We introduce a deep learning model that estimates the relative saliency of candidate positions, enabling gaze estimation through optimization within the learned representation. Our results demonstrate that attentional focus was successfully mapped in a 3D coordinate space from 5 participants, and low-frequency oscillations contributed more significantly to predictive performance. The model achieved robust accuracy in distinguishing gaze locations, highlighting the potential of EEG-based gaze estimation for attention tracking in 3D environments.}, archive = {J_NEUCOM}, author = {Dantong Qin and Yang Long and Xun Zhang and Zhibin Zhou and Yuting Jin and Pan Wang}, doi = {10.1016/j.neucom.2025.130577}, journal = {Neurocomputing}, month = {10}, pages = {130577}, shortjournal = {Neurocomputing}, title = {Towards stereoscopic vision: Attention-guided gaze estimation with EEG in 3D space}, volume = {648}, year = {2025}, }
• (2025). Individual vs. group dynamics: Impacts on equilibrium states in self-organizing systems. NEUCOM, 648, 130575. (www)

  @article{ , abstract = {Recent research highlights the critical need to learn complex structures from data and adapt to evolving data patterns. In this work, we introduce the Decentralized SyncMap model, which shifts the focus from group-based to individual-based interactions in order to reveal more meaningful relationships while operating in lower-dimensional spaces. To improve the model’s robustness and adaptability by incorporating past information, we further propose a temporal memory mechanism. The temporal memory of attractors enhances the model’s ability to capture rare but important features, while the temporal memory of repellers helps the model identify local chunk patterns. By leveraging these individual dynamics, the model naturally preserves internal structures that traditional group-level methods tend to overlook. Our experiments on synthetic chunking tasks show that the Decentralized SyncMap model achieves an average NMI of 0.876, outperforming Word2Vec and both the Original and Symmetrical SyncMap models by 12.6%, 10.6%, and 5.5%, respectively. On real-world datasets, it attains an average NMI of 0.840, performing better than competing approaches by 400.1%, 52.7%, and 4.6%. These results suggest that a model emphasizing individual dynamics provides a more effective means of uncovering latent patterns in complex, evolving data.}, archive = {J_NEUCOM}, author = {Xueyuan Li and Danilo Vasconcellos Vargas}, doi = {10.1016/j.neucom.2025.130575}, journal = {Neurocomputing}, month = {10}, pages = {130575}, shortjournal = {Neurocomputing}, title = {Individual vs. group dynamics: Impacts on equilibrium states in self-organizing systems}, volume = {648}, year = {2025}, }
• (2025). Distributed policy evaluation over multi-agent network with communication delays. NEUCOM, 648, 130562. (www)

  @article{ , abstract = {This paper investigates the multi-agent policy evaluation problem for distributed reinforcement learning on time-varying directed communication structure with communication delays. In a completely distributed setting, agents jointly learn the value of a given policy through private local evaluation and neighbors’ evaluation. First, we propose the Push-Sum Dual Averaging Algorithm (PS-DAA) to deal with the distributed policy evaluation problem with communication delays. By considering the inevitable communication delays, a more general time-varying directed communication structure, and more realistic state constraints, PS-DAA still achieves sublinear convergence. Further, considering the case where the full update information is unavailable, we extend PS-DAA to the bandit feedback setting, i.e., the values of the sampling points are used instead of the full gradient information. We prove that compared to the full information scheme, the bandit-feedback PS-DAA does not lead to performance degradation. Finally, we verify the effectiveness of the proposed algorithm through two simulation cases.}, archive = {J_NEUCOM}, author = {Yaoyao Zhou and Gang Chen and Changli Pu and Keyu Wu and Zhenghua Chen}, doi = {10.1016/j.neucom.2025.130562}, journal = {Neurocomputing}, month = {10}, pages = {130562}, shortjournal = {Neurocomputing}, title = {Distributed policy evaluation over multi-agent network with communication delays}, volume = {648}, year = {2025}, }
• (2025). Architectures, variants, and performance of neural operators: A comparative review. NEUCOM, 648, 130518. (www)

  @article{ , abstract = {In recent years, neural operators have emerged as effective alternatives to traditional numerical solvers. They are known for their efficient computation, excellent generalization, and high solving accuracy. Many researchers have shown interest in their design and application. This paper provides a comprehensive summary and analysis of neural operators. We categorize them into three types based on their architecture: deep operator networks (DeepONets), integral kernel operators, and transformer-based neural operators. We then discuss the basic structures and properties of these operator types. Furthermore, we summarize and discuss the various variants and extensions of these three types of neural operators from three directions: (1) operator basis-based neural operator variants; (2) physics-informed neural operator variants; and (3) application of neural operator variants in complex systems. We also analyze the characteristics and performance of different operator methods through numerical experiments. Taking into account these discussions and analyses, we provide perspectives and suggestions regarding the challenges and potential enhancements for different neural operators. This offers valuable guidance and suggestions for the practical application and development of neural operators.}, archive = {J_NEUCOM}, author = {Shengjun Liu and Yu Yu and Ting Zhang and Hanchao Liu and Xinru Liu and Deyu Meng}, doi = {10.1016/j.neucom.2025.130518}, journal = {Neurocomputing}, month = {10}, pages = {130518}, shortjournal = {Neurocomputing}, title = {Architectures, variants, and performance of neural operators: A comparative review}, volume = {648}, year = {2025}, }
• (2025). Cross-trajectory spatiotemporal fusion for user identity linkage. NEUCOM, 648, 130509. (www)

  @article{ , abstract = {As personally identifiable information remains central to privacy protection on online platforms, linking a person’s multiple identities across different platforms has emerged as a critical challenge. This problem, pivotal for cross-domain recommendation, behavior analysis, social network alignment, and the like, has garnered significant attention in academia and industry. This study addresses the user identity linkage problem by leveraging user check-in location data, tackling two key challenges: (1) Check-in records for the same physical location often exhibit slight geospatial discrepancies (e.g., coordinate variations), introducing semantic gaps when one-hot encoding is applied since the method inherently fails to capture spatial proximity relationships. (2) Users often exhibit heterogeneous check-in frequencies across platforms (e.g., frequent check-ins on Foursquare vs. sparse activity on Twitter), which complicates cross-platform user alignment and hinders the modeling of their multifaceted temporal–spatial behavioral patterns. To overcome these issues, we propose CS-Fusion (Cross-trajectory Spatiotemporal Fusion), a novel model that integrates two innovative modules. 1. Cross-trajectory Attention: Utilizes spatiotemporal correlations to merge check-ins associated with the same physical location, mitigating semantic gaps. 2. Mutual-Guidance Attention: Identifies inter-trajectory location relevance to enhance trajectory representation learning, enabling robust alignment of users with divergent check-in behaviors. Extensive experiments on three real-world datasets demonstrate that CS-Fusion outperforms state-of-the-art methods, achieving superior generalizability and robustness in cross-platform user identity linkage. The results highlight its effectiveness in capturing nuanced spatiotemporal patterns and resolving frequency disparities, offering practical value for trajectory-based cross-platform analytics.}, archive = {J_NEUCOM}, author = {Yue Yang and Yichao Zhang and Jihong Guan and Shuigeng Zhou and Wengen Li}, doi = {10.1016/j.neucom.2025.130509}, journal = {Neurocomputing}, month = {10}, pages = {130509}, shortjournal = {Neurocomputing}, title = {Cross-trajectory spatiotemporal fusion for user identity linkage}, volume = {648}, year = {2025}, }
• (2025). Language-guided alignment and distillation for source-free domain adaptation. NEUCOM, 648, 130501. (www)

  @article{ , abstract = {Source-free domain adaptation (SFDA) is a practical problem in which a pre-trained source model is adapted to an unlabeled target domain without accessing the labeled source data. Although recent studies have successfully incorporated vision–language models (VLMs) like CLIP into SFDA frameworks, the performance of existing methods may be limited due to their reliance on coarse-grained class prompts, in which fine-grained textual knowledge has not been fully exploited. To overcome this limitation, we develop a novel framework of Language-guided Alignment and Distillation (LAD) by integrating visual features with fine-grained textual descriptions generated by pre-trained captioning models. Our method consists of two innovative designs, i.e., category-aware modality alignment (CMA) and language-guided knowledge distillation (LKD). CMA aligns cross-modal feature representations with a gating function to filter out high-confidence same-class samples from negatives to preserve intra-class similarity. LKD better adapts the vision encoder to the target domain through adaptive modality fusion and dual-level distillation guided by both visual and textual modalities. Extensive experiments on five benchmarks, including both image and video recognition , demonstrate that our method consistently outperforms the state of the arts for SFDA, e.g., +2.1% in Office-Home and +4.3% in UCF-HMDB.}, archive = {J_NEUCOM}, author = {Jiawen Peng and Jiaxin Chen and Rong Pan and Andy J. Ma}, doi = {10.1016/j.neucom.2025.130501}, journal = {Neurocomputing}, month = {10}, pages = {130501}, shortjournal = {Neurocomputing}, title = {Language-guided alignment and distillation for source-free domain adaptation}, volume = {648}, year = {2025}, }
• (2025). Ensemble predicate decoding for unbiased scene graph generation. NEUCOM, 648, 130496. (www)

  @article{ , abstract = {Scene Graph Generation (SGG) aims to generate a comprehensive graphical representation that accurately captures the semantic information of a given scenario. However, the SGG model’s performance in predicting more fine-grained predicates is hindered by significant predicate bias. According to existing works, the long-tail distribution of predicates in training data results in the biased scene graph. However, the semantic overlap between predicate categories makes predicate prediction difficult, and the significant difference in the sample size of semantically similar predicates makes the predicate prediction more difficult. Therefore, higher requirements are put forward for the discriminative ability of the model. To address this problem, this paper proposes Ensemble Predicate Decoding (EPD), which employs multiple decoders to attain unbiased scene graph generation. Two auxiliary decoders trained on lower-frequency predicates are used to improve the discriminative ability of the model. In order to keep the model concise, the ensemble decoders reuse parameters except for those of the batch normalization layers. Extensive experiments are conducted on the VG dataset, and the experiment results show that EPD enhances the model’s representation capability for predicates. Compared with previous unbiased SGG methods, EPD has relatively superior predictive capability for high-frequency predicates while improving the prediction accuracy of low-frequency predicates.}, archive = {J_NEUCOM}, author = {Jiasong Feng and Lichun Wang and Hongbo Xu and Kai Xu and Baocai Yin}, doi = {10.1016/j.neucom.2025.130496}, journal = {Neurocomputing}, month = {10}, pages = {130496}, shortjournal = {Neurocomputing}, title = {Ensemble predicate decoding for unbiased scene graph generation}, volume = {648}, year = {2025}, }
• (2025). Small object matters in weakly supervised object localization. NEUCOM, 648, 130494. (www)

  @article{ , abstract = {Weakly-supervised object localization (WSOL) methods aim to capture the extent of the target object without full supervision such as bounding boxes or segmentation masks. Although numerous studies have been conducted in the research field of WSOL, we find that most existing methods are less effective at localizing small objects. In this paper, we first analyze why previous studies have overlooked this problem. Based on the analysis, we propose two remedies: (1) new evaluation metrics and a dataset to accurately measure localization performance for small objects, and (2) a novel consistency learning framework to zoom in on small objects so the model can perceive them more clearly. Our extensive experimental results demonstrate that the proposed method significantly improves small object localization on four different backbone networks and four different datasets, without sacrificing the performance of medium and large objects. In addition to these gains, our method can be easily applied to existing WSOL methods as it does not require any changes to the model architecture or data input pipeline. Code is available at https://github.com/dongjunhwang/small_object_wsol .}, archive = {J_NEUCOM}, author = {Dongjun Hwang and Seong Joon Oh and Junsuk Choe}, doi = {10.1016/j.neucom.2025.130494}, journal = {Neurocomputing}, month = {10}, pages = {130494}, shortjournal = {Neurocomputing}, title = {Small object matters in weakly supervised object localization}, volume = {648}, year = {2025}, }
• (2025). Detail retention and enhancement for camouflaged object detection. NEUCOM, 648, 130479. (www)

  @article{ , abstract = {The objective of camouflaged object detection is to segment objects that share similar visual attributes with their backgrounds. Existing methods leverage the consistency and variability of the object to pinpoint its location; however, when presented with samples exhibiting intricate edges, the model tends to overlook critical details. Toward this issue, we analyzed the reasons from the perspective of processing basic features and found that roughly compressing the feature channel could lead to the loss of fine features at the initial stage of the pipeline. Therefore, we introduce a multi-scale feature preprocessing approach named Soft Channel Compression prior to feature fusion, which retains key cues by enhancing the visibility of fine-grained features. We then utilize self-attention to fuse the results of multi-scale feature sampling, synthesizing contextual information between extended details and the main object. Furthermore, we propose the Hierarchical Receptive Field Complementarity module, using a divide-and-conquer strategy to enhance discriminative features with different scales in object details. This facilitates the integration of information across various ranges by employing a hierarchical manner to expand the receptive field, thereby achieving complementary information between local areas. Extensive experimental results on benchmark datasets demonstrate that the proposed method outperforms other state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Shangjing Chen and Feng Xu and Tao Zeng and Xin Li and Jiale Chen and Xin Lyu}, doi = {10.1016/j.neucom.2025.130479}, journal = {Neurocomputing}, month = {10}, pages = {130479}, shortjournal = {Neurocomputing}, title = {Detail retention and enhancement for camouflaged object detection}, volume = {648}, year = {2025}, }
• (2025). CoDeGAN: Contrastive disentanglement for generative adversarial network. NEUCOM, 648, 130478. (www)

  @article{ , abstract = {Disentanglement, a critical concern in interpretable machine learning, has garnered significant attention from the computer vision community. Many existing GAN-based class disentanglement (unsupervised) approaches, such as InfoGAN and its variants, primarily aim to maximize the mutual information (MI) between the generated image and its latent codes. However, this focus may lead to a tendency for the network to generate highly similar images when presented with the same latent class factor, potentially resulting in mode collapse or mode dropping. To alleviate this problem, we propose CoDeGAN (Contrastive Disentanglement for Generative Adversarial Networks), where we relax similarity constraints for disentanglement from the image domain to the feature domain. This modification not only enhances the stability of GANs training but also improves their disentangling capabilities. Moreover, we integrate self-supervised pre-training into CoDeGAN to learn semantic representations, significantly facilitating unsupervised disentanglement. Extensive experimental results demonstrate the superiority of our method over state-of-the-art approaches across multiple benchmarks. The code is available at https://github.com/learninginvision/CoDeGAN .}, archive = {J_NEUCOM}, author = {Zejia Liu and Lili Pan and Xiaohan Guo and Jiangwei Zhao}, doi = {10.1016/j.neucom.2025.130478}, journal = {Neurocomputing}, month = {10}, pages = {130478}, shortjournal = {Neurocomputing}, title = {CoDeGAN: Contrastive disentanglement for generative adversarial network}, volume = {648}, year = {2025}, }
• (2025). HRAA: Heuristic reclamation with agent allocation based on reinforcement learning for resource scheduling of financial agent-based modeling and simulation tasks. NEUCOM, 648, 130477. (www)

  @article{ , abstract = {Agent-Based Modeling and Simulation (ABMS) has emerged as a powerful tool for simulating complex systems, particularly in financial markets. However, the inherent complexity of financial systems, characterized by numerous interacting agents, requires the use of parallel simulation to enhance computational efficiency. Existing task placement and resource allocation methods often fall short in providing performance guarantees for such dynamic and resource-intensive simulations. This paper addresses these challenges by proposing a novel algorithm, Heuristic Reclamation and Agent Allocation (HRAA), which integrates reinforcement learning and heuristic strategies to optimize resource scheduling in real time. The HRAA algorithm can dynamically adjust resource allocation, reclaim resources from imbalanced sub-tasks, and optimize parallel execution and synchronization, thereby reducing computational time and improving resource utilization. Experimental results demonstrate that the HRAA algorithm can reduce scheduling overheads by nearly 50% compared to conventional approaches, enhancing the efficiency of financial simulations. This paper not only advances the field of ABMS in financial contexts but also provides a robust solution for resource scheduling in multi-agent parallel simulations.}, archive = {J_NEUCOM}, author = {Pengzhu Pang and Jian Yu and Yu Fang and Lu Liu and John Panneerselvam and Zhijun Ding and Changjun Jiang}, doi = {10.1016/j.neucom.2025.130477}, journal = {Neurocomputing}, month = {10}, pages = {130477}, shortjournal = {Neurocomputing}, title = {HRAA: Heuristic reclamation with agent allocation based on reinforcement learning for resource scheduling of financial agent-based modeling and simulation tasks}, volume = {648}, year = {2025}, }
• (2025). A probabilistic optimizer for binary neural networks. NEUCOM, 648, 130297. (www)

  @article{ , abstract = {Optimizing binarized neural networks (BNNs) offers efficiency advantages but presents unique optimization challenges due to the limited representational capacity of binary values. Many contemporary techniques optimize weights by introducing latent real-valued counterparts within conventional optimization frameworks, approximating the derivative of the potentially stochastic function round : R n → { ± 1 } n . However, some recent works focus on training BNNs by implementing weight flipping based on the sign and magnitude of accumulated gradients. In this paper, we thoroughly investigate the principles of weight sign reversals and find that the training becomes unstable when the accumulated gradients approach marginal values that trigger a flipping. To address this issue, we introduce an innovative probabilistic optimizer to determine weight signs. This optimizer leverages the Bernoulli distribution of accumulated gradients to make informed decisions about weight signs, thereby enhancing training stability. Furthermore, we implement measures to minimize the likelihood of sign reversals in weights that frequently undergo flips to reduce the training instability. We conduct an extensive series of experiments on the datasets CIFAR-10, CIFAR-100, and TinyImageNet, and our approach demonstrates significant improvements in accuracy when compared to state-of-the-art optimizers.}, archive = {J_NEUCOM}, author = {Xue He and Xue Geng and Tiancheng Zhang and Minghe Yu and Min Wu and Ge Yu and Yuhai Zhao}, doi = {10.1016/j.neucom.2025.130297}, journal = {Neurocomputing}, month = {10}, pages = {130297}, shortjournal = {Neurocomputing}, title = {A probabilistic optimizer for binary neural networks}, volume = {648}, year = {2025}, }
• (2025). GPPT: Graph pyramid pooling transformer for visual scene. NEUCOM, 647, 130729. (www)

  @article{ , abstract = {In the field of computer vision, network architectures are critical to the performance of tasks. Vision Graph Neural Network (ViG) has shown remarkable results in handling various vision tasks with their unique characteristics. However, the lack of multi-scale information in ViG limits its expressive capability. To address this challenge, we propose a Graph Pyramid Pooling Transformer (GPPT), which aims to enhance the performance of the model by introducing multi-scale feature learning. The core advantage of GPPT is its ability to effectively capture and fuse feature information at different scales. Specifically, it first generates multi-level pooled graphs using a graph pyramid pooling structure. Next, it encodes features at each scale using a weight-shared Graph Convolutional Neural Network (GCN). Then, it enhances information exchange across scales through a cross-scale feature fusion mechanism. Finally, it captures long-range node dependencies using a transformer module. The experimental results demonstrate that GPPT achieves exceptional performance across various visual scenes, including image classification, and object detection, highlighting its generality and validity.}, archive = {J_NEUCOM}, author = {Zhi-Peng Li and Wen-Jian Liu and Xin Sun and Yi-Jie Pan and Valeriya Gribova and Vladimir Fedorovich Filaretov and Anthony G. Cohn and De-Shuang Huang}, doi = {10.1016/j.neucom.2025.130729}, journal = {Neurocomputing}, month = {9}, pages = {130729}, shortjournal = {Neurocomputing}, title = {GPPT: Graph pyramid pooling transformer for visual scene}, volume = {647}, year = {2025}, }
• (2025). Few-shot learning framework based on classifier and domain adaptive alignment for hyperspectral classification. NEUCOM, 647, 130726. (www)

  @article{ , abstract = {The focus of cross-domain hyperspectral image (HSI) classification is on the model training phase, where the model can utilize a large number of labeled samples from the source domain (SD) to learn prior knowledge. And transfer the knowledge to the target domain (TD) classification task that contains a few labeled samples. Nevertheless, it is difficult to align the features of the SD and TD into a unified space using only a few samples to obtain domain-invariant feature representations. To address this problem, a cross-domain few-shot learning (FSL) framework based on classifier and domain-level adaptive alignment (CDLA) is proposed for HSI classification. The separation-selection-refinement strategy (SSRS) is used to design the parameter-sharing network, which can utilize features with a large amount of information and extract more representative fine-grained features. During the training stage, labels are determined using the covariance metric, and additional classifiers are added to design label consistency constraints to optimize the class prototype and strengthen the intra-class feature consistency. In addition, the domain adversarial network is used to mitigate the effects of the domain shift and learn collaborative representation. Comparative experiments are conducted on three public HSI datasets to validate the effectiveness of the CDLA model and its superiority in existing FSL methods for HSI classification.}, archive = {J_NEUCOM}, author = {Lili Yu and Xubing Zhang and Kai Wang}, doi = {10.1016/j.neucom.2025.130726}, journal = {Neurocomputing}, month = {9}, pages = {130726}, shortjournal = {Neurocomputing}, title = {Few-shot learning framework based on classifier and domain adaptive alignment for hyperspectral classification}, volume = {647}, year = {2025}, }
• (2025). Dual-alignment cross-domain few-shot learning with lightweight domain-specific attention for hyperspectral image classification. NEUCOM, 647, 130722. (www)

  @article{ , abstract = {Cross-domain few-shot learning (FSL) has achieved great success in tackling small-sample hyperspectral image (HSI) classification problems. However, most existing cross-domain FSL methods focused solely on the distribution alignment between domains from a high-level semantic feature perspective, while the low-level spectral feature alignment was somewhat neglected. Additionally, previous methods paid less attention to the domain-specific information when building domain-invariant representations, which may result in a loss of discriminability and give wrong guidance for HSI feature extraction. To address these issues, we propose a dual-alignment cross-domain FSL (DA-CFSL) framework, which collaborates with a novel lightweight domain-specific attention network (LDAN) that extracts discriminative features. In detail, we propose a dual-alignment strategy that includes asymmetric discrepancy-based spectral alignment and adversarial-based semantic alignment, imposing stronger constraints to mitigate domain shift in both low-level spectral feature and high-level semantic feature perspectives. As compensation for the loss of domain-specific information, the LDAN feature extractor is devised to build global dependencies among samples exclusively in the target domain. Therefore, it can not only explore domain-irrelevant information but also retain sensitivity to the target domain. Besides, to further enhance discriminative ability, the supervised contrastive loss and the multikernel triplet loss are utilized, thereby promoting intraclass compactness and interclass dispersion. Experimental results conducted on four HSI datasets show that the DA-CFSL achieves superior performance with lower computational complexity compared to other state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Ruiru Qin and Mingyun Lv and Chuanzhi Wang and Yongmei Wu and Huafei Du}, doi = {10.1016/j.neucom.2025.130722}, journal = {Neurocomputing}, month = {9}, pages = {130722}, shortjournal = {Neurocomputing}, title = {Dual-alignment cross-domain few-shot learning with lightweight domain-specific attention for hyperspectral image classification}, volume = {647}, year = {2025}, }
• (2025). Multivariate time series approach integrating cross-temporal and cross-channel attention for dysarthria detection from speech. NEUCOM, 647, 130708. (www)

  @article{ , abstract = {Speech analysis offers a non-invasive, low-cost approach to dysarthria detection. Studies have shown that the temporal correlations within speech signals and the interactions among the multidimensional feature variables derived from them can facilitate dysarthria detection. However, current studies either rely on pre-designed feature sets, which depend heavily on cumbersome feature engineering, or focus solely on spectral or high-dimensional audio vectors that capture temporal dependencies while neglecting the interactions between internal multivariate features. We propose an end-to-end method that utilizes audio pre-trained models as multivariate time series feature extractors, combined with InceptionTime and cross-temporal and cross-channel attention mechanisms, to fully capture temporal dependencies and interactions among variables within speech for accurate dysarthria detection. Results show that the proposed method achieves a detection accuracy of 92.06 % on a local Mandarin dysarthria dataset, which is at least 2.17 percentage points higher than previous studies, with the highest stability and the lowest time cost. Furthermore, it achieves an accuracy of 87.73 % on an external English dataset, demonstrating good cross-linguistic adaptability and generalizability. Additionally, experiments show that in connected speech tasks, structured tasks outperform unstructured ones in leveraging interactions, leading to more effective dysarthria detection. These findings validate the effectiveness of the proposed end-to-end dysarthria detection method, further advancing the development of speech analysis as a promising tool for dysarthria screening.}, archive = {J_NEUCOM}, author = {Zhenglin Zhang and Tengfei Wang and Zian Hu and Li-Zhuang Yang and Hai Li}, doi = {10.1016/j.neucom.2025.130708}, journal = {Neurocomputing}, month = {9}, pages = {130708}, shortjournal = {Neurocomputing}, title = {Multivariate time series approach integrating cross-temporal and cross-channel attention for dysarthria detection from speech}, volume = {647}, year = {2025}, }
• (2025). Damped sliding based mining of average utility-driven sequential patterns over uncertain data streams. NEUCOM, 647, 130669. (www)

  @article{ , abstract = {Stream pattern mining has attracted significant attention due to its broad application prospects. However, existing algorithms face two major challenges in practical applications: First, current methods cannot effectively handle uncertain data streams, limiting their applicability. Second, the temporal nature of data streams means that recent data is typically more valuable than historical data, yet existing algorithms fail to adequately account for the temporal sensitivity of data streams, lacking an effective mechanism to distinguish between new and old data, which compromises the practicality of mining results. To address these issues, this paper proposes a High Average-Utility Sequential Pattern mining algorithm for Uncertain Data Streams (HAUSP_UDS). The algorithm employs a probability model to quantify data uncertainty and incorporates an average utility measure to eliminate the impact of sequence length on mining results. Additionally, a sliding window-based time decay mechanism is designed to dynamically adjust data weights, reflecting the temporal value of sequences. To improve mining efficiency, a Dynamic Sequence List (DSL) structure is introduced to efficiently store and compute utility information, thereby accelerating the calculation of utility values and upper bounds. Experiments on both real-world and synthetic datasets demonstrate that, compared to state-of-the-art algorithms, the proposed method achieves significant improvements in recall, precision, time-space efficiency, and scalability.}, archive = {J_NEUCOM}, author = {Ruihua Zhang and Meng Han and Feifei He and Fanxing Meng and Chunpeng Li}, doi = {10.1016/j.neucom.2025.130669}, journal = {Neurocomputing}, month = {9}, pages = {130669}, shortjournal = {Neurocomputing}, title = {Damped sliding based mining of average utility-driven sequential patterns over uncertain data streams}, volume = {647}, year = {2025}, }
• (2025). HRHPE: FRoI guides heterogeneous relationship representation learning for precise head pose estimation. NEUCOM, 647, 130623. (www)

  @article{ , abstract = {The ability to effectively detect head positions has raised concerns in the field of computer vision. However, head pose estimation (HPE) is prone to problems, such as extreme angles, occlusion, and lighting. To effectively address this critical missing information gap in the field of HPE, we propose a heterogeneous relationship guided learning method for the representation of transannular layers that effectively captures and uses “Facial Regions of Interest” (FRoI) and heterogeneous relationships through FRoI morphable modeling. Two key observations are revealed: 1) the saliency of the facial regions of interest; and 2) the heterogeneous relationships between the adjacent postures. On this basis, three modules are proposed, namely Region Feature Generation (RFG), Hierarchical Structure Modeling (HSM) and Heterogeneous Relationship Mining (HRM). In particular, we introduce a regional attention mechanism in the RFG, which assigns a higher weighting to FRoI. In HSM, the concept of the “Rugby style” design is proposed as a model for the cross-layer structure. In HRM, we use the Transformer to explore the interdependencies between facial regions and semantic relationships in rotation of the head. Experiments with three real HPE datasets (300 W_LP, AFLW2000 and BIWI) show that our HRHPE is more efficient than the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Hai Liu and Shijia Qian and Tingting Liu and Zelin Cao and Minhong Wang and Jianping Ju and Zhaoli Zhang}, doi = {10.1016/j.neucom.2025.130623}, journal = {Neurocomputing}, month = {9}, pages = {130623}, shortjournal = {Neurocomputing}, title = {HRHPE: FRoI guides heterogeneous relationship representation learning for precise head pose estimation}, volume = {647}, year = {2025}, }
• (2025). HI-captioner: End-to-end image captioning based on hierarchical multi-scale encoding and cross-modal interactive decoding. NEUCOM, 647, 130622. (www)

  @article{ , abstract = {The image captioning task aims to accurately describe images using natural language. It is widely applied in assisting visually impaired individuals, enhancing human–computer interaction, and image retrieval. However, existing end-to-end image captioning methods typically rely on single-scale feature extraction and use simplistic cross-modal interactions. This restricts the model’s capacity to capture multi-level information and complex scenes, leading to captions that lack detailed expression and exhibit poor semantic coherence in context. In response, this paper proposes an innovative image captioning model HI-Captioner based on Swin Transformer to capture multi-scale features, hierarchical positional information and optimize inter-modal information interaction. HI-Captioner integrates Hierarchical Synergistic Attention Module (HSAM) and Cascading Cross-Modal Interaction Decoder (CCMID). The former combines a hierarchical multi-scale attention mechanism with a hierarchical positional encoding method during the encoding stage, enabling more effective capture of multi-scale features and multi-level positional information of images. This improves the model’s ability to capture intricate nuances and comprehend global semantics in the process of generating captions. The latter adopts a novel cross-modal interaction approach to further enhance information exchange between image and text modalities during the decoding stage, significantly improving the model’s efficiency in integrating and representing information from different modalities. Experimental results demonstrate that our proposed model yields substantial performance enhancements on the MSCOCO and Flickr datasets, particularly excelling in evaluation metrics such as BLEU4 and CIDEr. This demonstrates the great potential of the method in generating accurate and natural image captions. The code will be released at: https://github.com/Lwj-cap/HI_Captioner_main}, archive = {J_NEUCOM}, author = {Wenjing Li and Xue Li and Ziyang Li and Jiong Yu and Pengcheng Chen}, doi = {10.1016/j.neucom.2025.130622}, journal = {Neurocomputing}, month = {9}, pages = {130622}, shortjournal = {Neurocomputing}, title = {HI-captioner: End-to-end image captioning based on hierarchical multi-scale encoding and cross-modal interactive decoding}, volume = {647}, year = {2025}, }
• (2025). Bearing fault diagnosis under time-varying speeds with limited samples using frequency temporal series graph and graph generative classified adversarial networks. NEUCOM, 647, 130613. (www)

  @article{ , abstract = {Bearings are pivotal supporting components within the rotating machinery equipment, often running in acceleration, deceleration, and other time-varying speed conditions. This leads to the bearing frequently operating in unstable environments, increasing the risk of failure. Existing fault diagnosis techniques predominantly rely on data-driven approaches. However, in industrial scenarios, the scarcity of fault samples poses a significant challenge in obtaining sufficient data to train diagnostic models. Moreover, the variation of speed over time can cause confusion in fault characteristics within the data, further complicating the fault diagnosis task. To achieve accurate diagnosis under time-varying rotational speeds with limited fault samples, this study introduces a novel intelligent fault diagnosis method based on the Frequency Temporal Series Graph (FTSG) and Graph Generative Classified Adversarial Networks (GGCAN). Some frequencies in the spectrum are selected as the node features. The nodes are then sequentially connected through edges in time order to construct the FTSG that can map time-varying fault information. Subsequently, the GGCAN model is built. The model employs the generator to augment training samples and assists the classifier in mining fault characteristic information contained in the graph data for fault diagnosis. To attain optimal diagnostic performance, a loss function is designed to balance the discriminative loss and the classification loss. Experiments are conducted utilizing bearing fault data under time-varying rotational speeds. The results demonstrate that the proposed method achieves a diagnosis accuracy rate of over 90 % under conditions of limited samples and time-varying speeds, outperforming other comparative methods.}, archive = {J_NEUCOM}, author = {Sichao Sun and Xinyu Xia and Hua Zhou}, doi = {10.1016/j.neucom.2025.130613}, journal = {Neurocomputing}, month = {9}, pages = {130613}, shortjournal = {Neurocomputing}, title = {Bearing fault diagnosis under time-varying speeds with limited samples using frequency temporal series graph and graph generative classified adversarial networks}, volume = {647}, year = {2025}, }
• (2025). Dynamic heterogeneous graph contrastive learning based on multi-prior tasks. NEUCOM, 647, 130612. (www)

  @article{ , abstract = {Dynamic heterogeneous graph embedding aims to enable a variety of graph-related tasks by efficiently capturing structures and attributes over time in complex graph data. During recent decades, the self-supervised contrastive learning techniques has presented immense advantage as an approach to understanding dynamic heterogeneous graph. However, most forms of self-supervised contrastive learning approach for dynamic heterogeneous graph embedding nowadays emphasize single prior task, which results in its failure to capture multiscale knowledge in dynamic heterogeneous graph. Thus, this study presents a new self-supervised graph contrastive learning framework based on multi-prior tasks for multiscale knowledge in dynamic heterogeneous graph (MTDG). In the proposed model, this study first obtains four embedding vectors as contrastive samples using the local, global, long-term, and short-term encoders. Additionally, this study develops single-contrastive learning comprising four prior tasks to optimize the model’s ability to understand the knowledge on the local, global, long-term, and short-term of the dynamic heterogeneous graph. Besides, this study designs cross-contrastive learning comprising two prior tasks to achieve complementary knowledge between the four embedding vectors generated by the encoders. Furthermore, this study introduces slight random noise and a shuffling strategy to prevent the generation of similar contrastive samples. Ultimately, this study assesses MTDG on twelve dynamic graph datasets from the real world, and the experimental results show that the proposed model achieves the better results on all datasets compared to the eleven baselines.}, archive = {J_NEUCOM}, author = {Wenhao Bai and Liqing Qiu and Weidong Zhao}, doi = {10.1016/j.neucom.2025.130612}, journal = {Neurocomputing}, month = {9}, pages = {130612}, shortjournal = {Neurocomputing}, title = {Dynamic heterogeneous graph contrastive learning based on multi-prior tasks}, volume = {647}, year = {2025}, }
• (2025). Exploiting module evolution correlation relationship for fine-grained bird image classification with structural functional representation. NEUCOM, 647, 130609. (www)

  @article{ , abstract = {Fine-grained bird image classification (FBIC) targeting the classification of avian subspecies is facing challenges due to confusing background, partly occlusions and varied posture. In this paper, we proposed a novel module evolution correlation relationship modeling for FBIC task, which can learn structural functional representations among different functional feathers in a bird. Specially, the proposed EMECR model includes two modules, such as periodic topology mining (PTM), and multi-scale semantics alignment strategy (MSAS). The PTM module is proposed to reveal local periodic organizations with implicit functional expressions, and the MSAS is leveraged for better semantic modeling. In addition, a joint loss is designed to suppress the outliers and enforce semantic consistency. Furthermore, to better model the module evolution relationships between different functional representations and hierarchical context correlation, Mamba architecture is employed as the decoder with its linear computational complexity. Experiments on CUB-200–2011 and NABirds verify that our method can obtain robust results and significantly outperform the existing state-of-the-art FBIC methods. Extended experiments have been conducted on the Stanford Cars dataset to suggest the potential of generalizing our method on other fine-grained visual classification tasks.}, archive = {J_NEUCOM}, author = {Shuang Zeng and Hai Liu and Tingting Liu and Qiuxia Liu and Minhong Wang and Bing Yang and Zhaoli Zhang}, doi = {10.1016/j.neucom.2025.130609}, journal = {Neurocomputing}, month = {9}, pages = {130609}, shortjournal = {Neurocomputing}, title = {Exploiting module evolution correlation relationship for fine-grained bird image classification with structural functional representation}, volume = {647}, year = {2025}, }
• (2025). CSubBT: A modular execution framework with self-adjusting capability for mobile manipulation system. NEUCOM, 647, 130608. (www)

  @article{ , abstract = {Embodied intelligence is advancing the capability of intelligent agents to transition from controlled environments, such as factories, to unstructured real-world settings by integrating perception, planning, and physical interaction. Task and Motion Planning (TAMP) can guide agents in completing complex tasks in these unstructured environments. However, the execution of plans by agents is not merely the implementation of pre-defined instructions. The planned actions often fail due to discrepancies between the perceptual information used in planning and the actual conditions encountered. Existing robust execution systems fall short of providing a universal solution at the execution level, making them unsuitable as actuators for upstream task planners. In this paper, we propose the Conditional Subtree (CSubBT), a modular, self-adjusting execution framework for mobile manipulation systems based on Behavior Trees (BTs). CSubBT decomposes planned actions into sub-actions and leverages BTs to control their execution, addressing potential anomalies without the need for intervention from high-level planners. CSubBT treats common anomalies as constraint non-satisfaction problems and continuously guides the robot in performing tasks by sampling new action parameters in the constraint space when anomalies are detected. We validate the robustness of our framework through extensive manipulation experiments conducted in both simulated and real-world environments across different platforms.}, archive = {J_NEUCOM}, author = {Huihui Guo and Huizhang Luo and Huilong Pi and Mingxing Duan and Kenli Li and Chubo Liu}, doi = {10.1016/j.neucom.2025.130608}, journal = {Neurocomputing}, month = {9}, pages = {130608}, shortjournal = {Neurocomputing}, title = {CSubBT: A modular execution framework with self-adjusting capability for mobile manipulation system}, volume = {647}, year = {2025}, }
• (2025). AM-CD: Joint attention and mamba for remote sensing image change detection. NEUCOM, 647, 130607. (www)

  @article{ , abstract = {In recent years, remote sensing images change detection has been applied in a wide range of scenarios, and significant progress has emerged in the technology. Convolutional neural networks (CNNs) and Transformer detection methods have demonstrated excellent performance in several tasks. However, both methods have inherent limitations: CNNs struggle to capture long-range dependencies, while Transformers have comparatively weaker localized feature extraction. Mamba, a recently introduced state-space model, has demonstrated superior global context modeling compared to Transformers while maintaining lower parameter and computational costs. Based on this, this paper proposes an innovative model named AM-CD, which integrates CNNs and Mamba to effectively leverage attention for local feature extraction and Mamba for global context modeling. Specifically, we introduce the Attention-Mamba (AM) Block into a multi-scale dual-encoder architecture to enhance both local and global feature representation while mitigating Mamba’s inherent limitations in capturing fine-grained edges and details. Furthermore, to address feature discrepancies and loss in bi-temporal images, we propose the Feature Compensation Difference (FCD) Block. This Block incorporates a gating mechanism to reduce redundant information interference, facilitate feature fusion, and improve the discriminability of change regions. Finally, we adopt a Multi-Scale Progressive Decoder (MSPD) Block to further refine change reconstruction. Extensive experiments on three widely used RS datasets demonstrate that AM-CD outperforms existing state-of-the-art (SOTA) methods.}, archive = {J_NEUCOM}, author = {Yingting Guo and Yan Xu and Guo Tang and Zhiqun Yu and Qixian Zhao and Qunli Tang}, doi = {10.1016/j.neucom.2025.130607}, journal = {Neurocomputing}, month = {9}, pages = {130607}, shortjournal = {Neurocomputing}, title = {AM-CD: Joint attention and mamba for remote sensing image change detection}, volume = {647}, year = {2025}, }
• (2025). Towards high-accuracy digital detection in meter images: The ASS-YOLO algorithm and dig21000 dataset. NEUCOM, 647, 130606. (www)

  @article{ , abstract = {Automatic meter reading (AMR) in complex real-world scenarios faces significant challenges including dense small targets, background interference, and environmental distortions. This study presents a systematic solution through a generalized enhancement framework incorporating three dedicated mechanisms: 1) Adaptive Channel Importance Enhancement (ACIEBlock) for noise interference suppression and model generalization, 2) Split-Enhance-Reconstruct (SER) Module for robustness improvement in complex background scenarios, and 3) improved model architecture for small target enhancement. Crucially, these architectural improvements demonstrate universal applicability across YOLO variants. When implemented on YOLOv8 to YOLOv12 frameworks, the proposed enhancements consistently boost detection performance by 1.4–7.1 % in mAP@75 with almost no increase in computational overhead, establishing new state-of-the-art benchmarks while maintaining real-time processing at 183 FPS. To enable comprehensive evaluation, we introduce Dig21000 - the largest AMR benchmark containing 21,000 real-world digit meter images capturing different conditions including motion blur, multi-scale variations, and occlusions. Besides, extensive experiments across 13 detectors reveal ASS-YOLO's superiority, achieving 96.6 % mAP@50 and 64.4 % mAP@75. Our work provides both a robust technical framework and a high-quality evaluation benchmark to advance AMR research in real-world scenarios.}, archive = {J_NEUCOM}, author = {Wei Xie and Ying Guo and Zhe Wu and Yonghao Gu and Shibo Dong and Guokang Xie and Weifeng Xu}, doi = {10.1016/j.neucom.2025.130606}, journal = {Neurocomputing}, month = {9}, pages = {130606}, shortjournal = {Neurocomputing}, title = {Towards high-accuracy digital detection in meter images: The ASS-YOLO algorithm and dig21000 dataset}, volume = {647}, year = {2025}, }
• (2025). The status-based optimization: Algorithm and comprehensive performance analysis. NEUCOM, 647, 130603. (www)

  @article{ , abstract = {In recent years, real-world optimization problems have grown increasingly complex, challenging the effectiveness of traditional deterministic approaches. This paper introduces the Status-based Optimization (SBO), an efficient algorithm inspired by human desire for status advancement. By emulating how individuals seek proximity to and learn from or acquire resources from high-status figures, SBO translates these social patterns into a robust computational method for challenging optimization tasks. Comprehensive comparative analyses against various meta-heuristic algorithms, such as those based on human behaviors, classical methods, and state-of-the-art techniques, validate its strength across problem dimensions of 10, 30, 50, and 100 as defined by the IEEE CEC 2017 test suite. Moreover, SBO excels in feature selection by identifying fewer but more relevant features than 8 established binary optimizers across 9 high-dimensional datasets. The algorithm's adaptability is further demonstrated in a multi-threshold image segmentation application at 3 threshold levels. SBO was benchmarked against 7 advanced meta-heuristic methods to segment 9 breast cancer histology images, successfully identifying optimal threshold values to enhance image clarity and segmentation accuracy. These promising results underscore the potential of SBO for future research and practical applications in a wide array of optimization scenarios. The SBO source code is available at https://aliasgharheidari.com/SBO.html for interested researchers.}, archive = {J_NEUCOM}, author = {Jian Wang and Yi Chen and Chenglang Lu and Ali Asghar Heidari and Zongda Wu and Huiling Chen}, doi = {10.1016/j.neucom.2025.130603}, journal = {Neurocomputing}, month = {9}, pages = {130603}, shortjournal = {Neurocomputing}, title = {The status-based optimization: Algorithm and comprehensive performance analysis}, volume = {647}, year = {2025}, }
• (2025). A comprehensive review of fuzzy logic based interpretability and explainability of machine learning techniques across domains. NEUCOM, 647, 130602. (www)

  @article{ , abstract = {Machine Learning (ML) applications have achieved significant improvements across various domains. However, a key challenge limiting their full adoption in critical fields is their black-box nature. Fuzzy systems, particularly rule-based ones, are being extensively studied for their ability to enhance interpretability through linguistic representation. To investigate the effectiveness of these systems, this paper presents a systematic review of the use of fuzzy systems to improve the interpretability of black-box ML models, while also assessing their performance. The review covered papers published between 1994 and January 2024 across five major digital libraries (IEEE Xplore, ScienceDirect, PubMed, ACM Digital Library, Wiley, and Google Scholar), and 330 papers were selected, analyzed, and categorized based on key criteria, including publication year, channels and sources, research type, black-box systems, fuzzy systems, explanation types, data types, overall interpretability, and overall performance. Several conclusions can be drawn from this review's findings. 1) Most papers were published in conferences, with a rise from 2018 to 2021, focusing on solutions and domains like engineering and medical/biomedical fields. 2) Neural networks are the most studied black-box models, while adaptive neuro-fuzzy inference system and evolving neuro-fuzzy systems are the most used fuzzy techniques, offering the best performance-interpretability tradeoff. 3) Explanations are mainly rule-based, and tabular data is the most evaluated type. 4) While fuzzy logic based ML techniques excel in accuracy, adaptability, and robustness, they face challenges like computational complexity and scalability. Future work should address these issues and explore new interpretability metrics for complex applications.}, archive = {J_NEUCOM}, author = {Hafsaa Ouifak and Ali Idri}, doi = {10.1016/j.neucom.2025.130602}, journal = {Neurocomputing}, month = {9}, pages = {130602}, shortjournal = {Neurocomputing}, title = {A comprehensive review of fuzzy logic based interpretability and explainability of machine learning techniques across domains}, volume = {647}, year = {2025}, }
• (2025). A tiny defect detection method on stamped parts with feature aggregation-diffusion and wasserstein distance. NEUCOM, 647, 130601. (www)

  @article{ , abstract = {The surface of stamped parts often exhibits tiny defects due to limitations in production processes and industrial environments, which significantly affect product quality. These defects, with sparse feature information, are easily overlooked or lost during feature transmission, rendering traditional detection methods ineffective. To resolve this challenge, a novel defect detection framework is developed through the integration of feature aggregation-diffusion mechanisms with Wasserstein distance metrics. The architecture is implemented through a frequency-aware encoder consisting of multi-scale feature aggregation and cross-layer adaptive diffusion operations, where feature degradation during transmission is effectively prevented. Critical feature extraction is achieved in the early fusion stage through frequency-guided cluster kernel perception, while representation refinement is accomplished in the late fusion stage via decoupled high-low frequency attention encoding. Furthermore, a Wasserstein-IoU navigation loss function is constructed by combining normalized Wasserstein distance with Generalized IoU, supplemented with pixel-wise Gaussian distribution modeling to mitigate scale dependency. A washing machine front seal door (WM-FSD) dataset is built using surface smudge samples from a stamping line. Comparative experiments indicate the proposed method surpasses state-of-the-art methods, achieving an average precision (AP) of 44.4 % and an AP50 of 87.0 %, with a 9.3 % recall improvement over the baseline model, while the model's parameters are 28.6 M and the detection speed reaches 44.5 FPS, which can meet the requirements of industrial deployments. Furthermore, the evaluation of the public PCB dataset confirms its generalization ability.}, archive = {J_NEUCOM}, author = {Hao Zhong and Zhongxu Hu and Jie Liu and Haifeng Wang and Youmin Hu and Tielin Shi}, doi = {10.1016/j.neucom.2025.130601}, journal = {Neurocomputing}, month = {9}, pages = {130601}, shortjournal = {Neurocomputing}, title = {A tiny defect detection method on stamped parts with feature aggregation-diffusion and wasserstein distance}, volume = {647}, year = {2025}, }
• (2025). Multistage hierarchical incremental neural network based on stochastic configuration mechanism and dynamic inequality constraint. NEUCOM, 647, 130584. (www)

  @article{ , abstract = {As a class of potent randomized learning models, stochastic configuration networks have shown strong potential in data modeling. However, they still face challenges in terms of computational efficiency and modeling accuracy when modeling complex systems. Hence, this paper proposes a multistage hierarchical incremental neural network (MHINN) and designs a stochastic configuration-based parameter-learning method. A multistage hierarchical modeling strategy is proposed to guide its construction process, including the layer increment stage and aggregation stage. In the first stage, learning layers are constructed layer by layer to obtain feature representations. Then, the refined learning results are obtained in the aggregation stage. To assign appropriate learning parameters, a dynamic inequality constraint is designed for the stochastic configuration learning process, which dynamically adjusts the degree of constraint to reduce computational complexity. Meanwhile, the dynamic inequality constraint ensures the universal approximation property of MHINN. A series of experiments are conducted on several real-world datasets. Experimental results indicate that the MHINN outperforms other classical randomized learning models and deep learning models in terms of computational efficiency and modeling accuracy.}, archive = {J_NEUCOM}, author = {Jie Li and Degang Wang}, doi = {10.1016/j.neucom.2025.130584}, journal = {Neurocomputing}, month = {9}, pages = {130584}, shortjournal = {Neurocomputing}, title = {Multistage hierarchical incremental neural network based on stochastic configuration mechanism and dynamic inequality constraint}, volume = {647}, year = {2025}, }
• (2025). GRIF-PPGNN: Group equality informed ranking-based individual fairness for privacy-preserving graph neural network. NEUCOM, 647, 130583. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) play a crucial role in various high-stakes decision-making scenarios due to their superior learning performance and end-to-end design. However, concerns have arisen about GNNs making biased decisions against sensitive subgroups or individuals and their vulnerability to privacy attacks. To address these issues, researchers have proposed methods to enhance both privacy protection and fairness in GNNs. However, existing methods for promoting ranking-based individual fairness only focus on the ranking list with top-K nodes and the optimization tends to promote the fairness of individuals in the majority subgroups. This leads to drastically different levels of individual fairness among groups. Additionally, while some approaches aim for fair and privacy-preserving GNNs, they fail to protect inference privacy, exposing the graph structure during training and testing. We study a novel problem: promoting group equality-informed ranking-based individual fairness (i.e., ensuring both ranking-based individual fairness and fairness across groups) and defending against edge-stealing attacks without exposing the graph structure. To tackle this, we propose GRIF-PPGNN (Group informed Ranking-based Individual Fairness — Privacy-Preserving Graph Neural Network). Our framework comprises three modules: (i) a privacy-preserving module generating perturbed aggregations, (ii) a utility maximization module using these aggregations to train a classifier for downstream tasks, and (iii) a fairness promotion module optimizing neural network parameters to enhance individual fairness and equalize fairness across groups. Comprehensive experiments on real-world datasets demonstrate that GRIF-PPGNN achieves a good balance between group equality-informed individual fairness and model utility in differential private GNNs.}, archive = {J_NEUCOM}, author = {Xuemin Wang and Yunhui Li and Tianlong Gu and Xuguang Bao and Liang Chang and Guoyong Cai and Tieyuan Liu}, doi = {10.1016/j.neucom.2025.130583}, journal = {Neurocomputing}, month = {9}, pages = {130583}, shortjournal = {Neurocomputing}, title = {GRIF-PPGNN: Group equality informed ranking-based individual fairness for privacy-preserving graph neural network}, volume = {647}, year = {2025}, }
• (2025). FDESI: An open-source web application for visual bioinformatics pipeline design. NEUCOM, 647, 130582. (www)

  @article{ , abstract = {In bioinformatics, data analysis often involves complex workflows that require processing vast amounts of biological data. Pipelines play a crucial role in automating and organizing these sequences of computational tasks, ensuring reproducibility, efficiency, and scalability. In this paper, we presented F LENP DESI gner ( fDESI ), an open-source and user-friendly web application for visual bioinformatics pipeline design. It allows intuitive pipeline creation through a drag-and-drop interface, generating pipeline descriptions in a JSON-based meta-language. Furthermore, it provides an all-in-one environment to overcome the need for external tools, beyond the third-party software within the pipeline itself. Our contribution includes a flexible framework for pipeline modelling, built-in functionalities requiring no programming expertise, an integrated execution engine, and an open-source graphical interface for streamlined bioinformatics workflow design.}, archive = {J_NEUCOM}, author = {Pietro Cinaglia and Mario Cannataro}, doi = {10.1016/j.neucom.2025.130582}, journal = {Neurocomputing}, month = {9}, pages = {130582}, shortjournal = {Neurocomputing}, title = {FDESI: An open-source web application for visual bioinformatics pipeline design}, volume = {647}, year = {2025}, }
• (2025). Sampled-data-based synchronization of matrix-weighted multi-layer complex networks via pinning control. NEUCOM, 647, 130580. (www)

  @article{ , abstract = {This paper examines the synchronization issue for a class of matrix-weighted multi-layer complex networks. Both intra-layer and inter-layer synchronization are investigated by means of sampled-data-based pinning control. In order to better describe the multidimensional attributes of state connection, matrix coupling rather than the conventional scalar weights is introduced to specify the node connections within and between layers. To reduce the waste of communication resources and save the control cost, a sampled-data-based pinning control strategy is designed to reach the intra-layer and inter-layer synchronization of matrix-weighted multi-layer complex networks. To effectively implement the pinning control, it is assumed that for each unpinned node (layer), there exists at least one pinned node (layer) which has a path to the unpinned node (layer). In this context, some sufficient conditions are derived through the utilization of algebraic graph theory, the Lyapunov stability theory, and a modified Halanay inequality. Finally, one simulation example is provided to verify the effectiveness of the obtained theoretical results.}, archive = {J_NEUCOM}, author = {Xiaoli Zhu and Luyang Yu and Yurong Liu}, doi = {10.1016/j.neucom.2025.130580}, journal = {Neurocomputing}, month = {9}, pages = {130580}, shortjournal = {Neurocomputing}, title = {Sampled-data-based synchronization of matrix-weighted multi-layer complex networks via pinning control}, volume = {647}, year = {2025}, }
• (2025). Lightweight verifiable federated learning based on matrix theory. NEUCOM, 647, 130579. (www)

  @article{ , abstract = {Federated learning has been proposed to address modern data privacy and security challenges by enabling a global model to be trained through the sharing of local gradients instead of raw data. In recent years, numerous privacy-preserving and verifiable aggregation schemes for federated learning have been developed. However, these solutions often face limitations, including high computational costs and limited applicability to lightweight Internet of Things (IoT) environments. In this paper, we propose an efficient, lightweight, and verifiable federated learning protection algorithm designed to ensure local data security and resist collusive attacks. Our approach leverages linear secret sharing and blinding techniques to safeguard gradient privacy, while avoiding the computationally expensive bilinear pairing operations. Additionally, the encryption mechanism enables all participating terminals to verify the authenticity of aggregation results and respond appropriately to potential forgeries by the server. Comprehensive security analyses demonstrate that the proposed scheme achieves data confidentiality, authentication, and integrity. Furthermore, experimental evaluations show that the algorithm maintains low computational costs and communication overheads without compromising training accuracy.}, archive = {J_NEUCOM}, author = {Shufen Niu and Yan Zhang and Weifang Wang and Lin Hu and Xingxing Nan}, doi = {10.1016/j.neucom.2025.130579}, journal = {Neurocomputing}, month = {9}, pages = {130579}, shortjournal = {Neurocomputing}, title = {Lightweight verifiable federated learning based on matrix theory}, volume = {647}, year = {2025}, }
• (2025). H∞ control for cooperative multi-agent systems: Event-triggered off-policy reinforcement learning approach. NEUCOM, 647, 130576. (www)

  @article{ , abstract = {This paper investigates the H ∞ synchronization control problem for cooperative multi-agent systems using an event-triggered off-policy reinforcement learning (OffRL) approach. Based on the connection graph from a global optimization perspective, a system model is first constructed to reformulate the problem as solving the Hamilton–Jacobi–Isaacs (HJI) equation. To address the computation and communication challenges associated with extensive information exchanges in multi-agent systems, an event-triggered scheme is introduced. A triggering condition is proposed, and its feasibility is rigorously analyzed in terms of system stability and the exclusion of Zeno behavior. To solve the HJI equation, a model-free OffRL algorithm is derived from the Bellman equation, leveraging system dataset collection to bypass potential inaccuracies in the system dynamic model. Finally, the feasibility and effectiveness of the proposed algorithm are rigorously demonstrated through theoretical analysis and a simulation example.}, archive = {J_NEUCOM}, author = {Hongji Zhuang and Shufan Wu and Vladimir Yu. Razoumny and Yury N. Razoumny}, doi = {10.1016/j.neucom.2025.130576}, journal = {Neurocomputing}, month = {9}, pages = {130576}, shortjournal = {Neurocomputing}, title = {H∞ control for cooperative multi-agent systems: Event-triggered off-policy reinforcement learning approach}, volume = {647}, year = {2025}, }
• (2025). Multi-level aggregation based federated few-shot learning. NEUCOM, 647, 130574. (www)

  @article{ , abstract = {Federated learning (FL) enables clients to train a model collaboratively without data sharing. In real-world applications of FL, data scarcity is a common scenario. Existing research attempts to tackle this issue by combining few-shot learning and FL, which enables clients to learn rich prior knowledge from their local base datasets, thereby accurately predicting new classes with limited training data. However, federated few-shot learning frameworks still face challenges, such as overfitting the training models due to limited local client training data and the difficulty of effectively aggregating prior knowledge caused by non-IID data distribution among clients. We propose a multi-level aggregation based federated few-shot learning framework to address the above issues. By introducing student and teacher levels, the framework achieves fine-grained aggregation. We alleviate the overfitting of client models through a clustering aggregation method at the student level. Moreover, we mitigate the non-IID problem by using a bidirectional communication strategy between the teacher and student levels to achieve knowledge consensus among the models participating in aggregation. Experimental results on benchmark datasets show that our proposed framework outperforms existing methods. Moreover, our approach matches FL-MAML’s time complexity with minimal overall performance degradation, especially in 1-shot settings (average decrease: 0.6%). We also evaluated framework applicability on public datasets, finding robust performance even with unrelated data, consistently outperforming baseline methods. Our experiments demonstrate the effectiveness and efficiency of the proposed framework. It also has high applicability, while various public datasets can provide effective improvements.}, archive = {J_NEUCOM}, author = {Mingyue Li and Hong Zhu and Erxi Wang and Deyu Chen and Meiyi Xie}, doi = {10.1016/j.neucom.2025.130574}, journal = {Neurocomputing}, month = {9}, pages = {130574}, shortjournal = {Neurocomputing}, title = {Multi-level aggregation based federated few-shot learning}, volume = {647}, year = {2025}, }
• (2025). Text-attributed community detection in complex networks through LLMs and GNNs: A powerful fusion of language and graphs. NEUCOM, 647, 130573. (www)

  @article{ , abstract = {This paper introduces a novel framework for community detection in complex networks by considering advanced embedding techniques to integrate textual information associated with nodes and edges with graph structures. We propose a Text-Attributed Graph (TAG) approach, where textual data from nodes and edges, such as book descriptions and user reviews in book recommendation systems, is transformed into semantic embeddings using pre-trained language models (PLMs). Specifically, we employ the latest state-of-the-art embedding models, including E5-Base, variants of BERT (SBERT, DistiBERT, BERT-Base, and BERT-Large), and OpenAI’s text-embedding-3-large , and the cost-effective text-embedding-ada-002 , to enrich graph representations with meaningful contextual features as edge embeddings. These embeddings are integrated into graph neural networks (GNNs), enabling the model to exploit structural and textual contexts to improve community detection performance. The integration of textual embeddings and several GNNs in this manner offers a promising performance for enhancing community detection tasks in complex networks, opening new possibilities for applications in recommendation systems, information retrieval, predictive tasks, and beyond.}, archive = {J_NEUCOM}, author = {Sruthi K.S. and A. Sreekumar and Kannan Balakrishnan}, doi = {10.1016/j.neucom.2025.130573}, journal = {Neurocomputing}, month = {9}, pages = {130573}, shortjournal = {Neurocomputing}, title = {Text-attributed community detection in complex networks through LLMs and GNNs: A powerful fusion of language and graphs}, volume = {647}, year = {2025}, }
• (2025). From continuous pre-training to alignment: A comprehensive toolkit for large language models in federated learning. NEUCOM, 647, 130572. (www)

  @article{ , abstract = {The rapid success of Large Language Models (LLMs) has unlocked vast potential for AI applications in privacy-sensitive domains. However, the traditional centralized training of LLMs poses significant challenges due to privacy concerns regarding collecting sensitive data from diverse sources. This paper offers a promising and privacy-enhancing solution for LLMs: collaboratively training LLMs via Federated Learning (FL) across multiple clients, eliminating the need for raw data transmission. To this end, we present F4LLM, a new and comprehensive toolbox that supports the entire Federated LLM pipeline, from Continuous pre-training to alignment and LLM evaluation. F4LLM employs gRPC as the communication protocol to support various widely-used FL algorithms, ensuring efficient development and benchmarking in geo-distributed FL environments. Moreover, F4LLM offers both open-form and closed-form evaluation options via the efficient inference tool vLLM. The source code and documentation are at here .}, archive = {J_NEUCOM}, author = {Zhuo Zhang and Yukun Zhang and Guanzhong Chen and Lizhen Qu and Xun Zhou and Hui Wang and Zenglin Xu}, doi = {10.1016/j.neucom.2025.130572}, journal = {Neurocomputing}, month = {9}, pages = {130572}, shortjournal = {Neurocomputing}, title = {From continuous pre-training to alignment: A comprehensive toolkit for large language models in federated learning}, volume = {647}, year = {2025}, }
• (2025). Multi-order graph based clustering via dynamical low rank tensor approximation. NEUCOM, 647, 130571. (www)

  @article{ , abstract = {Graph based clustering involves learning a proximity matrix with explicit clustering structure. However, since the limited link inputs and insufficient graph fusion, current works always obtain poor graphs with suboptimal clustering results. To solve the problem, in this paper, we propose a novel multi-order graph based clustering model via dynamic low-rank tensor approximation (MCDLT). Firstly, we use high-order proximity to enrich the link relations for graph inputs. Then a graph selection mechanism and low rank tensor approximation method are used to dynamically learn the consistent information from the complex links in various-order graphs. Finally, we propose a doubly stochastic graph fusion method to directly learn a symmetrical graph that provides clustering results by its connectivity. A novel Augmented Lagrangian Multiplier (ALM) based method is proposed for the sub-problem of doubly stochastic constraints. Experiments shows that our method learns a graph with clearer data structure, achieving SOTA clustering performance and obtaining the GroundTruth for JAFFE data set. Code will be published at https://github.com/NianWang-HJJGCDX/MCDLT.git .}, archive = {J_NEUCOM}, author = {Nian Wang and Zhigao Cui and Aihua Li and Yuanliang Xue and Rong Wang and Feiping Nie}, doi = {10.1016/j.neucom.2025.130571}, journal = {Neurocomputing}, month = {9}, pages = {130571}, shortjournal = {Neurocomputing}, title = {Multi-order graph based clustering via dynamical low rank tensor approximation}, volume = {647}, year = {2025}, }
• (2025). Practical fixed-time bipartite containment for uncertain nonlinear multi-agent systems with switching topology and unknown hysteresis. NEUCOM, 647, 130570. (www)

  @article{ , abstract = {Practical fixed-time bipartite containment control of uncertain nonlinear multi-agent systems (UNMASs) accompanied by switching topologies and unknown hysteresis is investigated in this paper. Firstly, a radial basis function neural network (RBFNN) is used to approximate the unknown nonlinear function of UNMASs, and an adaptive law is designed to adjust the weight parameters online, while a Nussbaum function is introduced to deal with the unknown hysteresis. Secondly, combining the backstepping method with the improved fixed-time stability theory, this paper proposes an adaptive fixed-time control protocol, which ensures that the tracking error converges to the preset tight set in a fixed time and the convergence time is independent of the initial state, so as to realise the fixed-time control. Furthermore, bipartite containment control is aimed and realised taking into account the cooperative–competitive interaction. In addition, numerical examples are enumerated simulated to verify the effectiveness of the control protocol.}, archive = {J_NEUCOM}, author = {Siru li and Jie Wu and Xisheng Zhan and Tao Han and Lingli Cheng}, doi = {10.1016/j.neucom.2025.130570}, journal = {Neurocomputing}, month = {9}, pages = {130570}, shortjournal = {Neurocomputing}, title = {Practical fixed-time bipartite containment for uncertain nonlinear multi-agent systems with switching topology and unknown hysteresis}, volume = {647}, year = {2025}, }
• (2025). Uncertainty-aware contrastive learning for deep clustering. NEUCOM, 647, 130568. (www)

  @article{ , abstract = {Deep clustering aims to group unlabeled data into meaningful clusters by learning discriminative feature representations. However, ambiguous features often lead to noisy representations and inconsistent semantics, limiting improvements in clustering performance. To address this issue, we propose an Uncertainty-Aware Contrastive Learning (UACL) method for deep clustering, which achieves robustness by adaptively restricting the learning of ambiguous features. Specifically, we model pairwise similarity evidence via subjective logic theory, formulating co-cluster probabilities as a Dirichlet distribution to quantify epistemic uncertainty from feature ambiguity. Guided by this uncertainty, we design a dynamic weight-updating strategy that progressively extracts information from potential positives, enhancing the model’s ability to learn discriminative representations and semantically consistent clusters. Furthermore, to enforce attribute consistency, we develop an Attribute Distribution Alignment module that aligns similarity and uncertainty. Extensive experiments on five benchmark datasets demonstrate UACL outperforms current state-of-the-art methods, with an improved ACC of 3.5% for CIFAR-100 and 7.0% for ImageNet-Dogs. The source code is available at: https://github.com/YL616/UACL .}, archive = {J_NEUCOM}, author = {Luyao Chang and Leiting Chen and Chuan Zhou}, doi = {10.1016/j.neucom.2025.130568}, journal = {Neurocomputing}, month = {9}, pages = {130568}, shortjournal = {Neurocomputing}, title = {Uncertainty-aware contrastive learning for deep clustering}, volume = {647}, year = {2025}, }
• (2025). Adaptive multi-view decision model for group recommendation. NEUCOM, 647, 130567. (www)

  @article{ , abstract = {With the increasing richness and diversity of social interactions, group activities are becoming more prevalent. Current personalized recommender systems struggle to accommodate the complexity of multi-user joint decision-making, necessitating in-depth research on group recommendation. Most existing group recommendation models incorporate graph representation learning, which often faces the issue of over-smoothing. Additionally, group recommendation suffers from incomplete representation learning and biased group decision-making. To address these challenges, we propose an A daptive M ulti-view D ecision M odel ( AMDM ) for group recommendation. First, to alleviate the over-smoothing problem, we design a DualFocus Embedding Module, which models from two perspectives. The module bypasses infinite convolutional layers, directly constraining node convergence through loss functions. Second, to overcome the limitations of representation learning, we introduce a Multi-view Representation Module, which comprehensively models groups and items from multiple perspectives. We constrain different views through distribution loss and propose an adaptive fusion component to aggregate intermediate representations. Finally, we design an Individuality-Assisted Group Decision-Making mechanism to address the group decision-making challenge. This mechanism captures commonality features of group members as the subject of preferences and utilizes individual features to assist in modeling group preferences. We evaluate our AMDM on two real-world datasets and experimental results show that our model outperforms state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Ming He and Man Zhou and Changle Li and Yongjie Zheng and Xiaojing Wang}, doi = {10.1016/j.neucom.2025.130567}, journal = {Neurocomputing}, month = {9}, pages = {130567}, shortjournal = {Neurocomputing}, title = {Adaptive multi-view decision model for group recommendation}, volume = {647}, year = {2025}, }
• (2025). LAR-net: An enhanced edge-aware re-detection framework for detecting hazy runways. NEUCOM, 647, 130566. (www)

  @article{ , abstract = {Landing in hazy conditions poses a significant challenge to aircraft vision systems, necessitating precise and reliable detection mechanisms to ensure safety. However, the lack of datasets makes it challenging to design effective methods. Therefore, this paper introduces a unique, real-world Hazy Runway Dataset (HRD), consisting of 28 visible light sequences with clear and hazy weather conditions, as well as 1 infrared sequence. Furthermore, we present the enhanced edge Loss-Aware Re-detection Network (LAR-Net), a novel framework capable of performing both haze removal and hazy runway detection, in this paper to benchmark the performance of HRD. Specifically, LAR-Net integrates an edge-aware loss function to refine detection performance, enhancing the overall accuracy of the HRD dataset. Besides, we incorporate a re-detection module specially designed for the detection of small runway features. Our results on HRD demonstrate that LAR-Net surpasses current state-of-the-art methods, underscoring the effectiveness of LAR-Net in detecting runways under complex weather conditions.}, archive = {J_NEUCOM}, author = {Lichun Yang and Jianghao Wu and Hongguang Li and Chunlei Liu and Shize Wei}, doi = {10.1016/j.neucom.2025.130566}, journal = {Neurocomputing}, month = {9}, pages = {130566}, shortjournal = {Neurocomputing}, title = {LAR-net: An enhanced edge-aware re-detection framework for detecting hazy runways}, volume = {647}, year = {2025}, }
• (2025). Quantum AI for alzheimer’s disease early screening. NEUCOM, 647, 130565. (www)

  @article{ , abstract = {Quantum machine learning is a new research field combining quantum information science and machine learning. Quantum computing technologies appear to be particularly well-suited for addressing problems in the health sector efficiently. They have the potential to handle large datasets more effectively than classical models and offer greater transparency and interpretability for clinicians. Alzheimer’s disease is a neurodegenerative brain disorder that mostly affects elderly people, causing important cognitive impairments. It is the most common cause of dementia and it has an effect on memory, thought, learning abilities and movement control. This type of disease has no cure, consequently an early diagnosis is fundamental for reducing its impact. The analysis of handwriting can be effective for diagnosing, as many researches have conjectured. The DARWIN (Diagnosis AlzheimeR WIth haNdwriting) dataset contains handwriting samples from people affected by Alzheimer’s disease and a group of healthy people. Here we apply quantum AI to this use-case. In particular, we use this dataset to test classical methods for classification and compare their performances with the ones obtained via quantum machine learning methods. We find that quantum methods generally perform better than classical methods. Our results pave the way for future new quantum machine learning applications in early-screening diagnostics in the healthcare domain.}, archive = {J_NEUCOM}, author = {Giacomo Cappiello and Filippo Caruso}, doi = {10.1016/j.neucom.2025.130565}, journal = {Neurocomputing}, month = {9}, pages = {130565}, shortjournal = {Neurocomputing}, title = {Quantum AI for alzheimer’s disease early screening}, volume = {647}, year = {2025}, }
• (2025). Q&A driven conversational recommendation integrating multiple interest modeling. NEUCOM, 647, 130564. (www)

  @article{ , abstract = {The conversational recommendation aims to provide users with recommendations in an interaction form of “System Ask-User Respond”. Existing studies rarely consider combining users’ multi-interest at the attribute level for preference modeling, and conduct inefficient conversational recommendation strategy learning, which affects the recommendation performance and conversation performance. To this end, we proposed a Q-A driven conversational recommendation method integrating multiple interest modeling. Specifically, we integrate the user’s positive and negative feedback to model a session dynamic graph, then use the signed graph convolutional network for graph representation learning, and we model multiple interest sequences based on the attention mechanism, then obtain the user interest state representation by fusing of sequence representations to solve the problem of insufficient user preference representation. Besides, we proposed a personalized decision space optimization method to narrow the range of action candidates and train the model with a multi-agent reinforcement learning method integrating hierarchical decision debiasing to solve the problem of poor conversational recommendation strategy learning effect. Experimental results on three public datasets, LastFM, Yelp, and Book, show that compared with existing conversational recommendation methods, our method demonstrates consistent performance improvement across all datasets. In addition, the results of ablation experiments verify the effectiveness of each component in our method.}, archive = {J_NEUCOM}, author = {Haiping Zhu and Yuqi Sun and Shuwei Che and Yan Chen}, doi = {10.1016/j.neucom.2025.130564}, journal = {Neurocomputing}, month = {9}, pages = {130564}, shortjournal = {Neurocomputing}, title = {Q&A driven conversational recommendation integrating multiple interest modeling}, volume = {647}, year = {2025}, }
• (2025). A semi-supervised wise-attention weighted prototype network for rolling bearing fault diagnosis under noisy and limited labeled data conditions. NEUCOM, 647, 130563. (www)

  @article{ , abstract = {At present, real industrial scenarios face inevitable noise and scarcity of labeled samples, posing a severe challenge to deep learning methods. To address the aforementioned issues, a semi-supervised wise-attention weighted prototype network (SWAWPNet) was proposed for rolling bearing fault diagnosis. Specifically, a multi-scale attention feature extraction module based on depthwise separable convolution and large convolution kernel (DSMLN) was proposed to extract different levels of fault feature information from small samples. Secondly, a prototype update strategy method with wise-attention (WA) weighting was designed based on the differences in class representations. The method generates class weight vectors for prototypes and unlabeled samples through the WA module, refines prototype positions, and enhances attention to key features, reducing the impact of noisy samples on the network model and enhancing the model’s ability to recognize fault categories. Finally, our method validated the effectiveness of SWAWPNet on two publicly available datasets and a simulated experimental platform dataset, demonstrating its satisfactory fault diagnosis ability in addressing the challenge of scarce labeled samples in noisy environments.}, archive = {J_NEUCOM}, author = {Yuanyuan Chen and Jianhai Yue and Zhiming Liu and Jiawei Chen}, doi = {10.1016/j.neucom.2025.130563}, journal = {Neurocomputing}, month = {9}, pages = {130563}, shortjournal = {Neurocomputing}, title = {A semi-supervised wise-attention weighted prototype network for rolling bearing fault diagnosis under noisy and limited labeled data conditions}, volume = {647}, year = {2025}, }
• (2025). Slime mould metaheuristic for optimization and robot path planning. NEUCOM, 647, 130551. (www)

  @article{ , abstract = {Function optimization represents a remarkable challenge in industry and society, aiming to find reasonable solutions – even if they are suboptimal – for everyday problems. Metaheuristics drive the optimization search towards the goals using a specific algorithm inspired by different concepts: from industrial processes to the behaviour of living beings in nature, from mathematical ideas to physics notions. This research proposes a new metaheuristic inspired by the Slime Mould and its foraging behaviours. On the one hand, an exploitation stage mimics the greedy amoeba’s conduct when food is plenty. On the other hand, an exploration stage copies the fruity aggregation of the cells and the subsequent spore dissemination. This study compares the most cited metaheuristics and the Slime Mould Optimization in two different experimentation stages: on the one hand, the optimization of standard benchmarking functions; on the other hand, solving the robot path planning problem. Moreover, a hybridization of the SMO and the WOA is presented, which keeps the SMO’s convergence speed and the WOA’s good performance in finding the best solutions.}, archive = {J_NEUCOM}, author = {Enol García and José R. Villar and Javier Sedano and Camelia Chira and Enrique de la Cal and Luciano Sánchez}, doi = {10.1016/j.neucom.2025.130551}, journal = {Neurocomputing}, month = {9}, pages = {130551}, shortjournal = {Neurocomputing}, title = {Slime mould metaheuristic for optimization and robot path planning}, volume = {647}, year = {2025}, }
• (2025). Searching efficient network with lightweight optimization for image super-resolution. NEUCOM, 647, 130550. (www)

  @article{ , abstract = {Despite rapid advances in deep learning-based image super-resolution (SR), deploying them in resource-constrained scenarios (e.g., mobile devices) has remained a challenging task. Generally speaking, good performance of SR networks comes at the price of increased computational complexity and model parameters. To strike an improved trade-off, recent studies have considered the design of efficient SR networks. However, most existing methods only consider the network with fewer parameters and calculations, overlooking the actual inference latency time of the network on mobile devices. To meet this challenge of supporting real-time applications, a neural architecture search (NAS) method is developed to find an efficient SR network in this paper, where search space is designed, including efficient operations and lightweight attention blocks. Besides, a novel search method for channel numbers is proposed, called filter grouping search. And a subnet sampling-updating method is designed to accelerate the search. Finally, the weight pruning and adaptive weighted uncertainty-driven loss combined with contrastive distillation are used to optimize the final target network. The experimental results show that our method is superior to the existing state-of-the-art methods. Additionally, the proposed method is also tested on mobile devices and found that it has more advantages on the Kirin 990 5G ARM CPU, supporting its potential in mobile computing. The code is available at https://github.com/douzaikongcheng/EMNAS .}, archive = {J_NEUCOM}, author = {Ruina Shen and Zhangheng Peng and Le Dong and Jun Wang and Weisheng Dong and Guangming Shi}, doi = {10.1016/j.neucom.2025.130550}, journal = {Neurocomputing}, month = {9}, pages = {130550}, shortjournal = {Neurocomputing}, title = {Searching efficient network with lightweight optimization for image super-resolution}, volume = {647}, year = {2025}, }
• (2025). Memristor-based multilayer neural network with edge learning and weight update. NEUCOM, 647, 130549. (www)

  @article{ , abstract = {Memristor crossbar array (MCA) is a promising computing-in-memory architecture for accelerating vector-matrix multiplication (VMM). However, existing memristor-based circuits primarily perform forward computations without local learning capabilities. To address the issue, a memristor-based multilayer neural network (MNN) with edge learning is proposed, enabling error backpropagation and one-step weight update. In particular, the overall circuit design of a three-layer MNN is outlined, comprising three operational stages: forward computation, backpropagation, and weight update. Subsequently, the schematics of four adopted circuit blocks are detailed. By integrating the memory and weight update blocks with MCA, the proposed three-layer memristor-based MNN exhibits scalability and universality, allowing it to be extended to the deep neural network (DNN) architecture. Circuit simulations validate the effectiveness of the proposed scheme, while offline-online co-simulations demonstrate its strength in compensating for variations and non-ideal characteristics of physical devices. Comparative results further highlight the advantages of the proposed scheme in realizing full-analog edge learning circuits and performing error backpropagation processes efficiently.}, archive = {J_NEUCOM}, author = {Ningye Jiang and Jupeng Xie and Mingxuan Jiang and Haoen Huang and Zhigang Zeng}, doi = {10.1016/j.neucom.2025.130549}, journal = {Neurocomputing}, month = {9}, pages = {130549}, shortjournal = {Neurocomputing}, title = {Memristor-based multilayer neural network with edge learning and weight update}, volume = {647}, year = {2025}, }
• (2025). Stochastic configuration networks with robust supervised least squares regression. NEUCOM, 647, 130548. (www)

  @article{ , abstract = {Stochastic Configuration Networks (SCNs) are widely used in regression modeling to fit data distributions due to their fast convergence and powerful learning capabilities. However, in practical industrial applications, labeled outliers in the dataset inevitably appear, which can destroy the internal structure and connection of the data, and not only seriously reduce the prediction accuracy of SCNs, but may also increase the risk of overfitting. To cope with this problem, this paper proposes a new algorithm - Robust Supervised Stochastic Configuration Networks (RSSCNs). In this paper, firstly, the loss function is constructed based on the concave function of the error, which extends the applicability of the robust supervised least squares regression. Second, robust supervised least squares regression is introduced into the SCN algorithm for the first time, which enhances the robustness of the model and reduces the risk of model overfitting. Finally, the inequality constraint of the SCN algorithm is reconstructed according to the loss function, which optimizes the parameter selection and accelerates the convergence of the model. In the experimental section, RSSCNs are applied to a function fitting experiment, several benchmark datasets, and an industrial dataset from a dense medium coal preparation process, and the performance is compared with other algorithms. The experimental results show that for these datasets, RSSCNs perform favorably in terms of accuracy and robustness, validating the effectiveness of the algorithm.}, archive = {J_NEUCOM}, author = {Fei Chu and Zihang Sun and Yu Wang and Yong Zhang}, doi = {10.1016/j.neucom.2025.130548}, journal = {Neurocomputing}, month = {9}, pages = {130548}, shortjournal = {Neurocomputing}, title = {Stochastic configuration networks with robust supervised least squares regression}, volume = {647}, year = {2025}, }
• (2025). From hippocampal neurons to broad spiking neural networks. NEUCOM, 647, 130547. (www)

  @article{ , abstract = {Spiking Neural Network (SNN) provides an irreplaceable mechanism for series prediction, particularly in scenarios where time and computing resources are both critical, exhibiting superiority over other deep learning models with sparse and spike-based communication inherent to SNN methods. However, as the depth of the network increases, a critical challenge arises: excessive diffusion of signals in the deeper network can seriously affect its accuracy and efficiency. Inspired by the organizational principles of hippocampal circuits, this study proposes a broad learning framework inspired by the brain ( B 2 L ), offering an alternative SNN architecture to mitigate the problem of excessive diffusion. It adopts a broad, incrementally extendable structure while adopting sparse coding for data representation, dedicated to: (1) Construction of the B 2 L SNN : Combining preliminary feature extraction with enhancement processing, a dual-layer broad structure establishes random vector functional-link (RVFL) mapping between rate-coded spike sequences and output labels, effectively preventing excessive diffusion to upstream neurons; (2) Extendable Optimization of B 2 L SNN : Drawing from the concept of incremental learning, B 2 L SNN employs a weight optimization method to search for optimal configurations by expanding RVFL mappings from initial architecture to static datasets through incremental neuron integration. Experiments across six benchmarks (MNIST, CIFAR-10, CIFAR-100, NORB, N-MNIST, and BCI IV 2b) demonstrate: (1) B 2 L SNN achieves 20.3×faster training time than four SNN variants (4058 seconds → 197 s) via random vector functional-link mapping; (2) 96.97% NORB accuracy (+3.64% over SNN2ANN) with 2.79 ms inference time (3.06×speedup), also achieving SOTA 89.52% accuracy on BCI IV 2b and leading N-MNIST inference speed (3.85ms); (3) Scalability: With 1.5×expansion in feature layers (PF: 60 → 100, EF: 7k → 11k), B 2 L SNN achieves 1.04% accuracy improvement on MNIST alongside 12.3% faster training than BLS baselines.}, archive = {J_NEUCOM}, author = {Yaodong Wang and Yiping Zuo and Dan Chen and Weiping Tu and Albert Y. Zomaya and Xiaoli Li}, doi = {10.1016/j.neucom.2025.130547}, journal = {Neurocomputing}, month = {9}, pages = {130547}, shortjournal = {Neurocomputing}, title = {From hippocampal neurons to broad spiking neural networks}, volume = {647}, year = {2025}, }
• (2025). Infrared small target detection using the global low-rank and local smoothness coupled representation with local structure. NEUCOM, 647, 130546. (www)

  @article{ , abstract = {Infrared small target detection is crucial in both military and civilian applications. However, existing low-rank and sparse decomposition (LRSD) methods often suffer from noise residues caused by inaccurate background estimation. This is because the low-rank and smoothness of the background exhibit inherent coupling properties. It is usually difficult to accurately fit complicated background using a single regularization term or their additive hybrid model. This paper tackles this issue by proposing a coupled tensor model that incorporates global low-rank and local smoothness. Furthermore, to suppress potential “grid artifacts”, which are usually brought on by the infrared device’s pixel array characteristics and total variation, another regularization term that focuses on the minimum absolute structure of the tensor’s gradient in the local region is constructed. The proposed model is then solved using an optimization framework based on the alternating direction method of multipliers (ADMM). Finally, comparative experiments on three public datasets demonstrate that the proposed model outperforms existing state-of-the-art LRSD methods in terms of suppressing complicated background and sparse “grid artifacts”.}, archive = {J_NEUCOM}, author = {Junying Li and Xiaorong Hou and Yajian Zeng}, doi = {10.1016/j.neucom.2025.130546}, journal = {Neurocomputing}, month = {9}, pages = {130546}, shortjournal = {Neurocomputing}, title = {Infrared small target detection using the global low-rank and local smoothness coupled representation with local structure}, volume = {647}, year = {2025}, }
• (2025). HRSGD: High-order recurrent medical image fusion via self-supervised generative distillation. NEUCOM, 647, 130545. (www)

  @article{ , abstract = {Medical image fusion (MMIF) is essential for enhancing the precision of clinical diagnoses. Self-supervised knowledge distillation (SSL-KD) aims to address the limitation of poor performance of self-supervised learning on small models. However, existing methods rely on student networks that imitate the output of the teacher network, which cannot achieve effective knowledge transfer when network structures differ. This study introduces a self-supervised generative distillation framework (HRSGD) for MMIF that enhances knowledge transfer across different structures by incorporating structural prior and mask generation mechanisms. HRSGD undergoes three-stage training with three distinct networks. Firstly, the teacher network conducts self-supervised training to transfer multi-resolution semantic information to the following stages. Subsequently, we introduce an implicit multi-resolution block (IMB) in the feature extraction of the student network to obtain the prior knowledge of the source image and aid in restoring multi-resolution semantics during transmission. By employing masked generative distillation (MGD), the student network restores the complete features of the teacher from the partial pixels of the student through distillation training, which focuses on key semantic information and avoids blind imitations. Finally, the recovered features of the student network utilize a high-order recurrent cross-attention (HRCA) during fusion training, which captures the deep dependencies between modalities through continuous high-order semantic interaction. The fusion experiment for qualitative analysis and quantitative analysis indicates that our method achieves efficient knowledge transfer.}, archive = {J_NEUCOM}, author = {Xiaoyu Li and Rencan Nie and Jinde Cao and Kaixiong Qing and Jiang Zuo and Wenmiao Shi}, doi = {10.1016/j.neucom.2025.130545}, journal = {Neurocomputing}, month = {9}, pages = {130545}, shortjournal = {Neurocomputing}, title = {HRSGD: High-order recurrent medical image fusion via self-supervised generative distillation}, volume = {647}, year = {2025}, }
• (2025). RFBLS: A robust rough fuzzy broad learning system with local neighborhood structure. NEUCOM, 647, 130544. (www)

  @article{ , abstract = {As an extension of random vector functional link neural network (RVFLNN), the broad learning system (BLS) has gradually been applied in various fields due to its advantages of simple network structure, rapid training, and fast model updating. Despite its remarkable achievements, there are still some difficulties and challenges that restrict its further application in practical tasks. In BLS, all samples equally contribute to constructing the optimal classification boundary. However, when training samples contain noise or outliers, it will result in the obtained classification boundary not being the true optimal classification boundary. In other words, BLS is sensitive to noise points or outliers. To address it, we propose a robust rough fuzzy broad learning system with local neighborhood structure (RFBLS). In RFBLS, fuzzy sets are utilized by associating a membership degree with each sample. This incorporation ensures diverse contributions of different samples during the minimization of the objective function. Samples with noise or outliers are assigned smaller weights, effectively eliminating the influence of noise and outlier samples in the learning process. We employ the rough set theory to explore the influence-decision degree of each conditional attribute on the decision attribute, aiming to eliminate the impact of redundant or interfering features. Subsequently, we comprehensively consider the distance between samples and class centers, as well as the local neighborhood structure relationship between samples. Initially, we calculate the initial membership degree of each sample based on the distance to class centers. Then, utilizing fuzzy K-nearest neighbors, we calculate and rank the membership degrees of K nearest neighbors for each sample. These rankings are then weighted with the initial membership degree, achieving the fusion of relationships and membership degree calculations between samples. Finally, extensive comparative experiments are conducted on various UCI datasets and image datasets, and the results indicate that our method exhibits superior noise resistance and classification performance.}, archive = {J_NEUCOM}, author = {Yonghe Chu and Yanlong Guo and Peng Li and Weiping Ding and Witold Pedrycz and Heling Cao}, doi = {10.1016/j.neucom.2025.130544}, journal = {Neurocomputing}, month = {9}, pages = {130544}, shortjournal = {Neurocomputing}, title = {RFBLS: A robust rough fuzzy broad learning system with local neighborhood structure}, volume = {647}, year = {2025}, }
• (2025). Heterogeneous agents trajectory prediction with dynamic interaction relational reasoning. NEUCOM, 647, 130543. (www)

  @article{ , abstract = {Accurate trajectory prediction for different types of agents in complex environments is crucial for enabling safe navigation planning. However, current trajectory prediction methods usually ignore the fact that the behaviors of different types of traffic participants exhibit first-order discontinuities. For example, at intersections, the movement behavior can abruptly shift between stopping, going straight, turning right, or turning left, due to the frequent interactions occurring between the agents and the constraints imposed by traffic rules. Their behavior is directly affected by their interactions with the surrounding agents and the environment. To address these challenges, we propose a trajectory prediction method for heterogeneous agents with dynamic interaction relational reasoning. We utilize type-specific encoders to extract dynamic features of agents from their historical states. Interactions between heterogeneous agents are abstracted as heterogeneous graphs with directed edge features, then processed by a novel graph attention network with dynamic relational reasoning designed to extract spatio-temporal interaction features. To capture dynamic interactions, the graph is evolved into a topologically and representationally dynamic graph. Spatial interaction discontinuities are handled by reconstructing subgraphs for different agents with dynamic and changeable features. Furthermore, to reason about interactions, a two-element relational representation is proposed to obtain dynamic relational reasoning. Finally, we conduct a validation test of the proposed model utilizing real-world datasets. The experimental results from different aspects demonstrate that our method can effectively capture the dynamic interaction features between heterogeneous agents, realize the trajectory prediction of heterogeneous agents, and achieve state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Nianwen Ning and Shihan Tian and Hengji Li and Wei Li and Chang Liu and Yi Zhou and XiaoZhi Gao}, doi = {10.1016/j.neucom.2025.130543}, journal = {Neurocomputing}, month = {9}, pages = {130543}, shortjournal = {Neurocomputing}, title = {Heterogeneous agents trajectory prediction with dynamic interaction relational reasoning}, volume = {647}, year = {2025}, }
• (2025). A collaborative content moderation framework for toxicity detection based on multitask neural networks and conformal estimates of annotation disagreement. NEUCOM, 647, 130542. (www)

  @article{ , abstract = {Content moderation typically combines the efforts of human moderators and machine learning models. However, these systems often rely on data where significant disagreement occurs during moderation, reflecting the subjective nature of toxicity perception. Rather than dismissing this disagreement as noise, we interpret it as a valuable signal that highlights the inherent ambiguity of the content—an insight missed when only the majority label is considered. In this work, we introduce a novel content moderation framework that emphasizes the importance of capturing annotation disagreement. In this work, we propose a novel content moderation framework that prioritizes capturing annotation disagreement. Our approach leverages multitask neural networks with transformer architectures as their backbone, where toxicity classification serves as the primary task and annotation disagreement is modelled as an auxiliary task. By framing disagreement as a predictive problem within the multitask learning architecture, our method effectively captures the nuanced ambiguity of content toxicity. Additionally, we leverage uncertainty estimation techniques, specifically Conformal Prediction, to account for the model’s inherent uncertainty in predicting toxicity and annotation disagreement. The framework also allows moderators to adjust thresholds for annotation disagreement, offering flexibility in determining when ambiguity should trigger a review. We demonstrate that our joint approach enhances model performance, calibration, and uncertainty estimation, while offering greater parameter efficiency and improving the review process in comparison to single-task methods.}, archive = {J_NEUCOM}, author = {Guillermo Villate-Castillo and Javier Del Ser and Borja Sanz}, doi = {10.1016/j.neucom.2025.130542}, journal = {Neurocomputing}, month = {9}, pages = {130542}, shortjournal = {Neurocomputing}, title = {A collaborative content moderation framework for toxicity detection based on multitask neural networks and conformal estimates of annotation disagreement}, volume = {647}, year = {2025}, }
• (2025). Few-shot action recognition with temporal feature enhancement in fine-tuning. NEUCOM, 647, 130541. (www)

  @article{ , abstract = {In the few-shot video action recognition task, parameter-efficient fine-tuning methods can mitigate the risk of overfitting caused by the limited training samples. However, existing frameworks that utilize these methods directly on pre-trained models still need to be improved to enhance the temporal feature extraction capabilities of the model. Additionally, the global property of the attention mechanism may diminish high-frequency information critical for describing motion characteristics of actions. To address these issues, we propose a novel few-shot video action recognition model. The model mainly perform more comprehensive parameter-efficient fine-tuning of the pre-trained image model through the STSKLoRA and SWDST-Adapter modules. The STSKLoRA module employs a multi-path aggregation architecture. By mapping the fitted 3D weight changes to the 2D level to act as the weight changes of the model, this module can enhance the ability of the model to reason about temporal information. The SWDST-Adapter module fuses the stationary wavelet transform in the up- and down-sampling structures of the adapter so that this module can use information from different frequency bands to represent video actions accurately. We conduct experiments on four public video action recognition datasets and achieve the accuracy of 79.1% (HMDB51), 97.5% (UCF101), 88.2% (Kinetics), and 75.2% (SSv2) for 5-way 5-shot, respectively, which demonstrate the superiority of our method compared to state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Zhiwei Xie and Yanxiang Gong and Feiyang Sun and Mei Xie and Xin Ma}, doi = {10.1016/j.neucom.2025.130541}, journal = {Neurocomputing}, month = {9}, pages = {130541}, shortjournal = {Neurocomputing}, title = {Few-shot action recognition with temporal feature enhancement in fine-tuning}, volume = {647}, year = {2025}, }
• (2025). NeuroTorch: DRAM-free nonvolatile memory-based hybrid training compute-in-memory system simulator. NEUCOM, 647, 130539. (www)

  @article{ , abstract = {We present NeuroTorch, a simulator for nonvolatile memory-based compute-in-memory. It integrates the parallel outer product update and hybrid training, eliminating the need for weight gradient computation units and DRAM. Conventional on-chip training using weight gradient computation units and DRAM does not fully leverage the analog capabilities of for nonvolatile memory-based compute-in-memory during weight gradient computation and weight update phases. NeuroTorch assumes minimized usage of additional hardware modules for feed-forward, backpropagation, weight gradient computation, and weight update phases, offering insights into new, energy-efficient compute-in-memory. NeuroTorch shows that the new compute-in-memory approach improves energy efficiency by over 20 times compared to that of conventional on-chip training. (Code: https://github.com/SMDLGITHUB/NeuroTorch )}, archive = {J_NEUCOM}, author = {Jonghyun Ko and Inseok Lee and Kyu-Ho Lee and Joon Hwang and Min-Kyu Park and Jiseong Im and Jeonghyun Kim and Ryun-Han Koo and Sung-Ho Park and Jae-Joon Kim and Jong-Ho Lee}, doi = {10.1016/j.neucom.2025.130539}, journal = {Neurocomputing}, month = {9}, pages = {130539}, shortjournal = {Neurocomputing}, title = {NeuroTorch: DRAM-free nonvolatile memory-based hybrid training compute-in-memory system simulator}, volume = {647}, year = {2025}, }
• (2025). Bayesian deep multi-instance learning for student performance prediction based on campus big data. NEUCOM, 647, 130538. (www)

  @article{ , abstract = {Problem: Predicting student performance using campus big data and deep learning methods has emerged as a promising alternative to traditional psychological assessments, which are often delayed and subjective. However, existing deep learning approaches face significant challenges, including the sparsity of campus big data and susceptibility to overfitting. Objective: To address these issues, this study aims to develop a robust and effective framework for predicting student performance by overcoming data sparsity and overfitting problems. Method: We propose a novel Bayesian Deep Multi-Instance Learning (Bay-DeepMIL) framework. The method integrates multi-instance learning (MIL) to handle data sparsity and incorporates a Bayesian framework to treat network parameters as probabilistic distributions, enhancing robustness and mitigating overfitting. Additionally, a Bayesian multi-head attention mechanism is introduced to dynamically assign importance to different instances, improving the extraction of key information from multi-instance data. Results: Extensive experiments demonstrate that Bay-DeepMIL outperforms state-of-the-art methods in prediction accuracy. The framework also provides uncertainty estimates for each prediction, offering valuable confidence measures and decision support for educational stakeholders. Conclusion: The proposed Bay-DeepMIL framework not only advances the technical capabilities of predictive models but also provides practical tools for enhancing educational decision-making. This study underscores the effectiveness of integrating Bayesian inference with multi-instance learning to address core challenges such as data sparsity and model overfitting in student performance prediction.}, archive = {J_NEUCOM}, author = {Jiayang Huang and Keyi Yang and Quan Wang and Pengfei Yang and Ziling Ruan and Jiaxi Wang and Zhi-Qiang Zhang}, doi = {10.1016/j.neucom.2025.130538}, journal = {Neurocomputing}, month = {9}, pages = {130538}, shortjournal = {Neurocomputing}, title = {Bayesian deep multi-instance learning for student performance prediction based on campus big data}, volume = {647}, year = {2025}, }
• (2025). Finite-time synchronization of delayed quaternion-valued neural networks by non-decomposition and improved two-step methods. NEUCOM, 647, 130537. (www)

  @article{ , abstract = {This paper explores the finite-time (F-T) synchronization of quaternion-valued neural networks (QVNNs) with bounded time-varying delay. To simplify the theoretical derivation, the delayed QVNNs are addressed without any decomposition by introducing 1-norm, 2-norm and ∞ -norm of quaternion. To handle time-delay term, the two-step method is employed. In the first step, a generalized Halanay-type inequality is utilized to converge the system synchronization error decreases from initial values V ( t 0 ) to δ V ( t 0 ) , where 0 < δ ≤ 1 is a free parameter. This technique can remove the constraint of differentiability of time delays and avoid the construction of complex Lyapunov functionals. In the second step, by using existing F-T stability theorems, the synchronization error is further decreased to 0. Furthermore, a set of F-T synchronization criteria are acquired to achieve the F-T synchronization between the considered system. Eventually, two numerical examples are derived to demonstrate the practicability of the obtained results.}, archive = {J_NEUCOM}, author = {Jingjing Wang and Song Zhu}, doi = {10.1016/j.neucom.2025.130537}, journal = {Neurocomputing}, month = {9}, pages = {130537}, shortjournal = {Neurocomputing}, title = {Finite-time synchronization of delayed quaternion-valued neural networks by non-decomposition and improved two-step methods}, volume = {647}, year = {2025}, }
• (2025). Verifying ambiguous claims with a reasoner-translator framework. NEUCOM, 647, 130536. (www)

  @article{ , abstract = {Fact-checking is a task that verifies the claim’s veracity based on evidence. During fact-checking, the ambiguity in natural language often poses a significant challenge, which leads to disagreement among fact-checkers even when presented with the same evidence (e.g., simultaneously determining the veracity label as Support and Neutral). Tracing where and how ambiguity occurs is crucial for resolving these disagreements, yet this area remains unexplored. In response, we introduce the entailment tree, a novel tree-structured representation that traces ambiguity in the fact-checking process. Specifically, we represent the evidence-based entailment reasoning process as a tree with premise and conclusion nodes, with ambiguity captured through “forked nodes”. A shared conclusion highlights the ambiguous point in forked nodes, while diverging premise branches reflect different interpretations. To automate the construction of entailment trees, we propose a Reasoner-Translator framework based on Large Language Models (LLMs). This framework operates in two stages: a Reasoner conducts Chain-of-Thought (CoT) reasoning, and a Translator translates the reasoning into the corresponding entailment tree. We evaluate our framework on the AmbiFC dataset. Compared to baseline methods that directly produce trees, the Reasoner-Translator framework achieves better reasoning performance and generates higher-quality trees. Additionally, we present a linguistic analysis of forked nodes in the case study to benefit further research.}, archive = {J_NEUCOM}, author = {Xiucheng Lyu and Mingwei Sun and Chengyu Cao and Bin Liang and Ruifeng Xu}, doi = {10.1016/j.neucom.2025.130536}, journal = {Neurocomputing}, month = {9}, pages = {130536}, shortjournal = {Neurocomputing}, title = {Verifying ambiguous claims with a reasoner-translator framework}, volume = {647}, year = {2025}, }
• (2025). A survey of few-shot relation extraction combining meta-learning with prompt learning. NEUCOM, 647, 130534. (www)

  @article{ , abstract = {Few-shot Relation Extraction (FSRE) is challenging in Natural Language Processing (NLP). It requires models to accurately identify and categorize the relation between entities in texts based on limited annotated samples. Researchers have explored meta-learning and prompt-learning methods to address this challenge. Meta-learning aims to accelerate the model’s learning rate for new tasks by training across different tasks, while prompt learning seeks to enhance the model’s understanding of context. However, both methods face limitations in few-shot scenarios. To overcome these restrictions, studies combining meta-learning with prompt learning have emerged, garnering widespread academic interest. This paper presents a comprehensive survey of FSRE combining Meta-Learning with Prompt Learning. We first categorize existing research into three main groups: meta-learning-based, prompt learning-based, and hybrid approaches. Then, we provide a detailed analysis and discussion of the latest advancements in each category. Moreover, we extensively describe datasets commonly used in the FSRE task, evaluation metrics, and comparisons among various models. Finally, we explore current challenges in the field and forecast potential future research trends, offering insights for subsequent studies.}, archive = {J_NEUCOM}, author = {Tingting Hang and Wei Wu and Jun Feng and Hamza Djigal and Jun Huang}, doi = {10.1016/j.neucom.2025.130534}, journal = {Neurocomputing}, month = {9}, pages = {130534}, shortjournal = {Neurocomputing}, title = {A survey of few-shot relation extraction combining meta-learning with prompt learning}, volume = {647}, year = {2025}, }
• (2025). AVS-net: Point sampling with adaptive voxel size for 3D scene understanding. NEUCOM, 647, 130533. (www)

  @article{ , abstract = {The recent advancements in point cloud learning have enabled intelligent vehicles and robots to comprehend 3D environments better. However, processing large-scale 3D scenes remains a challenging problem, such that efficient downsampling methods play a crucial role in point cloud learning. Existing downsampling methods either require a huge computational burden or sacrifice fine-grained geometric information. For such purpose, this paper presents an advanced sampler that achieves both high accuracy and efficiency. The proposed method utilizes voxel centroid sampling as a foundation but effectively addresses the challenges regarding voxel size determination and the preservation of critical geometric cues. Specifically, we propose a Voxel Adaptation Module that adaptively adjusts voxel sizes with the reference of point-based downsampling ratio. This ensures that the sampling results exhibit a favorable distribution for comprehending various 3D objects or scenes. Meanwhile, we introduce a network compatible with arbitrary voxel sizes for sampling and feature extraction while maintaining high efficiency. The proposed approach is demonstrated with 3D object detection and 3D semantic segmentation. Compared to existing state-of-the-art methods, our approach achieves better accuracy on outdoor and indoor large-scale datasets, e.g. Waymo and ScanNet, with promising efficiency. Code is available at https://github.com/yhc2021/GPR .}, archive = {J_NEUCOM}, author = {Hongcheng Yang and Dingkang Liang and Dingyuan Zhang and Zhe Liu and Zhikang Zou and Xingyu Jiang and Yingying Zhu}, doi = {10.1016/j.neucom.2025.130533}, journal = {Neurocomputing}, month = {9}, pages = {130533}, shortjournal = {Neurocomputing}, title = {AVS-net: Point sampling with adaptive voxel size for 3D scene understanding}, volume = {647}, year = {2025}, }
• (2025). Global and diverse ensemble model for regression. NEUCOM, 647, 130520. (www)

  @article{ , abstract = {Diversity is a fundamental component in ensemble methods, crucial for enhancing the overall performance and robustness of predictive models. In bagging and boosting, diversity is implicitly generated through the data sampling process. In stacking, diversity is introduced by incorporating heterogeneous machine learning models as base learners, and an error function is created to focus on minimizing the meta-learner’s overall errors. Some models promote diversity directly within the error function; however, they often prioritize generating competitive and diverse individual learners, neglecting the necessity of creating an ensemble that is collectively accurate, as seen in stacking, whilst ensuring diversity among its members. Motivated by this point, two ensemble models are proposed in this manuscript, named Global and Diverse Ensemble Methods. These models incorporate implicit diversity through data, diversity in the heterogeneity of base learners, and an error function designed to produce a competitive overall ensemble with diverse individuals. The two diversity proposals included in the models are negative correlation and squared Pearson correlation. In both cases, the error function incorporates the minimization of the ensemble’s overall error (global error measure) in addition to promoting diversity. The proposed methods have been rigorously tested on 45 publicly available regression datasets using shallow base learners, as well as on 7 additional datasets using deep base learners, yielding very promising results. These findings underscore the importance of integrating these diversity-promoting elements and minimizing the global errors of the ensemble, rather than focusing solely on the errors of individual base learners, in the design of ensemble methods.}, archive = {J_NEUCOM}, author = {Antonio M. Durán-Rosal and Thomas Ashley and Javier Pérez-Rodríguez and Francisco Fernández-Navarro}, doi = {10.1016/j.neucom.2025.130520}, journal = {Neurocomputing}, month = {9}, pages = {130520}, shortjournal = {Neurocomputing}, title = {Global and diverse ensemble model for regression}, volume = {647}, year = {2025}, }
• (2025). A dual medical ontology and relational graph framework with neural ordinary differential equations for diagnostic prediction. NEUCOM, 647, 130519. (www)

  @article{ , abstract = {Accurate diagnostic prediction from Electronic Health Records (EHRs) is critical for advancing clinical decision-making and improving patient outcomes. Current works often struggle to systematically integrate medical domain knowledge, leverage unstructured data effectively, and model irregular temporal patterns in patient encounters. To address these limitations, we propose a dual medical ontology and relational graph framework with neural ordinary differential equations (DORI) for diagnostic prediction. DORI integrates hierarchical medical ontology structures and disease co-occurrence relationships to refine medical code embeddings, employs graph-based relational learning to capture intricate disease interactions, and simulates the continuous evolution of patient health states using neural ordinary differential equations (ODEs). Empirical evaluations on the MIMIC-III and MIMIC-IV datasets show that DORI achieves state-of-the-art performance in weighted F 1 score and recall, confirming its enhanced ability to process irregular time intervals and capture complex disease relationships. Our framework increases interpretability by explicitly modeling multi-source medical knowledge and dynamic patient states, underscoring its potential for real-world clinical implementation. These results further underscore the importance of integrating multi-source medical knowledge and dynamic patient states for the precise prediction and interpretation of health events. Our code is released on https://github.com/1245505490/DORI .}, archive = {J_NEUCOM}, author = {Shibo Li and Fanglin Dong and Runze Li and Weihua Li}, doi = {10.1016/j.neucom.2025.130519}, journal = {Neurocomputing}, month = {9}, pages = {130519}, shortjournal = {Neurocomputing}, title = {A dual medical ontology and relational graph framework with neural ordinary differential equations for diagnostic prediction}, volume = {647}, year = {2025}, }
• (2025). Attaining hardware-efficient inference in memristor-based transformer accelerator via network redesign. NEUCOM, 647, 130517. (www)

  @article{ , abstract = {Transformer Network (TN) has achieved remarkable success in artificial intelligence, largely attributed to its formidable attention mechanism performance. Memristive neuromorphic computing architectures have been employed for hardware acceleration of the TN to alleviate the enormous hardware resource consumption caused by the extensive vector-matrix multiplication (VMM). However, the attention mechanism in the TN entails numerous dynamic VMM, which is unsuitable for memristor crossbars acceleration and requires additional multipliers to perform the multiplication and accumulation operations. Consequently, the memristor-based TN suffers from significant peripheral circuit overhead, progressively becoming the principal bottleneck as the network scale increases. To address these challenges, we redesign the computational core of the memristor-based TN by integrating the retention mechanism into the memristive neuromorphic computing system and propose the Memristor-based Retentive Network (M-RN), which enables recurrent inference, thereby improving computational efficiency and reducing resource consumption. Furthermore, we present a fully analog circuit implementation of M-RN and validate the proposed system’s efficacy, robustness, and scalability through simulation analysis. In character recognition experiments, the results demonstrate that compared to a memristor-based TN, M-RN achieves significant improvements of up to 2 × , 73 . 92 × , and 11 . 24 × in terms of latency, energy consumption, and area overhead, while maintaining comparable or even superior inference accuracy. Notably, these improvements become more pronounced when applied to larger-scale networks, positioning M-RN as one of the most promising solutions for hardware acceleration of large-scale neural networks.}, archive = {J_NEUCOM}, author = {Junzhe Xu and Haoqin Hong and Yue Zhou and Xiaofang Hu and Shukai Duan}, doi = {10.1016/j.neucom.2025.130517}, journal = {Neurocomputing}, month = {9}, pages = {130517}, shortjournal = {Neurocomputing}, title = {Attaining hardware-efficient inference in memristor-based transformer accelerator via network redesign}, volume = {647}, year = {2025}, }
• (2025). From detection to segmentation: Combining foundation model with bounding boxes for efficient fire detection. NEUCOM, 647, 130513. (www)

  @article{ , abstract = {In the context of fire detection in industries, segmentation has been found to be particularly effective considering very dynamic nature and small size of fires. However, annotation of the pixel-level segmentation mask is costly, especially when labeling large-scale industrial datasets. To address this issue, we have proposed the SAM2Segment approach that leverages foundation models to train a segmentation network without spending the expensive effort of annotating pixel-level mask annotations. The approach combines the computationally efficient bounding box annotation with foundation models to generate the segmentation masks, thereby outperforming detection-based state-of-the-art methods with the same effort of annotating boxes. To the best of our knowledge, SAM2Segment is the first work to incorporate a foundation model to generate annotations and train a segmentation model for a fire detection use-case. Extensive experiments validate the effectiveness of our approach, achieving the superior fire detection performance on the industrial and publicly available datasets across established evaluation metrics.}, archive = {J_NEUCOM}, author = {Keyur Joshi and Xin Li and Tjark Windisch and Markus König}, doi = {10.1016/j.neucom.2025.130513}, journal = {Neurocomputing}, month = {9}, pages = {130513}, shortjournal = {Neurocomputing}, title = {From detection to segmentation: Combining foundation model with bounding boxes for efficient fire detection}, volume = {647}, year = {2025}, }
• (2025). HDR-NFlow: High dynamic range imaging with normalizing flow. NEUCOM, 647, 130512. (www)

  @article{ , abstract = {High Dynamic Range (HDR) imaging aims to generate a high-quality HDR image by fusing multi-exposure Low Dynamic Range (LDR) images. When input LDR images have large object motion and severe saturation, previous methods suffer from ghosting artifacts, which results in unpleasant HDR images and hinders real-world applications. To address this critical issue to reconstruct high-quality HDR images, we propose a novel HDR imaging framework based on the normalizing flow (called HDR-NFlow), which regards HDR imaging as a conditional generation task and consists of a conditional encoder and an invertible flow network. Specifically, the conditional encoder utilizes the proposed Composite Attention Merge Module (CAMM) to capture long-range context and fusion dependency of multi-exposed frames to align the large object motions and an Asymmetric Selective Kernel Detail (ASKD) module to capture texture information via locally stripy extraction. With the extracted features as reasonable conditions, the invertible flow network hallucinates realistic content for saturated regions and generates an HDR image by realizing the conversion of Gaussian distribution to HDR image distribution. We conduct extensive experiments on commonly used benchmark datasets to demonstrate that our method achieves state-of-the-art performance both quantitatively and qualitatively.}, archive = {J_NEUCOM}, author = {Shuaikang Shang and Xuejing Kang and Anlong Ming}, doi = {10.1016/j.neucom.2025.130512}, journal = {Neurocomputing}, month = {9}, pages = {130512}, shortjournal = {Neurocomputing}, title = {HDR-NFlow: High dynamic range imaging with normalizing flow}, volume = {647}, year = {2025}, }
• (2025). MorphoSeg: An uncertainty-aware deep learning method for biomedical segmentation of complex cellular morphologies. NEUCOM, 647, 130511. (www)

  @article{ , abstract = {Deep learning has revolutionized medical and biological imaging, particularly in segmentation tasks. However, segmenting biological cells remains challenging due to the high variability and complexity of cell shapes. Addressing this challenge requires high-quality datasets that accurately represent the diverse morphologies found in biological cells. Existing cell segmentation datasets are often limited by their focus on regular and uniform shapes. In this paper, we introduce a novel benchmark dataset of Ntera-2 (NT2) cells, a pluripotent carcinoma cell line, exhibiting diverse morphologies across multiple stages of differentiation, capturing the intricate and heterogeneous cellular structures that complicate segmentation tasks. To address these challenges, we propose an uncertainty-aware deep learning framework for complex cellular morphology segmentation (MorphoSeg) by incorporating sampling of virtual outliers from low-likelihood regions during training. Our comprehensive experimental evaluations against state-of-the-art baselines demonstrate that MorphoSeg significantly enhances segmentation accuracy, achieving up to a 7.74% increase in the Dice Similarity Coefficient (DSC) and a 28.36 reduction in the Hausdorff Distance. These findings highlight the effectiveness of our dataset and methodology in advancing cell segmentation capabilities, especially for complex and variable cell morphologies. The dataset and source code is publicly available at https://github.com/RanchoGoose/MorphoSeg .}, archive = {J_NEUCOM}, author = {Tianhao Zhang and Heather J. McCourty and Berardo M. Sanchez-Tafolla and Anton Nikolaev and Lyudmila S. Mihaylova}, doi = {10.1016/j.neucom.2025.130511}, journal = {Neurocomputing}, month = {9}, pages = {130511}, shortjournal = {Neurocomputing}, title = {MorphoSeg: An uncertainty-aware deep learning method for biomedical segmentation of complex cellular morphologies}, volume = {647}, year = {2025}, }
• (2025). A band grouping-based hybrid convolution for hyperspectral image super-resolution. NEUCOM, 647, 130510. (www)

  @article{ , abstract = {Hyperspectral images (HSIs) exhibit high spectral resolution but suffer from limited spatial resolution due to hardware limitations. While deep learning has been widely applied for single-image HSI super-resolution (HSI SR), challenges persist when auxiliary high-resolution images are unavailable. Existing convolutional neural networks(CNNs)-based methods encounter two primary issues: inputting all spectral bands simultaneously limits network design depth and includes redundant information from similar bands, hindering the full exploitation of HSI’s rich spatial–spectral details. Additionally, grouped reconstruction methods that prioritize spatial features using 2D CNNs often compromise spectral fidelity. To address these problems, we introduce BGSR, a novel approach for single HSI SR. This method strategically groups HSIs, enhances spectral diversity and reduces redundancy through group sampling, and introduces complementary information from distant symmetric groups to enhance details. The intra-group fusion module consists of parallel branches based on hybrid convolutions, utilizing a serial structure to better capture and utilize complementary intra-group spatial–spectral features. Moreover, a cross-group fusion module equipped with an attention-based sparse mask mechanism is introduced for comprehensive inter-group information integration and global optimization. Extensive evaluations on five public datasets demonstrate that BGSR outperforms existing methods, achieving state-of-the-art results. The code is available at https://github.com/LTTdouble/BGSR .}, archive = {J_NEUCOM}, author = {Tingting Liu and Tong Jiang and Chuncheng Zhang and Yuan Liu and Xiubao Sui and Qian Chen}, doi = {10.1016/j.neucom.2025.130510}, journal = {Neurocomputing}, month = {9}, pages = {130510}, shortjournal = {Neurocomputing}, title = {A band grouping-based hybrid convolution for hyperspectral image super-resolution}, volume = {647}, year = {2025}, }
• (2025). RankSync: Synergistic learning of embeddings for coarse and fine ranking in recommendation. NEUCOM, 647, 130506. (www)

  @article{ , abstract = {Coarse and fine ranking are two critical tasks in recommender system. The former focuses on efficiency, often employing low-dimensional embeddings to quickly filter candidates, while the latter emphasizes accuracy, utilizing high-dimensional embeddings for more granular sorting. Traditionally trained separately, these ranking models not only incur high training costs but also fail to fully exploit the synergy between tasks, limiting overall recommendation performance. Besides, many methods utilize pre-trained language model (PLM) to embed natural language text (such as reviews) for more accurate ranking. However, directly mapping these embeddings into lower-dimensional spaces, especially for coarse ranking, can weaken their representational capacity. So we propose a framework named RankSync, which aims to leverage PLMs to train coarse and fine ranking models simultaneously. RankSync jointly learns low-dimensional embeddings for coarse ranking and high-dimensional embeddings for fine ranking in a single process, outperforming independently trained fixed-dimensional embeddings. Specifically, RankSync learns embeddings at different granularities, then computes losses for each granularity separately. These losses are then combined with appropriate weights to optimize the whole model. Experiments show RankSync outperforms state-of-the-art baselines in both recommendation quality and training efficiency.}, archive = {J_NEUCOM}, author = {Yuqiao Liu and Nan Zheng and Song Zhang and Jiayue Li and Dongyuan Lu}, doi = {10.1016/j.neucom.2025.130506}, journal = {Neurocomputing}, month = {9}, pages = {130506}, shortjournal = {Neurocomputing}, title = {RankSync: Synergistic learning of embeddings for coarse and fine ranking in recommendation}, volume = {647}, year = {2025}, }
• (2025). Alignment-assisted frequency fusion network for RGB-infrared vehicle detection. NEUCOM, 647, 130505. (www)

  @article{ , abstract = {Fusion complementary information from RGB and infrared (IR) modalities enables CNN-based detectors to learn a more robust vehicle representation, thus enhancing detection performance. However, the inherent modality gap between RGB and IR hinders effective feature fusion and limits detection performance. In this work, we decouple the RGB-IR vehicle detection task into two sub-sequential tasks: mitigate the modality gap, adaptive fusion of multimodal features. Specifically, we introduce a novel Alignment-assisted Frequency Fusion vehicle detection Network (AFFNet), which comprises two innovative blocks: the Detail-aware Semantic Alignment Block (DSAB) and the Adaptive Frequency domain Fusion Block (AFFB). The DSAB alleviates the modality gap and improves the semantic consistency by exploring the feature correspondence through explicit bidirectional multimodal feature alignment. Furthermore, the AFFB decouples key vehicle features with 2D Discrete Fourier Transform (DCT) in the frequency domain and leverages the proposed fusion units to adaptively integrate fine-grained, frequency-domain vehicle features from different modalities, thereby enhancing the quality of the fusion representations. We conducted experiments on the VEDAI and DroneVehicle multimodal vehicle datasets, achieving improvements of 4% and 1.2% on mAP50 over the comparison methods, respectively. Both the quantitative and qualitative results show the effectiveness of our method.}, archive = {J_NEUCOM}, author = {Zhenshuai Chen and Wei Xiang and Zhiyuan Lin and Kaixuan Yang and Yunpeng Liu and Zelin Shi}, doi = {10.1016/j.neucom.2025.130505}, journal = {Neurocomputing}, month = {9}, pages = {130505}, shortjournal = {Neurocomputing}, title = {Alignment-assisted frequency fusion network for RGB-infrared vehicle detection}, volume = {647}, year = {2025}, }
• (2025). Borrowing treasures from neighbors: In-context learning for multimodal learning with missing modalities and data scarcity. NEUCOM, 647, 130502. (www)

  @article{ , abstract = {Multimodal machine learning with missing modalities is an increasingly relevant challenge arising in various applications such as healthcare. This paper extends the current research into missing modalities to the low-data regime, i.e., a downstream task has both missing modalities and limited sample size issues. This problem setting is particularly challenging and also practical as it is often expensive to get full-modality data and sufficient annotated training samples. We propose to use retrieval-augmented in-context learning to address these two crucial issues by unleashing the potential of a transformer’s in-context learning ability. Diverging from existing methods, which primarily belong to the parametric paradigm and often require sufficient training samples, our work exploits the value of the available full-modality data, offering a novel perspective on resolving the challenge. The proposed data-dependent framework exhibits a higher degree of sample efficiency and is empirically demonstrated to enhance the classification model’s performance on both full- and missing-modality data in the low-data regime across various multimodal learning tasks. When only 1% of the training data are available, our proposed ICL-CA method outperforms the best baseline by 5.9%, 5.9%, 5.3% and 10.8% on four datasets across various missing states. Notably, our method also reduces the performance gap between full-modality and missing-modality data compared with the baseline. Code is available 1 .}, archive = {J_NEUCOM}, author = {Zhuo Zhi and Ziquan Liu and Moe Elbadawi and Adam Daneshmend and Mine Orlu and Abdul Basit and Andreas Demosthenous and Miguel Rodrigues}, doi = {10.1016/j.neucom.2025.130502}, journal = {Neurocomputing}, month = {9}, pages = {130502}, shortjournal = {Neurocomputing}, title = {Borrowing treasures from neighbors: In-context learning for multimodal learning with missing modalities and data scarcity}, volume = {647}, year = {2025}, }
• (2025). Category-wise fine-tuning: Resisting incorrect pseudo-labels in multi-label image classification with partial labels. NEUCOM, 647, 130498. (www)

  @article{ , abstract = {Large-scale image datasets are often partially labeled, where only a few categories’ labels are known for each image. Assigning pseudo-labels to unknown labels to gain additional training signals has become prevalent for training deep classification models. However, some pseudo-labels are inevitably incorrect, leading to a notable decline in the model classification performance. In this paper, we propose a new method called Category-wise Fine-Tuning (CFT), aiming to reduce model inaccuracies caused by the wrong pseudo-labels. In particular, CFT employs known labels without pseudo-labels to fine-tune the logistic regressions of trained models individually to calibrate each category’s model predictions. Genetic Algorithm, seldom used for training deep models, is also utilized in CFT to maximize the classification performance directly. CFT is applied to well-trained models, unlike most existing methods that train models from scratch. Hence, CFT is general and compatible with models trained with different methods and schemes, as demonstrated through extensive experiments. CFT requires only a few seconds for each category for calibration with consumer-grade GPUs. We achieve state-of-the-art results on three benchmarking datasets, including the CheXpert chest X-ray competition dataset (ensemble mAUC 93.33%, single model 91.82%), partially labeled MS-COCO (average mAP 83.69%), and Open Image V3 (mAP 85.31%), outperforming the previous bests by 0.28%, 2.21%, 2.50%, and 0.91%, respectively. The single model on CheXpert has been officially evaluated by the competition server, endorsing the correctness of the result. The outstanding results and generalizability indicate that CFT could be substantial and prevalent for classification model development. Code is available at: https://github.com/maxium0526/category-wise-fine-tuning .}, archive = {J_NEUCOM}, author = {Chak Fong Chong and Xinyi Fang and Jielong Guo and Pedro Henriques Abreu and Yapeng Wang and Xu Yang and Wei Ke and Sio-Kei Im}, doi = {10.1016/j.neucom.2025.130498}, journal = {Neurocomputing}, month = {9}, pages = {130498}, shortjournal = {Neurocomputing}, title = {Category-wise fine-tuning: Resisting incorrect pseudo-labels in multi-label image classification with partial labels}, volume = {647}, year = {2025}, }
• (2025). Simplified transformer. NEUCOM, 647, 130497. (www)

  @article{ , abstract = {Based on two interesting mathematical observations that almost all square matrices are diagonalizable and invertible, this paper proposes a simplified transformer called SimplTrans, which eliminates parameters in the attention mechanism. It is known that the success of the Transformer can be attributed to its embedded attention mechanism, which plays a crucial role in the network. However, the introduction of a large number of parameters with this attention mechanism hinders the efficiency of learning and computation in the Transformer. The proposed SimplTrans leverages this parameter-free attention mechanism, leading to a significant reduction in the number of parameters. As a result, the proposed SimplTrans achieves improved learning and computation efficiency compared to the original Transformer. Furthermore, the complexity of SimplTrans is also reduced to that of the original Transformer, resulting in superior performance across various tasks, as demonstrated in our experiments. The associated code can be found at https://github.com/IntelligentInterdisciplinary/SimplTrans .}, archive = {J_NEUCOM}, author = {Lei Xu and Haiying Luo and Zhang Yi}, doi = {10.1016/j.neucom.2025.130497}, journal = {Neurocomputing}, month = {9}, pages = {130497}, shortjournal = {Neurocomputing}, title = {Simplified transformer}, volume = {647}, year = {2025}, }
• (2025). A novel zeroing neural network with activation function-enhanced convergence for efficient data matrix decomposition. NEUCOM, 647, 130495. (www)

  @article{ , abstract = {Given the critical role of zeroing neural networks (ZNN) in various fields, this paper introduces a novel activation function-based dynamic composite zeroing neural network matrix decomposition (AF-DCZNN-MD) model aimed at accelerating the solution of matrix decomposition problems. This method does not involve any matrix inversion operations. The error-tracking function defined within the model mitigates overfitting. By leveraging the monotonic increasing property of odd functions, the model expedites convergence and facilitates efficient physical implementation using low-complexity analog circuits. The finite-time stable convergence of the proposed model is theoretically supported by Lyapunov’s theory. During numerical experimentation, initial validation was conducted on simple matrices to evaluate the effects of different activation functions (AFs), parameter settings, and the absence of AFs on convergence speed. Additionally, the model’s robustness against noise interference was validated in noisy environments. Subsequently, a set of real matrices with dimensions 89351 × 150 was decomposed, yielding a reconstruction root mean square error (RMSE) of 0.1182 using the proposed model. This result demonstrated a 91.3158% reduction in RMSE compared to the initial state, thereby confirming the applicability of the proposed neural network model in handling matrices in real-world scenarios.}, archive = {J_NEUCOM}, author = {Yiguo Yang and Pin Wu and Duc Truong Pham and Shuai Li and Weibing Feng}, doi = {10.1016/j.neucom.2025.130495}, journal = {Neurocomputing}, month = {9}, pages = {130495}, shortjournal = {Neurocomputing}, title = {A novel zeroing neural network with activation function-enhanced convergence for efficient data matrix decomposition}, volume = {647}, year = {2025}, }
• (2025). SEG-doc: A simple yet efficient graph neural network framework for document key information extraction. NEUCOM, 647, 130493. (www)

  @article{ , abstract = {In recent years, Graph Neural Networks (GNNs) have gained significant attention as an effective approach in the field of document key information extraction. However, existing research introduces a significant amount of negative edges in their graph construction strategies, leading to a severe class-imbalance problem in relation extraction tasks. Additionally, popular GNNs usually regard all neighbor nodes as having the same weight during message passing, potentially resulting in the loss of critical information. To address these issues, we propose SEG-Doc : a simple yet efficient graph neural networks framework for document key information extraction. Firstly, the Text Histogram-based Graph Construction module (THGC) utilizes text histogram feature to determine the edge set during graph construction. Our strategy not only balances the positive and negative edge samples but also captures key–value relationships more effectively. Secondly, we propose the Layout Attention Mechanism (LAM), which computes layout attention between nodes and their neighbors to assign different weights during message passing. This mechanism assists GNNs in learning richer node representations and mitigating the impact of noise on network updates. We evaluated the effectiveness of SEG-Doc on three benchmark datasets: FUNSD, SROIE and RVL-CDIP. Comprehensive experiments demonstrate that the proposed SEG-Doc achieves state-of-the-art performance than competitive baselines on several public downstream benchmarks.}, archive = {J_NEUCOM}, author = {Yangyang Hui and Jiahao Liu and Qi Zhang and Tianyi Zhou and Yonghong Song}, doi = {10.1016/j.neucom.2025.130493}, journal = {Neurocomputing}, month = {9}, pages = {130493}, shortjournal = {Neurocomputing}, title = {SEG-doc: A simple yet efficient graph neural network framework for document key information extraction}, volume = {647}, year = {2025}, }
• (2025). A zeroing neural dynamics algorithm for discrete-time nonlinear optimization with linear equality constraint and perturbation inhibition. NEUCOM, 647, 130492. (www)

  @article{ , abstract = {As neural network technology gallops forward, more and more applied mathematical problems can be solved by such an intelligent algorithm. The traditional zeroing neural dynamics (ZND) can only solve problems effectively without perturbation. However, in the practical environment, the existence of perturbation may seriously affect the computational performance of intelligent algorithm. This paper investigates the discrete time-varying nonlinear optimization problem with linear equality constraints in the perturbation environment. With the help of integration technique, the model has the excellent performance of perturbation inhibition. First of all, the Lagrange multiplier method is used to construct a continuous-time ZND algorithm. Secondly, the widely-used discretization formulas (i.e., Euler-type discretization formula, Taylor-type discretization formula and five-instant-type discretization formula) are used to obtain different discrete-time ZND (DT-ZND) algorithms. Finally, in the experiments, compared with traditional DT-ZND algorithms, the proposed DT-ZND algorithms (i.e., Eu-DT-ZND algorithm, Ta-DT-ZND algorithm and FI-DT-ZND algorithm) have high accuracy and excellent performance.}, archive = {J_NEUCOM}, author = {Chao Jiang and Jiacheng Wang and Ruxin Zhao and Dimitrios K. Gerontitis and Yang Shi}, doi = {10.1016/j.neucom.2025.130492}, journal = {Neurocomputing}, month = {9}, pages = {130492}, shortjournal = {Neurocomputing}, title = {A zeroing neural dynamics algorithm for discrete-time nonlinear optimization with linear equality constraint and perturbation inhibition}, volume = {647}, year = {2025}, }
• (2025). MGTAB: A multi-relational graph-based twitter account detection benchmark. NEUCOM, 647, 130490. (www)

  @article{ , abstract = {The advancement of social media user stance detection and bot identification methodologies is fundamentally dependent on large-scale and high-quality benchmarks. Current benchmarks, however, frequently fall short in supporting graph-based account detection research due to limitations in annotation quality and incomplete user relationship data. To counter these issues, we present the Multi-Relational Graph-Based Twitter Account Detection Benchmark (MGTAB), the inaugural graph-based benchmark that simultaneously annotates both bot accounts and user stances. To our knowledge, MGTAB represents the most extensive original dataset in stance detection research, encompassing over 1.55 million users and 130 million tweets. This benchmark incorporates 10,199 expert-annotated users and seven distinct relationship types, thereby ensuring both annotation reliability and relational diversity. We have additionally identified and extracted the 20 most informative user property features alongside user tweet features to construct comprehensive user profiles. Our experiments revealed that graph-based approaches are generally more efficacious than feature-based approaches and perform better when incorporating multiple relations. By analyzing experiment results, we identify effective approaches for account detection and suggest potential future research directions in this field. Our benchmark and standardized evaluation procedures are freely available at: https://github.com/GraphDetec/MGTAB .}, archive = {J_NEUCOM}, author = {Shuhao Shi and Kai Qiao and Zihao Liu and Jie Yang and Chen Chen and Jian Chen and Bin Yan}, doi = {10.1016/j.neucom.2025.130490}, journal = {Neurocomputing}, month = {9}, pages = {130490}, shortjournal = {Neurocomputing}, title = {MGTAB: A multi-relational graph-based twitter account detection benchmark}, volume = {647}, year = {2025}, }
• (2025). R2-diff: Denoising by diffusion as a refinement of retrieved motion for image-based motion prediction. NEUCOM, 647, 130489. (www)

  @article{ , abstract = {Diffusion models are promising in robot motion prediction to learn diverse motions from demonstrations. However, not only the diversity but also the spatial fidelity is essential in achieving robot manipulation tasks. Conventional diffusion models often fail to achieve sufficient fidelity. We identify that this limitation arises from the reliance on sampled initial noises, which are typically too far from any correct motion that achieves the task, making it difficult to accurately denoise them. To overcome this limitation, we propose a novel approach that integrates image-based motion retrieval with diffusion-based motion refinement. Our image-based motion retrieval gives a near-correct motion, and our diffusion model refines it instead of denoising a sampled random noise. However, we found that a naive integration does not improve the fidelity. To achieve effective integration, we propose two essential techniques: (1) automatically tuning the noise scale based on a nearest neighbor motion retrieval during training to enable accurate refinement from retrieved motions, and (2) bridging the domain gap in inference by adding small noise to the retrieved clean motions, aligning them with the learned noisy motion distribution. Experiments on various robot manipulation tasks demonstrate that our method outperforms state-of-the-art methods by over 6.3%. Importantly, we also demonstrate that diffusion models trained with typical noise scales cannot improve the performance when retrieved motions are used as initial motions, highlighting the necessity of our two novel techniques.}, archive = {J_NEUCOM}, author = {Takeru Oba and Norimichi Ukita}, doi = {10.1016/j.neucom.2025.130489}, journal = {Neurocomputing}, month = {9}, pages = {130489}, shortjournal = {Neurocomputing}, title = {R2-diff: Denoising by diffusion as a refinement of retrieved motion for image-based motion prediction}, volume = {647}, year = {2025}, }
• (2025). A positive-unlabeled learning approach for industrial anomaly detection based on self-adaptive training. NEUCOM, 647, 130488. (www)

  @article{ , abstract = {The availability of large amounts of data offers potential for condition monitoring and anomaly detection in complex industrial systems. However, a major obstacle arises from the prevalent issue of insufficient and sparse labeling, particularly of anomalous data. This scarcity poses a significant challenge in efficiently utilizing a large yet weakly-labeled dataset. Our work addresses this challenge by framing the problem as a highly unlabeled, imbalanced, and positive-unlabeled(PU) learning scenario, which integrates massive unlabeled data with rare anomalous data into a single training set. We introduce a self-adaptive training based approach, termed SatPU, designed for training anomaly detection models in this context. Our approach incorporates a novel pseudo labeling scheme coupled with a new dynamic re-weighting technique to facilitate self-adaptive training. These augmentations address issues related to confirmation bias and the risk of converging to trivial outcomes in the PU learning setup, thereby enhancing classification performance. Comparative experiments conducted on two industrial datasets against state-of-the-art PU learning methods validate the effectiveness of the proposed approach. The results demonstrate that SatPU exhibits superior performance when trained on industrial data that violate the select completely at random (SCAR) assumption and contain a high proportion of unlabeled samples. Furthermore, an ablation study is conducted to substantiates the necessities of the introduced augmentations for self-adaptive training.}, archive = {J_NEUCOM}, author = {Keyu Song and Chao Liu and Dongxiang Jiang}, doi = {10.1016/j.neucom.2025.130488}, journal = {Neurocomputing}, month = {9}, pages = {130488}, shortjournal = {Neurocomputing}, title = {A positive-unlabeled learning approach for industrial anomaly detection based on self-adaptive training}, volume = {647}, year = {2025}, }
• (2025). Interpretable deep model pruning. NEUCOM, 647, 130485. (www)

  @article{ , abstract = {We propose a deep learning model pruning method based on the mutual information and the total correlation computation. The solid information theory foundation suggests us the transparent and interpretable pruning strategy for the much-wanted goal of trustworthy AI. We focus on the pruning of the representation layer(s) in an unsupervised manner by measuring the feature redundancy between neurons in the layer(s). The redundancy computation relies on the mutual information then total correlation estimation to measure how two sets of neurons own overlapped information for the pruning to suggest a model with less complexity. Overall, we horizontally focus on the narrowest, assumed to be the most representable layer(s) for the pruning and in the layer(s), find the neuron(s) with the highest redundancy to prune vertically, if having the network models go from the inputs to the outputs horizontally. In short, we propose an interpretable pruning method and one can understand how and why some neurons are pruned from the network. A further issue to address is to have a proper estimation on the mutual information and total correlation. Some theoretical and empirical results confirm the value of the proposed method and the evaluation shows that the proposed pruning method can indeed perform superior or competitively to other state-of-the-art methods even when some of the pruning methods are considered supervised ones. We have tested the proposed method on various network structures such as fully-connected and convolutional networks given real-world datasets for the evaluation. The code for this work is available via https://github.com/NTUST-ML-lab/Interpretable-Deep-Model-Pruning .}, archive = {J_NEUCOM}, author = {Hanjuan Huang and Hsing-Kuo Pao}, doi = {10.1016/j.neucom.2025.130485}, journal = {Neurocomputing}, month = {9}, pages = {130485}, shortjournal = {Neurocomputing}, title = {Interpretable deep model pruning}, volume = {647}, year = {2025}, }
• (2025). A variational framework based on synergy between illumination and reflectance for retinex decomposition. NEUCOM, 647, 130480. (www)

  @article{ , abstract = {The primary assumption of Retinex theory is that an image can be decomposed into its illumination and reflectance components. Traditional variational methods typically estimate the illumination and reflectance through a L2 fidelity term combined with different regularizations. However, this may ignore the collaboration mechanism between the illumination and reflectance, leading to inaccurate estimations. To address this challenge, this paper proposes a novel variational framework for Retinex, which takes full advantages of the synergistic collaboration mechanism between the illumination and reflectance. Then we use the alternating minimization algorithm to solve the proposed variational model. To demonstrate the generality of our framework, the proposed Retinex framework are applied in different applications including low-light image enhancement and tone mapping of high dynamic range (HDR) images. In each application, we provide an approach to construct the corresponding Retinex model by setting the synergistic collaboration indicator respectively. Extensive experimental results show that the proposed method can maintain visual naturalness and achieve high objective measurements on different representative metrics compared to the mainstream methods.}, archive = {J_NEUCOM}, author = {Xiaoting Zhang and Wei Wang}, doi = {10.1016/j.neucom.2025.130480}, journal = {Neurocomputing}, month = {9}, pages = {130480}, shortjournal = {Neurocomputing}, title = {A variational framework based on synergy between illumination and reflectance for retinex decomposition}, volume = {647}, year = {2025}, }
• (2025). New results on quasi-stability/synchronization of quaternion-valued fuzzy memristive neural networks. NEUCOM, 647, 130472. (www)

  @article{ , abstract = {This paper investigates the quasi-stability/synchronization problem for quaternion-valued fuzzy memristive system subject to fuzzy terms. Via a new exponentially quasi-stability definition, one can estimate the convergence rate more effectively, besides, the system can adjust the converge rate by modifying the controller. Then, with aim to utilize a new lemma generalized from Halanay inequality, some improved results are presented to achieve the quasi-stability/synchronization conclusions, which not only avoid considering the “magnitude” of quaternion, but also gives a more flexible criteria. Finally, example is given to substantiate the feasibility of the developed scheme.}, archive = {J_NEUCOM}, author = {Ruoxia Li and Chenyu Sun and Zhengwen Tu and Jinde Cao}, doi = {10.1016/j.neucom.2025.130472}, journal = {Neurocomputing}, month = {9}, pages = {130472}, shortjournal = {Neurocomputing}, title = {New results on quasi-stability/synchronization of quaternion-valued fuzzy memristive neural networks}, volume = {647}, year = {2025}, }
• (2025). SAGE: Self-evolving agents with reflective and memory-augmented abilities. NEUCOM, 647, 130470. (www)

  @article{ , abstract = {Large language models (LLMs) have made significant advances in natural language processing but face persistent challenges in dynamic environments, including continuous decision-making, long-term memory retention, and constrained context windows. To address these limitations, we propose SAGE (Self-evolving Agents with Reflective and Memory-augmented Abilities), a novel framework comprising three collaborative agents (User, Assistant, and Checker) that integrate iterative feedback, reflective reasoning, and memory optimization based on the Ebbinghaus forgetting curve. SAGE enhances agents’ multi-tasking and long-context capabilities through adaptive strategy adjustment, optimized information storage, and reduced cognitive load. Evaluations across AgentBench and long-text tasks demonstrate transformative improvements: closed-source models like GPT-4 achieve 2.26 × performance gains in database operations, while open-source models exhibit 5.0–48.0 absolute percentage point improvements, with smaller models (e.g., Qwen-1.8B) closing the gap to GPT-3.5-level performance. Key innovations include a memory syntax mechanism reducing overhead by 47% and a reflection module resolving 73.6% of ambiguous references in dialog tasks. These advances position SAGE as a scalable solution for real-world applications such as customer support (45.3% WebShop task improvement) and healthcare diagnostics (22.06 F1 on HotpotQA). Our results underscore SAGE’s ability to bridge theoretical LLM capabilities with practical deployment needs, enabling robust, self-evolving AI systems.}, archive = {J_NEUCOM}, author = {Xuechen Liang and Meiling Tao and Yinghui Xia and Jianhui Wang and Kun Li and Yijin Wang and Yangfan He and Jingsong Yang and Tianyu Shi and Yuantao Wang and Miao Zhang and Xueqian Wang}, doi = {10.1016/j.neucom.2025.130470}, journal = {Neurocomputing}, month = {9}, pages = {130470}, shortjournal = {Neurocomputing}, title = {SAGE: Self-evolving agents with reflective and memory-augmented abilities}, volume = {647}, year = {2025}, }
• (2025). Multi-scale hybrid mamba-LSTM experts for long-short term human trajectory prediction under CVAE framework. NEUCOM, 647, 130464. (www)

  @article{ , abstract = {Pedestrian trajectory prediction plays an important role in the design of autonomous driving systems and video surveillance. Despite substantial advancements in this domain, the accuracy of predictive models for long-term trajectories has experienced a notable decline. Despite significant progress in pedestrian trajectory prediction, the accuracy of prediction models significantly decreases when predicting long-term trajectories. Our research finds that this challenge arises from the local irregularities in pedestrian trajectories, which increase the randomness and complexity of the trajectories. Existing literature on social modeling based on the distance between pedestrians tends to focus more on the local interactions between pedestrians, neglecting the modeling of global trajectory features. In this paper, we point out that pedestrian trajectories can be effectively decomposed into global and local changes, which can be captured from the trajectory data in both global and local views. We introduce the State Space Attention LSTM (SSAL). SSAL leverages GAT-Mamba as an expert to dynamically extract the global changes of long-range pedestrian trajectories at a coarse-grained level. The Attention-LSTM, as another expert, extracts the local changes of short-range pedestrian trajectories at a fine-grained level. The Mamba component, based on a State Space Model (SSM), employs a selectable state mechanism with an input-dependent mechanism that selectively retains long-term patterns while filtering out fluctuations through global spatio-temporal attention. In contrast, the Attention-LSTM component enhances its perception of local changes through attention mechanisms, effectively capturing the historical trajectories and social relationships of pedestrians. The Global–Local Gate facilitates the adaptive integration of global and local patterns. To address the uncertainty in pedestrian trajectory prediction and generate multi-modal trajectories, we employ a deep generative model known as the conditional variational autoencoder (CVAE). Comprehensive experiments demonstrate that SSAL not only achieves state-of-the-art results in long-term predictions and excels in short-term predictions. Notably, on the ETH-UCY benchmark dataset, SSAL achieves improvement margins of 35% and 42% in terms of Average Displacement Error (ADE) and Final Displacement Error (FDE), respectively.}, archive = {J_NEUCOM}, author = {Hong Zhang and YuCheng Shi and Lei Shi and Chengming Liu and Zhenghan Gao and Yamei Xu and Liyong Chen}, doi = {10.1016/j.neucom.2025.130464}, journal = {Neurocomputing}, month = {9}, pages = {130464}, shortjournal = {Neurocomputing}, title = {Multi-scale hybrid mamba-LSTM experts for long-short term human trajectory prediction under CVAE framework}, volume = {647}, year = {2025}, }
• (2025). An efficient scale-aware model based on the improved RT-DETR for pomegranate growth stage detection. NEUCOM, 647, 130462. (www)

  @article{ , abstract = {The phenotypic characteristics of pomegranates at different growth stages significantly influence yield, making real-time detection essential for optimizing production. Additionally, complex environmental conditions complicate this process. Current research on pomegranate growth stage detection is scarce, with many studies ignoring environmental interference. To address these gaps, this study proposes a hybrid and efficient scale-aware pomegranate detection transformer model (HESP-DETR) based on Real-Time DEtection TRansformer (RT-DETR). Firstly, reparameterized cheap-operation layer aggregation network (RCLAN) was proposed to aggregate multi-layer semantic information, thereby reducing redundant feature processing and optimizing inference efficiency. Secondly, diverse branch block fusion module (DBFusion) was proposed to merge local and global semantics through multiple branch paths, preserving information in complex scenes while alleviating the computational burden. Finally, multi-scale multi-head self-attention (MMSA) was improved and introduced into the transformer encoder to enhance multi-scale feature processing and strengthen the interaction between local and global features, thereby boosting the model’s global perception ability. Experimental results show that HESP-DETR achieves a mean Average Precision (mAP 50 ) of 93.4%, outperforming RT-DETR by increasing mAP 50 by 1% and reducing parameter count, Giga Floating-point Operations Per Second (GFLOPs), weight size and inference time by 37.5%, 36.5%, 37% and 34% respectively. Robustness experiments further confirm that HESP-DETR demonstrates superior anti-interference capability compared to other mainstream models. Overall, HESP-DETR achieves a favorable balance between accuracy, model size, and speed, delivering a positive impact on the practical detection tasks of pomegranate growth stages.}, archive = {J_NEUCOM}, author = {Song Wang and Shaocong Dong and Shiyu Chen and Mingyu Liu}, doi = {10.1016/j.neucom.2025.130462}, journal = {Neurocomputing}, month = {9}, pages = {130462}, shortjournal = {Neurocomputing}, title = {An efficient scale-aware model based on the improved RT-DETR for pomegranate growth stage detection}, volume = {647}, year = {2025}, }
• (2025). Event-triggered h∞ control for highly dissipative spatiotemporal process via ADP: Application to industrial sintering process. NEUCOM, 647, 130461. (www)

  @article{ , abstract = {Highly dissipative spatiotemporal processes (HDSP) widely exist in industrial preparation, and the precise control usually plays a critical role in guaranteeing preparation quality. Since external disturbance exists in general, an event-triggered H ∞ control method for spatiotemporal processes that are formulated by parabolic partial differential equation (PPDE) systems is studied. First, the problem presentation for the event-triggered H ∞ control is provided by incorporating the reduce-then-design thought and the time-triggered mechanism. A novel event-triggered condition is proposed from the structure of the Hamilton–Jacobi-Isaacs (HJI) equation, and is proved to be effective in guaranteeing the PPDE system stability and the boundedness of interexecution times. Then, adaptive dynamic programming (ADP) is adopted via neural network approximate technology, where the optimal value function of HJI equation is approached by a critic network, the event-triggered control is realized by an actor network governed by the designed condition, and the disturbance strategy is also contained. A verification study is conducted within an industrial sintering process to illustrate the effectiveness and convergence of the PPDE system event-triggered H ∞ control method.}, archive = {J_NEUCOM}, author = {Binyan Li and Jiayuan Gao and Ning Chen and Biao Luo and Shaosheng Fan and Jiayao Chen and Weihua Gui}, doi = {10.1016/j.neucom.2025.130461}, journal = {Neurocomputing}, month = {9}, pages = {130461}, shortjournal = {Neurocomputing}, title = {Event-triggered h∞ control for highly dissipative spatiotemporal process via ADP: Application to industrial sintering process}, volume = {647}, year = {2025}, }
• (2025). Neural networks-based terminal sliding mode fault tolerant control to quadruped robots with actuator fault. NEUCOM, 647, 130457. (www)

  @article{ , abstract = {A composite fault-tolerant control framework is proposed for quadruped robot systems characterized by actuator faults and uncertain dynamic parameters. The framework integrates neural network approximation with adaptive nonsingular fast sliding mode control. Specifically, it includes: (1) employs the radial basis function neural networks to approximate the system’s nonlinear dynamics online, effectively compensating for parameter uncertainties; (2) designs an adaptive anti-fault estimator that dynamically estimates the upper bound of faults, thereby enhancing the ability to mitigate actuator faults; and (3) introduces a nonsingular fast terminal sliding mode structure, enabling finite-time convergence of tracking errors, significantly improving response speed. Simulation results demonstrate that this algorithm effectively enhances the control accuracy of the quadruped robot systems in the presence of actuator faults.}, archive = {J_NEUCOM}, author = {Pengxin Yang and Shuang Zhang and Xinbo Yu and Nan Feng and Yi Xing and Wei He}, doi = {10.1016/j.neucom.2025.130457}, journal = {Neurocomputing}, month = {9}, pages = {130457}, shortjournal = {Neurocomputing}, title = {Neural networks-based terminal sliding mode fault tolerant control to quadruped robots with actuator fault}, volume = {647}, year = {2025}, }
• (2025). Frequency-embedded watermarking: Defending vision models from extraction attacks. NEUCOM, 647, 130456. (www)

  @article{ , abstract = {The success of deep learning models in visual tasks has also raised concerns over protecting their intellectual property (IP). To this end, model watermarking techniques have been proposed to verify ownership. However, many existing watermarking methods lack robustness against model extraction attacks, where adversaries query the original model and use the responses to create surrogate models, bypassing ownership verification techniques. To tackle this problem, we propose a novel Frequency-Embedded Watermarking (FEW) for protecting vision models. Specifically, FEW combines high-level semantic information and low-level visual features of two source classes from the training set in frequency domain, generating the watermark samples that naturally align with the input distribution of the model. Since the model processes watermark samples and normal inputs in the same way, a surrogate model trained through extraction attacks to replicate the main functionality inherently retains the ability to recognize watermark inputs. Additionally, owing to its preservation of the original image’s spatial structure, FEW is well-suited to both classification and dense prediction tasks. We systematically conduct comprehensive experiments on the FEW using six widely used datasets under both optimal model extraction conditions for the attacker and more realistic extraction scenarios. The results demonstrate that our watermarks are effectively preserved in the surrogate models and remain robust against various watermark removal and detection techniques.}, archive = {J_NEUCOM}, author = {Haohua Zhou and Xinyu Liu and Zhidong Shen}, doi = {10.1016/j.neucom.2025.130456}, journal = {Neurocomputing}, month = {9}, pages = {130456}, shortjournal = {Neurocomputing}, title = {Frequency-embedded watermarking: Defending vision models from extraction attacks}, volume = {647}, year = {2025}, }
• (2025). Defect image generation through feature disentanglement using StyleGAN2-ADA. NEUCOM, 647, 130455. (www)

  @article{ , abstract = {The limited availability of real defect samples and the high cost associated with manual defect generation present significant challenges in automated defect detection systems. In this paper, we propose a novel few-shot defect generation approach based on StyleGAN2-ADA, designed to produce high-quality defect samples for improving defect and anomaly detection performance. The proposed methodology comprises two stages: first, two independent few-shot generation models learn the distributions of defect-rich and defect-free datasets, respectively. Next, an indirect decoupling strategy is employed to incorporate prior knowledge of defect distributions, while still facilitating the unconditional generation of defect samples. To further enhance the generation process, we introduce a defect mask attention mechanism that specifically targets defect regions and the objects of interest, effectively preventing the generation of defect-free images and reducing background degradation. This method not only enhances the precision of defect generation but also allows for the flexible and realistic simulation of random defect variations. Extensive experiments on the MVTec AD and VisA dataset demonstrate the superiority of our approach, outperforming existing GAN-based defect generation methods in few-shot scenarios. Specifically, the proposed method excels across 35 different defects and 80 evaluation metrics, with 55 metrics surpassing all previous methods and 28 ranking second. Our approach achieves an average improvement of 23% on the KID metric.}, archive = {J_NEUCOM}, author = {Zhiquan He and Kangxing Wu and Yang Wen}, doi = {10.1016/j.neucom.2025.130455}, journal = {Neurocomputing}, month = {9}, pages = {130455}, shortjournal = {Neurocomputing}, title = {Defect image generation through feature disentanglement using StyleGAN2-ADA}, volume = {647}, year = {2025}, }
• (2025). Prompt-based hybrid supervised contrastive learning for emotion recognition in conversation. NEUCOM, 647, 130453. (www)

  @article{ , abstract = {Emotion Recognition in Conversation has garnered significant attention due to technological advances such as social networks, virtual assistants, and human–computer interaction. Existing ERC models often struggle with the nuances of context semantics and are notably challenged by the issue of unbalanced category distribution within datasets. In this paper, a Prompt-based Hybrid Supervised Contrastive Learning model is designed to enhance the correlation between semantics and emotions, effectively addressing imbalanced sample distribution scenarios. Specifically, we explore a prompt-based contextual emotion structure that leverages a transformer encoder to generate superior dialogue-level emotion embedding from each utterance. Subsequently, we design a hybrid supervised contrastive learning loss that integrates normalized supervised contrastive learning (NSCL) loss and cross-entropy (CE) loss. The NSCL loss facilitates the learning of nuanced emotional features, while the CE loss efficiently optimizes the classifiers. Additionally, an adaptive curriculum strategy is implemented to dynamically adjust the training difficulty based on the model’s learning progress. The proposed method is compared with 15 state-of-the-art ERC models on the MELD and IEMOCAP datasets. The experimental results demonstrate that our model outperforms other competitors.}, archive = {J_NEUCOM}, author = {Chuangxin Cai and Kao Zhang and Zhihua Hu and Xianxuan Lin and Zhigeng Pan}, doi = {10.1016/j.neucom.2025.130453}, journal = {Neurocomputing}, month = {9}, pages = {130453}, shortjournal = {Neurocomputing}, title = {Prompt-based hybrid supervised contrastive learning for emotion recognition in conversation}, volume = {647}, year = {2025}, }
• (2025). Time series is not enough: Financial transformer reinforcement learning for portfolio management. NEUCOM, 647, 130451. (www)

  @article{ , abstract = {Existing portfolio management research is primarily focused on analyzing the relationships between price fluctuations of individual assets to optimize trading strategies. Most reinforcement learning research have several limitations in extracting inter-asset correlations, such as the inability to effectively simulate long-term market dynamics, interference in gradient propagation due to misaligned optimization objectives across sequentially connected modules, and difficulties in capturing complex spatiotemporal dependencies. These limitations lead to an overemphasis on local information during the feature extraction process, ultimately hindering the framework’s ability to achieve global optimization and exacerbating the risk of overfitting. To address these limitations, we propose an innovative deep reinforcement learning framework, termed Financial Transformer Reinforcement Learning (FTRL). We evaluate the efficacy of FTRL against three learning variants, fourteen traditional strategies, and nine reinforcement learning methods across four Dow Jones and four NASDAQ datasets. Our results show that FTRL consistently outperforms all strategies in return, Sharpe ratio, and Sortino ratio, with at least 3.9%, 1.5%, and 5.3% improvements on Dow Jones datasets, respectively. On the NASDAQ datasets, FTRL achieves at least 40% higher average cumulative returns. We further validate the generalization of FTRL on the Greek stock market, where it delivers at least a 3.7% return advantage over other strategies, highlighting its applicability in emerging markets.}, archive = {J_NEUCOM}, author = {Xiaotian Ren and Ruoyu Sun and Zhengyong Jiang and Angelos Stefanidis and Hongbin Liu and Jionglong Su}, doi = {10.1016/j.neucom.2025.130451}, journal = {Neurocomputing}, month = {9}, pages = {130451}, shortjournal = {Neurocomputing}, title = {Time series is not enough: Financial transformer reinforcement learning for portfolio management}, volume = {647}, year = {2025}, }
• (2025). Adaptive synthetic oversampling with poincaré distance and two-stage weight assignment. NEUCOM, 647, 130436. (www)

  @article{ , abstract = {In machine learning, the class imbalance problem(CIP) presents a persistent challenge, negatively impacting the training efficacy and the model’s generalization capabilities. Oversampling has emerged as a widely adopted approach for addressing CIP. However, most existing oversampling methods exhibit limited performance, arising from the within-class distribution imbalance in imbalanced datasets. This paper proposes a Poincaré distance and Two-stage weight assignment-based adaptive synthetic oversampling (PTASO), a novel method designed to address the issue of within-class distribution imbalance and improve the distribution of synthetic samples. Firstly, PTASO utilizes Poincaré distance to extract the distance and density information of datasets; Secondly, to enhance the clarity of the decision boundary, overlapping majority samples are removed by leveraging distance information. Thirdly, minority samples are clustered into boundary and safe regions using distance information. Subsequently, each region is assigned a global weight based on its distance information, while local weights are assigned to individual samples in inverse proportion to their density information. Finally, PTASO establishes an adaptive generation strategy to ensure synthetic samples conform to their respective regions. Specifically, Boundary-enhanced Affine Shadowsampling (BEAS) is developed for boundary region; while the Overlapping-avoided Gamma-distribution-based Synthetic oversampling (OGS) is designed for safe region. Experiments conducted on 35 UCI imbalanced datasets validate the superiority of PTASO in enhancing classification performance.}, archive = {J_NEUCOM}, author = {Chenxiao Zhou and Lianying Zou}, doi = {10.1016/j.neucom.2025.130436}, journal = {Neurocomputing}, month = {9}, pages = {130436}, shortjournal = {Neurocomputing}, title = {Adaptive synthetic oversampling with poincaré distance and two-stage weight assignment}, volume = {647}, year = {2025}, }
• (2025). All on board: Efficient reinforcement learning with milestone aggregation in asynchronous distributed training for RTS games. NEUCOM, 647, 130430. (www)

  @article{ , abstract = {Reinforcement Learning has been used with great success in real-time strategy (RTS) games. Training a successful game agent typically involves multiple steps: the agent plays a game match, collects sample data to update the model, and an upgraded version of the agent interacts with the environment. This iterative training process is usually accompanied by data hungry and is very time-consuming, which may lead to the availability of high-performance devices being a necessity for relevant research and the inability to utilize a large number of low-profile devices to assist in training. Furthermore, asynchronous training of different clients may result in asynchronous learning progress, which can impair the performance of the server model. In response to these challenge, we propose a milestone aggregation based asynchronous federated reinforcement learning method (MA-AFRL). The method is based on the idea of curriculum-learning and unites multiple clients for joint training to alleviate the data scarcity. Meanwhile, in this paper, we propose a milestone model ensemble algorithm (M-Avg) for the multi-model ensemble challenge under curriculum learning, which is more robust to the ensemble of multiple local models and can adapt to the asynchronous training challenge under distributed systems without sharing local sample data. In our experiments, we evaluate our MA-AFRL method by using the different non-cheating levels under StarCraft II (SC2) as a representative of RTS games and compare it with various benchmark strategies. The results show that the MA-AFRL is able to achieve a standard win rate against the built-in AI with the help of multiple clients utilizing fewer iterations, while avoiding centralized uploading of sample data. Furthermore, MA-AFRL also has better performance in dealing with multi-client asynchronous training environments. The proposed method offers a promising direction for training high-quality models for RTS games in a distributed environment while simultaneously enhancing training efficiency.}, archive = {J_NEUCOM}, author = {Dayu Zhang and Weidong Bao and Ji Wang and Xiongtao Zhang and Jingxuan Zhou and Yaohong Zhang}, doi = {10.1016/j.neucom.2025.130430}, journal = {Neurocomputing}, month = {9}, pages = {130430}, shortjournal = {Neurocomputing}, title = {All on board: Efficient reinforcement learning with milestone aggregation in asynchronous distributed training for RTS games}, volume = {647}, year = {2025}, }
• (2025). Advancing ensemble learning against unlearnable data. NEUCOM, 647, 130422. (www)

  @article{ , abstract = {The unauthorized acquisition of image data for training deep neural network (DNN) models raises significant privacy concerns. A potential solution to address this concern is to render data unlearnable, disrupting the training process of DNN models. Given the unpredictable nature of the machine learning techniques that unauthorized users may employ, unlearnable data should possess the capability to guard against various advanced machine learning approaches, including ensemble learning. In this research, we propose frameworks to advance conventional ensemble learning methods, including stacking, boosting, and bagging, for defeating twelve existing data protection approaches by making their generated unlearnable data into learnable ones. Our extensive experiments reveal that the model ensembles resulted from the proposed frameworks attain an average test accuracy surpassing 89% on unlearnable CIFAR-10 datasets generated by the twelve existing approaches. These resulting findings suggest that all twelve existing approaches fail to the avoidance of data unauthorized use, and there is an urgent need to develop a new data protection approach for preventing data from such misuse.}, archive = {J_NEUCOM}, author = {Thushari Hapuarachchi and Kaiqi Xiong}, doi = {10.1016/j.neucom.2025.130422}, journal = {Neurocomputing}, month = {9}, pages = {130422}, shortjournal = {Neurocomputing}, title = {Advancing ensemble learning against unlearnable data}, volume = {647}, year = {2025}, }
• (2025). Neural network-based fully actuated trajectory tracking control for space manipulator with system uncertainties. NEUCOM, 647, 130399. (www)

  @article{ , abstract = {Space manipulator is regarded as one of the technologies with the greatest potential in various on-orbit service missions. By leveraging the fully actuated system approach (FASA) and neural network (NN), this paper explores the trajectory tracking control problem of space manipulator with the dynamical model error and external disturbance. The designed controller consists of three parts: the FASA-based part and two radial basis function NN (RBFNN) auxiliary inputs, the FASA-based part provides the desired eigenstructure and the global asymptotic stability for the nominal error system. It solves the problem that the trajectory tracking effect of space manipulator is poor under the working conditions of model error and external disturbance. This paper adopts a combination of offline training and online learning to comprehensively deal with uncertainties. By leveraging particle swarm optimization (PSO) method, we utilize abundant offline data to train the first RBFNN so as to approximate the inherent model error, and adopt an online adaptive learning strategy to update the weight of the second RBFNN. Based on Lyapunov’s direct method, stability analysis validates that the tracking error and the estimation error of RBFNN weight would finally converge to a small neighborhoods about the origin. Finally, simulation results with a planar free-flying space manipulator show that the proposed controller demonstrates more rapid convergence and more accurate tracking precision in contrast to the classical adaptive RBFNN controller.}, archive = {J_NEUCOM}, author = {Jing Xu and Kai Zhang and Zhaoke Ning and Zhihua Chen}, doi = {10.1016/j.neucom.2025.130399}, journal = {Neurocomputing}, month = {9}, pages = {130399}, shortjournal = {Neurocomputing}, title = {Neural network-based fully actuated trajectory tracking control for space manipulator with system uncertainties}, volume = {647}, year = {2025}, }
• (2025). Hessian-based parameter perceptual fingerprint characterization for hyperspectral image classification models. NEUCOM, 647, 130359. (www)

  @article{ , abstract = {Deep neural networks (DNNs) are highly susceptible to integrity attacks, due to their complexity and nonlinearity. Existing DNN integrity verification methods mainly focus on RGB image classification models. Addressing the problem of pixel-level classification for hyperspectral images with low spatial resolution and high spectral resolution, hyperspectral image classification models often employ 3-D convolutions and possess complex network structures. In this paper, we propose a novel Hessian-based parameter perceptual fingerprint characterization method for hyperspectral image classification model integrity verification. In our method, a sample augmentation strategy is proposed to increase the parameter activation amount of individual sample. Then, we propose using the Hessian matrix to enhance the samples’ ability of capturing the changes of model parameters. Finally, a parameter-aware sample selection algorithm is designed to enhance the efficiency of verification. Experimental results demonstrate that the proposed method is not only effective but also efficient, with verification efficiency improved by at least 10 times compared to only using original images, even over 200 times for some attacks. The proposed method can detect integrity attacks such as backdoor attacks, model pruning, and model quantization using fewer than 3 samples.}, archive = {J_NEUCOM}, author = {Shiqiang Dong and Song Xiao and Yaqi Liu and Jiahui Zhao and Wenqian Dong and Limeng Lu and Chenglei Luo}, doi = {10.1016/j.neucom.2025.130359}, journal = {Neurocomputing}, month = {9}, pages = {130359}, shortjournal = {Neurocomputing}, title = {Hessian-based parameter perceptual fingerprint characterization for hyperspectral image classification models}, volume = {647}, year = {2025}, }
• (2025). A hyperparameter-free scene-motion aware cross attention network for video anomaly detection in surveillance. NEUCOM, 647, 130348. (www)

  @article{ , abstract = {Video Anomaly Detection is a critical component of intelligent surveillance systems, focusing on identifying unusual human behaviors or abnormal objects in video sequences. However, most existing approaches, which typically rely on reconstruction or prediction techniques, face two major issues: (1) extensive hyperparameter tuning is required across different datasets and sample types, limiting their generalization; (2) they struggle to effectively capture spatiotemporal relationships, especially in complex, multi-object scenes. To address these, we propose the Scene-Motion Aware Cross Attention Network (SMACAN), a hyperparameter-free model for video anomaly detection, based on the characteristics of video surveillance. Our model incorporates three key modules: the Scene-Object Refinement Module, which adaptively adjusts motion information based on the varying scales of scenes and objects; the Cross-Flow Attention Module, which integrates motion and appearance features to better distinguish normal from anomalous behaviors; and the Motion Feature Distribution Module, which employs a parameter-free distance-based measurement for efficient anomaly detection. Extensive experiments on several public datasets demonstrate that SMACAN achieves state-of-the-art performance, significantly improving accuracy in video anomaly detection tasks.}, archive = {J_NEUCOM}, author = {Qiyue Sun and Yang Yang and Yunxia Liu and Hongjun Wang}, doi = {10.1016/j.neucom.2025.130348}, journal = {Neurocomputing}, month = {9}, pages = {130348}, shortjournal = {Neurocomputing}, title = {A hyperparameter-free scene-motion aware cross attention network for video anomaly detection in surveillance}, volume = {647}, year = {2025}, }
• (2025). Adaptive RBFNN-based fault-tolerant control for robotic manipulators with prescribed performance. NEUCOM, 647, 130322. (www)

  @article{ , abstract = {This paper presents an adaptive sliding mode control (SMC) with prescribed performance (PP) integrating radial basis function neural network (RBFNN) to address the challenges of actuator faults, parameter uncertainties in robotic manipulator systems. The proposed method leverages the adaptive learning capability of RBFNN to estimate and compensate for nonlinear dynamics and fault-induced deviations, enhancing the fault tolerance of the control systems. The approach reduces reliance on precise system models and ensures robust and accurate trajectory tracking. PP is embedded within the SMC framework, guaranteeing transient and steady-state performance by confining state errors within predefined bounds, and ensuring rapid error convergence. Stability of the proposed method is rigorously analyzed using Lyapunov theory. Compared simulation results demonstrate that the proposed strategy achieves precise tracking accuracy, robustness, and rapid response, particularly under severe actuator fault scenarios.}, archive = {J_NEUCOM}, author = {Shuang Zhang and Yunong Bi and Xinbo Yu and Yifan Wu and Chenghuan Li}, doi = {10.1016/j.neucom.2025.130322}, journal = {Neurocomputing}, month = {9}, pages = {130322}, shortjournal = {Neurocomputing}, title = {Adaptive RBFNN-based fault-tolerant control for robotic manipulators with prescribed performance}, volume = {647}, year = {2025}, }
• (2025). Hierarchical cross-modal interaction network for multimodal fake news detection. NEUCOM, 647, 130308. (www)

  @article{ , abstract = {In the rapidly evolving digital information age, multimodal fake news detection has become a critical field of study to ensure the veracity of content that blends text and imagery. Existing methods utilize attention mechanism to fuse the features extracted from unimodal encoder and perform surface-level cross-modal interaction, which is inefficient due to misalignment of multimodal features and limited ability to interpret the consistency between textual and visual information. To address these issues, we present the H ierarchical C ross- M odal I nteraction N etwork (HCMIN) for multimodal fake news detection, a novel architecture that robustly achieve intra- and inter-modality interaction. Specifically, we design dual unimodal branches that independently enhance textual and visual representation, coupled with unimodal-guided co-attention mechanisms that ensure complex intra- and inter-modality feature interactions are captured. Additionally, we propose a hierarchical mutual contrastive learning strategy, which refines the alignment and interactions of modalities at both feature and decision levels, utilizing contrastive loss to spotlight discriminative characteristics of fake news. Extensive experiments on three benchmarks demonstrate the superiority of our HCMIN.}, archive = {J_NEUCOM}, author = {Zihao Li and Xiaofei Li and Jiaxin Yang and Xianghan Wang and Shuohao Li and Jun Zhang}, doi = {10.1016/j.neucom.2025.130308}, journal = {Neurocomputing}, month = {9}, pages = {130308}, shortjournal = {Neurocomputing}, title = {Hierarchical cross-modal interaction network for multimodal fake news detection}, volume = {647}, year = {2025}, }
• (2025). Informed machine learning: Excess risk and generalization. NEUCOM, 646, 130521. (www)

  @article{ , abstract = {Machine Learning (ML) has transformed both research and industry by offering powerful models capable of capturing complex phenomena. However, these models often require large, high-quality datasets and may struggle to generalize beyond the distributions on which they are trained. Informed Machine Learning (IML) tackles these challenges by incorporating domain knowledge at various stages of the ML pipeline, thereby reducing data requirements and enhancing generalization. Building on statistical learning theory, we present some theoretical comparison and insights about ML and IML excess risk and generalization performance. We then illustrate how these theoretical insights can be leveraged in practice through some practical examples. Our findings shed some light on the mechanisms and conditions under which IML can outperform traditional ML, offering valuable guidance for effective implementation in real-world settings.}, archive = {J_NEUCOM}, author = {Luca Oneto and Sandro Ridella and Davide Anguita}, doi = {10.1016/j.neucom.2025.130521}, journal = {Neurocomputing}, month = {9}, pages = {130521}, shortjournal = {Neurocomputing}, title = {Informed machine learning: Excess risk and generalization}, volume = {646}, year = {2025}, }
• (2025). Hamilton–Jacobi based policy-iteration via deep operator learning. NEUCOM, 646, 130515. (www)

  @article{ , abstract = {The framework of deep operator network (DeepONet) has been widely exploited thanks to its capability of solving high dimensional partial differential equations. In this paper, we incorporate DeepONet with a recently developed policy iteration scheme to numerically solve optimal control problems and the corresponding Hamilton–Jacobi–Bellman (HJB) equations. A notable feature of our approach is that once the neural network is trained, the solution to the optimal control problem and HJB equations with different terminal functions can be inferred quickly thanks to the unique feature of operator learning. Furthermore, a quantitative analysis of the accuracy of the algorithm is carried out via comparison principles of viscosity solutions. The effectiveness of the method is verified with various examples, including 10-dimensional linear quadratic regulator problems (LQRs).}, archive = {J_NEUCOM}, author = {Jae Yong Lee and Yeoneung Kim}, doi = {10.1016/j.neucom.2025.130515}, journal = {Neurocomputing}, month = {9}, pages = {130515}, shortjournal = {Neurocomputing}, title = {Hamilton–Jacobi based policy-iteration via deep operator learning}, volume = {646}, year = {2025}, }
• (2025). Bipartite consensus of matrix-weighted multi-agent systems with dynamic event-triggered mechanism: An adaptive observer-based approach. NEUCOM, 646, 130508. (www)

  @article{ , abstract = {This paper deals with the bipartite consensus issue for matrix-weighted nonlinear multi-agent systems (MASs). To accurately characterize the multi-dimensional attribute of state connections between agents, matrix weights are utilized to represent the connectivity relationships among agents instead of traditional scalar weights. Additionally, both collaborative and antagonistic interactions are considered on the matrix-weighted networks, which are inscribed with positive and negative semi-definite matrices, respectively. In the development of the distributed control protocol, observers are constructed to estimate the state of agents. Due to the fact that the full states of the nodes of a network are not always available, an observer-based distributed control protocol with matrix weights is constructed to achieve the desirable bipartite consensus. Meanwhile, to avoid using the global information of the communication network, dynamic coupling strengths are designed to obey an adaptive scheme. Also, a novel dynamic event-triggered mechanism is considered for the sake of reducing the communication burden. Based on the assumption that there exists a positive-negative spanning tree in matrix-weighted signed networks, employing algebraic graph theory and Lyapunov stability theory, sufficient conditions are derived to ensure that the matrix-weighted nonlinear MASs can achieve bipartite consensus. Moreover, it is demonstrated that the Zeno phenomenon can be ruled out under the proposed dynamic event-triggered mechanism. Finally, a simulation example is provided to verify the effectiveness of the proposed design method.}, archive = {J_NEUCOM}, author = {Xiaoli Zhu and Yurong Liu}, doi = {10.1016/j.neucom.2025.130508}, journal = {Neurocomputing}, month = {9}, pages = {130508}, shortjournal = {Neurocomputing}, title = {Bipartite consensus of matrix-weighted multi-agent systems with dynamic event-triggered mechanism: An adaptive observer-based approach}, volume = {646}, year = {2025}, }
• (2025). Advancing unsupervised graph anomaly detection: A multi-level contrastive learning framework to mitigate local consistency deception. NEUCOM, 646, 130507. (www)

  @article{ , abstract = {Graph anomaly detection plays a crucial role in various fields such as social networks, financial transactions, and cybersecurity. Due to the high cost of acquiring labeled node data, contrastive learning, an important paradigm in unsupervised learning, has gained significant attention and demonstrated effective performance in practical applications. Traditional contrastive learning methods evaluate anomalies by detecting inconsistencies between nodes and their neighbors or subgraphs, often expanding these scopes using techniques like Random Walk with Restart (RWR) to gather more information. However, in real-world anomaly datasets, normal and anomalous nodes often exhibit complex hybrid distributions. This complexity can deceive the consistency discrimination of contrastive learning methods under local views, leading to substantial performance drops. To address this issue, we propose TCL-GAD, a novel solution that incorporates global information for graph anomaly detection. TCL-GAD uses a Simple GCN module to extract node and neighborhood information while employing a Hierarchical Transformer to prevent over-reliance on higher-order node information, which could negatively impact the model, while efficiently extracting valuable global information. The Multi-Level Contrastive Learning framework enhances negative node sampling and effectively leverages multi-level graph information to boost overall detection performance. We evaluated TCL-GAD against thirteen state-of-the-art baseline methods across seven diverse datasets. Results show that TCL-GAD achieves excellent operational efficiency and consistently demonstrates superior performance, showing significant improvements over the runner-up methods especially on complex real-world datasets.}, archive = {J_NEUCOM}, author = {Wensen Wu and Yijun Gu}, doi = {10.1016/j.neucom.2025.130507}, journal = {Neurocomputing}, month = {9}, pages = {130507}, shortjournal = {Neurocomputing}, title = {Advancing unsupervised graph anomaly detection: A multi-level contrastive learning framework to mitigate local consistency deception}, volume = {646}, year = {2025}, }
• (2025). Synergy-driven multi-modal prompting for weakly supervised semantic segmentation. NEUCOM, 646, 130504. (www)

  @article{ , abstract = {Weakly supervised semantic segmentation has advanced rapidly by using weak labels, significantly reducing costs and labor demands. Current methods often rely on class activation maps (CAMs) from classifiers as supervision signals, but these tend to focus on highly discriminative regions, resulting in incomplete coverage. To address this, we leverage the cross-modal consistency of CLIP’s language and vision encoders, enabling natural integration of textual and visual features for improved image classification and localization. We propose SDMMPS, which uses Text-synergized visual multimodal prompts and contrastive learning to enhance object recognition and localization. The model consists of three stages: Firstly, generating initial CAMs with a traditional classification network as a foundation for optimization. Secondly, creating “multi-modal prompts” by integrating CLIP’s text embeddings via a mapping function for richer semantic information. Finally, applying a prompt-driven suppression strategy to refine the focus on target regions while suppressing the background. Extensive experiments were conducted on PASCAL VOC 2012 and MS COCO 2014, achieving state-of-the-art performance. Code will be released soon.}, archive = {J_NEUCOM}, author = {Yaping Zhao and Chengyun Liu and Zhenyan Wang and Xiaona Peng and Zhenxue Chen and Q.M. Jonathan Wu}, doi = {10.1016/j.neucom.2025.130504}, journal = {Neurocomputing}, month = {9}, pages = {130504}, shortjournal = {Neurocomputing}, title = {Synergy-driven multi-modal prompting for weakly supervised semantic segmentation}, volume = {646}, year = {2025}, }
• (2025). Improving consistency of marine mammals tracking in challenging drone recordings through visual particle filter integration. NEUCOM, 646, 130503. (www)

  @article{ , abstract = {Tracking marine mammals is essential for gaining insights into their health, behaviour, and population dynamics. With the growing use of drones in ecological research, the demand for efficient, automated video analysis has increased. This paper introduces a novel tracking algorithm designed explicitly for online marine mammals tracking in UAV-recorded videos, addressing unique challenges such as complex seabed textures, water reflections, and irregular animal movements. By integrating particle filters with the widely used Simple Online and Real-time Tracking (SORT) algorithm, we tackle issues of missed detections in marine contexts, enhancing tracking robustness. We also present a new publicly available dataset for drone-based porpoise tracking and evaluate our SORT-PF against state-of-the-art tracking methods. The results demonstrate significant improvements, including a 9.3% increase in the HOTA metric, simultaneously keeping the lowest ID switches compared to baseline methods, highlighting the potential of the proposed SORT-PF for automated wildlife monitoring and other challenging tracking applications.}, archive = {J_NEUCOM}, author = {Bartosz Ptak and Henrik Skov Midtiby and Marek Kraft}, doi = {10.1016/j.neucom.2025.130503}, journal = {Neurocomputing}, month = {9}, pages = {130503}, shortjournal = {Neurocomputing}, title = {Improving consistency of marine mammals tracking in challenging drone recordings through visual particle filter integration}, volume = {646}, year = {2025}, }
• (2025). Correntropy based label loss for multi-classification on deep neural networks. NEUCOM, 646, 130500. (www)

  @article{ , abstract = {The success of deep learning heavily relies on large-scale labeled datasets, but manually labeled datasets that inevitably have errors. However, the network is highly susceptible to noisy labels, which seriously degenerate the learning performance of the network. Thus, training neural networks on datasets with noisy labels is a substantial challenge. As a nonlinear and local similarity metric, correntropy is not sensitive to outliers. Based on the features of correntropy, this paper proposes a novel loss function called Correntropy based Label Loss (CLL). The CLL can connect the output of the Softmax layer and utilize the properties of the layer’s output; thus, it is suitable for multi-classification problems with one-hot encoded labels and is naturally applied to networks with the Softmax layer. Specifically, when the distance between two random variables exceeds a particular threshold, the influence on the network can be alleviated by appropriately selected kernel bandwidth. Therefore, for data contaminated by noisy labels, the loss function CLL can alleviate the effects of the noisy labels in the data and let the network effectively learn information from the data with correct labels. We give theoretical and gradient analyses of the CLL loss to prove that CLL is robust to the noisy labels. In the MNIST dataset with a symmetric noise 60%, the model trained by CLL has an accuracy of up to 96.81%, which is 43.43% higher than the CE. Furthermore, the experiments conducted on the five publicly available datasets illustrate that the network with CLL loss outperforms the other state-of-the-art robust loss in most cases.}, archive = {J_NEUCOM}, author = {Qing Deng and Nan Zhou and Wenjun Luo and Yuanhua Du and Kaibo Shi and Badong Chen}, doi = {10.1016/j.neucom.2025.130500}, journal = {Neurocomputing}, month = {9}, pages = {130500}, shortjournal = {Neurocomputing}, title = {Correntropy based label loss for multi-classification on deep neural networks}, volume = {646}, year = {2025}, }
• (2025). Hybrid graph structure learning for improving semantic dependency parsing with robust graph neural networks. NEUCOM, 646, 130482. (www)

  @article{ , abstract = {Graph neural networks (GNN) can learn powerful representations of graphs, but their effectiveness is heavily influenced by the quality of the graph structure used. Graph structure learning (GSL) was developed to extract useful graph structures from data and learn better graph representations using GNN. GSL creates an adjacency matrix with two multilayer perceptrons and a biaffine attention network. However, it does not directly incorporate the principles of homophily, sparsity, and degree distribution, which are important in many real-life situations. Differentiable graph structure learning neural networks (DGSLN), which adhere to these basic principles, have been proven to improve GNN performance. This study proposes a GNN-based semantic dependency parsing model that incorporates the GSL adapted from the DGSLN approach to generate an initial graph. For the graph representation learning stage, we use GNN variants, GIN and GATv2, which have shown robust performance in previous research. The proposed model significantly improves previous best models, achieving an average F1 score of 93.77% on in-domain and 92.27% on out-of-domain datasets. However, our model’s parsing performance is slower than prior top-performing models. The model utilizing GIN exhibits the highest level of performance and shows exceptional performance in semantic dependency parsing on English datasets.}, archive = {J_NEUCOM}, author = {Irkham Huda and Moh Edi Wibowo and Reza Pulungan}, doi = {10.1016/j.neucom.2025.130482}, journal = {Neurocomputing}, month = {9}, pages = {130482}, shortjournal = {Neurocomputing}, title = {Hybrid graph structure learning for improving semantic dependency parsing with robust graph neural networks}, volume = {646}, year = {2025}, }
• (2025). DUGCN: Deep unfolding gradient consistency network for fast MRI reconstruction. NEUCOM, 646, 130481. (www)

  @article{ , abstract = {Compressed Sensing (CS) theory is a powerful technique to recover signal from highly under-sampled acquisition. It has been widely applied in fast Magnetic Resonance Imaging (MRI) reconstruction, effectively reducing MRI scan times and improving the patient experience. However, existing CS-MRI methods typically focus on minimizing the pixel-wise reconstruction error, and neglect the MR image characteristics, failing to restore high-frequency information. In this paper, we propose an efficient and reliable Deep Unfolding Gradient Consistency Network for fast MRI reconstruction, dubbed as DUGCN. First, to solve the ill-posed inverse problem, we consider a novel generalized CS energy model with implicit regularization by introducing a gradient consistency term, which encourages the model to ensure high-frequency consistency and pay more attention to delicate structure details. Then, we unfold the novel CS energy model using Half-Quadratic-Splitting algorithm and obtain four iterative sub-problems. We further develop corresponding deep architectures to effectively solve these subproblems with learnable parameters, and construct a end-to-end trainable DUGCN that combines the interpretable model-based methods and data-driven learning-based methods. Extensive experiments on various sampling patterns and acceleration rates demonstrate that our DUGCN can achieve high-quality MRI reconstruction with more accurate details, and outperforms the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Hui Fang and Le Hu and Wenbin Yin and Faming Fang and Shaoxin Li}, doi = {10.1016/j.neucom.2025.130481}, journal = {Neurocomputing}, month = {9}, pages = {130481}, shortjournal = {Neurocomputing}, title = {DUGCN: Deep unfolding gradient consistency network for fast MRI reconstruction}, volume = {646}, year = {2025}, }
• (2025). AgentBuilder: Automating agent creation via large language model-driven systems. NEUCOM, 646, 130476. (www)

  @article{ , abstract = {Building agents capable of handling diverse tasks and evolving themselves autonomously has been a long-term goal in the AI community. Large Language Models (LLMs) are considered a promising foundation for constructing such agents due to their generalized capabilities. Current approaches either rely on LLM-based agents imitating expert-provided workflows step-by-step, requiring human supervision, which hinders scalability and exploration; or allow agents to explore and learn in isolated environments, resulting in specialist agents with limited generalization. In this paper, we take a significant step toward building self-evolving, generally capable agents using an automated agent generation framework. Our key contributions include: (1) a Role Builder that automatically generates task-specific agents capable of self-evolution and adaptable across diverse tasks; (2) a LLM Manager that dynamically selects and integrates multiple LLMs for optimizing agent performance in varied environments; and (3) an evolution mechanism that allows agents to refine themselves based on task execution outcomes. Experimental results show that agents generated by our framework achieve competitive performance in diverse environments while continuously improving through self-evolution.}, archive = {J_NEUCOM}, author = {Wang Tang and Heng Wei Zhang and Jianuo Huang and Shaowen Wang and Feifan Yu and Hui Yang and Yu Wang}, doi = {10.1016/j.neucom.2025.130476}, journal = {Neurocomputing}, month = {9}, pages = {130476}, shortjournal = {Neurocomputing}, title = {AgentBuilder: Automating agent creation via large language model-driven systems}, volume = {646}, year = {2025}, }
• (2025). Anxiety recognition based on multimodal social media data and cross-attention mechanism. NEUCOM, 646, 130473. (www)

  @article{ , abstract = {Anxiety disorder is a common mental illness that involves persistent and recurrent episodes of intense anxiety and sudden feelings of fear or terror, which seriously affect the patient’s study, work and life. Social media contains a large amount of multimodal data reflecting people’s inner activities and emotional states, providing a new way for recognizing anxiety disorders. However, existing methods tend to excessively focus on the text data, and neglect the role of other types of data such as images. To address this issue, this paper proposes a Multimodal Anxiety Recognition model (MAR-IDCA) based on Image Description and Cross-Attention mechanism. Firstly, for each image of the user’s post, caption generation model is applied to obtain the abstract semantic information of the image, and the possible character information in the image is obtained through optical character recognition technology, so as to generate the description of the image. Secondly, through an image encoder and two text encoders, the visual feature of the image, the textual features of the image description and the post text are extracted respectively. Then, a multi-head cross-attention mechanism is employed to fuse the visual feature, image description feature and post text feature. Finally, combined with other auxiliary information of user posts, the final multimodal data representation is obtained and classified. Experiments on two multimodal anxiety recognition datasets show that, compared with the existing models, MAR-IDCA has better anxiety recognition performance.}, archive = {J_NEUCOM}, author = {Jianghong Zhu and Zhenwen Zhang and Zepeng Li and Bin Hu}, doi = {10.1016/j.neucom.2025.130473}, journal = {Neurocomputing}, month = {9}, pages = {130473}, shortjournal = {Neurocomputing}, title = {Anxiety recognition based on multimodal social media data and cross-attention mechanism}, volume = {646}, year = {2025}, }
• (2025). Adaptive continuous fractional-order nonsingular terminal sliding mode control based on neural network for PMLM system with actuator saturation. NEUCOM, 646, 130468. (www)

  @article{ , abstract = {This paper addresses the trajectory tracking control problem for permanent magnet linear motor (PMLM) systems, presenting a novel fractional-order nonsingular terminal sliding mode control (SMC) strategy integrated with a radial basis function (RBF) neural network approximator. The proposed method leverages a fractional-order sliding surface to achieve superior convergence precision compared to conventional nonsingular SMC approaches in finite time. To handle unknown disturbances and lumped uncertainties, an RBF neural network is employed for adaptive approximation, complemented by a dynamically adjustable compensation term to mitigate approximation errors under varying conditions. Actuator saturation, a practical constraint, is explicitly addressed through an auxiliary system design, ensuring robust motion performance. Compared with integer-order sliding mode control, the proposed scheme demonstrates faster convergence and enhanced tracking accuracy, validated theoretically via Lyapunov stability analysis. Simulation results are provided to confirm the efficacy of the control strategy against state-of-the-art methods, with root mean square error (RMSE) serving as a quantitative metric to support the experimental findings. This work has made obvious advances by integrating fractional-order dynamics, adaptive neural approximation, and saturation compensation, offering a robust solution with clear performance improvements for PMLM system.}, archive = {J_NEUCOM}, author = {Haoran Zhang and Jinyong Yu and Pengwei Shi and Shenglin Hu and Lin Zhao}, doi = {10.1016/j.neucom.2025.130468}, journal = {Neurocomputing}, month = {9}, pages = {130468}, shortjournal = {Neurocomputing}, title = {Adaptive continuous fractional-order nonsingular terminal sliding mode control based on neural network for PMLM system with actuator saturation}, volume = {646}, year = {2025}, }
• (2025). NN based adaptive FTC for fractional-order time-varying delays system with actuator faults. NEUCOM, 646, 130466. (www)

  @article{ , abstract = {In this paper, neural-network (NN) based on adaptive fault-tolerant control (FTC) for fractional-order time-varying delay systems (FOTVS) with actuator faults is studied. First, a unified linear matrix inequality (LMI) structure is proposed to stabilize FOTVS across the full fractional-order range 0 < α < 2 , resolving the critical limitation of existing methods that demand separate analyses for 0 < α < 1 and 1 < α < 2 . By integrating pseudo-state matrices and fractional Lyapunov functionals, this novel LMI framework is designed to eliminate redundant decision variables and entirely avoid complex-valued matrix operations, thereby reducing computational complexity compared to conventional partitioned approaches. Then, a critic NN with adaptive laws is proposed to dynamically approximate unknown nonlinearities and compensate for actuator faults . In addition a fractional-order Razumikhin theorem rigorously ensures robustness against time-varying delays and disturbances without conservative assumptions. Lastly, to show that the presented procedure is effective, the outcome of the simulation and the numerical example of fractional L u ̈ systems are presented.}, archive = {J_NEUCOM}, author = {Shuhang Yu and Huaguang Zhang and Jiayue Sun and Juan Zhang}, doi = {10.1016/j.neucom.2025.130466}, journal = {Neurocomputing}, month = {9}, pages = {130466}, shortjournal = {Neurocomputing}, title = {NN based adaptive FTC for fractional-order time-varying delays system with actuator faults}, volume = {646}, year = {2025}, }
• (2025). Optimized category relation and category individuation for source-free domain adaptation. NEUCOM, 646, 130465. (www)

  @article{ , abstract = {Regarding the closed-set source-free domain adaptation, there exists an intrinsic correlation between the source and target domains. Consequently, it is feasible to develop a target model that performs effectively on the target domain by leveraging knowledge transferred from the source model, even in the absence of source data. The source and target domains share an identical set of categories, and there exists an inherent relation between these categories. Consequently, the category relations in both domains exhibit a degree of correlation. However, it is not appropriate to naively assume that the category relations are identical across the two domains. Simply enforcing consistency between the two domains is not an effective solution to the problem of domain shifts. Therefore, it is essential to identify a more appropriate intermediary mechanism to bridge the source and target domains. Based on this consideration, we design an O ptimized C ategory R elation ( OCR ) which not only retains the knowledge of the source domain but also aligns with the data style of the target domain. Specifically, the mean-teacher architecture is employed, leveraging its inherent dual-model structure. Herein,the student model is responsible for learning the knowledge of the target domain, and the teacher model provides the target information for OCR. At the same time, the source category relations also constitute a crucial component. Furthermore, we conduct the C ategory I ndividuation ( CI ) operation on the target data to enhance the discrimination among different categories. We achieved excellent classification accuracies on four image datasets, which verified the effectiveness of our method.}, archive = {J_NEUCOM}, author = {Shumin Liang and Xiaorong Hou and Yajian Zeng}, doi = {10.1016/j.neucom.2025.130465}, journal = {Neurocomputing}, month = {9}, pages = {130465}, shortjournal = {Neurocomputing}, title = {Optimized category relation and category individuation for source-free domain adaptation}, volume = {646}, year = {2025}, }
• (2025). Updating KernelCanvas for weightless graph classification. NEUCOM, 646, 130458. (www)

  @article{ , abstract = {In a landscape where graph neural networks establish themselves as a standard in graph learning, we introduce a contrasting weightless graph learning architecture. We propose a novel representation for shallow graph learning based on weightless neural networks (WNN) and extend current WNN quantization techniques. Additionally, we assess this new architecture across widely adopted datasets. Our findings indicate that employing our representation and WNN models can yield superior results to those attained with extra-large graph neural networks or heavyweight kernel methods while imposing fewer demands on computer resources and memory. Furthermore, our results suggest that other classifiers can leverage the proposed representation for improved performance.}, archive = {J_NEUCOM}, author = {Raul Bezerra Barbosa and Diego Leonel Cadette Dutra and Priscila Machado Vieira Lima and Diego Carvalho and Felipe Maia Galvão França}, doi = {10.1016/j.neucom.2025.130458}, journal = {Neurocomputing}, month = {9}, pages = {130458}, shortjournal = {Neurocomputing}, title = {Updating KernelCanvas for weightless graph classification}, volume = {646}, year = {2025}, }
• (2025). Neural network-based adaptive prescribed-time bipartite flocking for uncertain networked multi-agent systems. NEUCOM, 646, 130452. (www)

  @article{ , abstract = {Flocking is a fundamental self-organizing behavior observed in networked agent systems (NASs), wherein agents achieve coordinated group dynamics through mutual interactions. While in dynamic environmental contexts, the demand for flocking behavior to demonstrate rapid responsiveness and robust stability becomes more critical. With this in mind, this paper addresses the adaptive bipartite flocking control problem for NASs, particularly in the presence of compound uncertainties and convergence time constraints. A robust adaptive neural network-based prescribed-time bipartite flocking controller is developed to ensure that, despite uncertainties, all agents achieve flocking behavior within a predefined time. Notably, the settling time can be predefined and remains independent of system parameters such as controller gain, initial agent states, and the communication topology among agents. Additionally, by analyzing the stability conditions of the closed-loop error system, an adaptive weight update law for the neural network estimator is formulated. This updated law allows for effective uncertainty estimation through backpropagation of the flocking control error. Finally, the effectiveness and superiority of the proposed prescribed-time bipartite flocking control scheme are validated through numerical simulations.}, archive = {J_NEUCOM}, author = {Xian Qing and Weihao Li and Bowen Chen and Boxian Lin and Mengji Shi and Kaiyu Qin}, doi = {10.1016/j.neucom.2025.130452}, journal = {Neurocomputing}, month = {9}, pages = {130452}, shortjournal = {Neurocomputing}, title = {Neural network-based adaptive prescribed-time bipartite flocking for uncertain networked multi-agent systems}, volume = {646}, year = {2025}, }
• (2025). Accelerated weka: GPU machine learning with weka workbench. NEUCOM, 646, 130432. (www)

  @article{ , abstract = {Recent advancements in Machine Learning (ML) have driven increased interest and demand in the field. However, the complexity and mathematical background required for many techniques can be challenging for newcomers and enthusiasts. For many years, Weka has been a cornerstone in introductory ML courses worldwide, offering a user-friendly interface that facilitates understanding of various methods. As an open-source software with an active community, Weka has continued to evolve with new attributes. With the growing size of datasets, users are increasingly looking for ways to accelerate computations. GPUs have become a preferred solution for efficiently building and deploying ML models. This paper presents Accelerated Weka, a software framework that integrates GPU-accelerated methods within Weka’s intuitive graphical user interface, significantly reducing execution times for large datasets while preserving Weka’s accessibility. Benchmark results indicate that Accelerated Weka can achieve 2,198x speedup on an A100 chip. The framework retains Weka’s GPL 3.0 license and offers a straightforward installation process through the Conda environment.}, archive = {J_NEUCOM}, author = {Guilherme Weigert Cassales and Justin Jia Liu and Albert Bifet}, doi = {10.1016/j.neucom.2025.130432}, journal = {Neurocomputing}, month = {9}, pages = {130432}, shortjournal = {Neurocomputing}, title = {Accelerated weka: GPU machine learning with weka workbench}, volume = {646}, year = {2025}, }
• (2025). Aesthetic-aware adversarial learning network for artistic style transfer. NEUCOM, 646, 130431. (www)

  @article{ , abstract = {Although artistic style transfer has been widely studied, the existing methods are either difficult to effectively convey aesthetic information from the style or unable to effectively maintain content features. To alleviate this issue, in this work we propose an aesthetic-aware adversarial learning network for artistic style transfer (AAALNet). Specifically, we first introduce an artistic style attention module, which can adaptively fuse style and content features by taking into account the global channel-wise distribution of style images and the local spatial semantics of content images. In addition, we introduce an residual pyramid artistic discriminator through adversarial learning, which not only acts as a discriminator to determine whether the generated image belongs to a real art work, but also acts as a feature extractor to learn universal artistic features from a large number of art paintings. Finally, we introduce attention-based content loss to enhance content retention by aligning content images with stylized images’ self-attention maps. Compared with the current eight popular methods, AAALNet achieves the highest content fidelity of 0.7579, the second best results of content loss and neural image evaluation (NIMA) are 1.2814 and 4.6348, respectively, and the stylization effect also achieves competitive results. Extensive experiments show that our proposed AAALNet achieves good content retention while maintaining the artistic sense. Our code is available at https://github.com/yixiu351/AAALNet .}, archive = {J_NEUCOM}, author = {Yongsheng Dong and Songtao Liu and Yaxin Li and Lintao Zheng}, doi = {10.1016/j.neucom.2025.130431}, journal = {Neurocomputing}, month = {9}, pages = {130431}, shortjournal = {Neurocomputing}, title = {Aesthetic-aware adversarial learning network for artistic style transfer}, volume = {646}, year = {2025}, }
• (2025). Enhancing intent understanding for ambiguous prompt: A human–machine co-adaption strategy. NEUCOM, 646, 130415. (www)

  @article{ , abstract = {Modern image generation systems have demonstrated the ability to produce realistic and high-quality visuals. However, user prompts often contain ambiguities, making it challenging for these systems to accurately interpret user intent. As a result, users frequently need to revise their prompts multiple times to achieve the desired output. While existing approaches aim to improve prompt quality, they often fail to address the nuanced requirements of non-expert users effectively. We introduce Visual Co-Adaptation (VCA), a novel framework designed to iteratively refine user prompts and align generated images with user preferences. VCA leverages a pre-trained language model fine-tuned via reinforcement learning and integrates multi-turn dialogues for prompt disambiguation. The framework employs a three-step process to align user intent with generated images. First, an interactive dialogue system (Incremental Context-Enhanced Dialogue Block) refines prompts progressively by asking users targeted follow-up questions and incorporating their feedback to reduce ambiguities. Second, to resolve complex or niche requests, the system(Semantic Exploration and Disambiguation Module (SESD)) cross-references external databases and uses a visual-text matching algorithm (CLIP-based scoring) to rank generated images based on how closely they reflect the prompt’s description. Finally, a detail optimization mechanism (Pixel Precision and Consistency Optimization Module (PPCO)) fine-tunes textures and lighting while ensuring all elements in the image logically coexist. This process combines reinforcement learning with multi-layered consistency checks to balance creativity and coherence. A human-in-the-loop feedback mechanism further enhances model performance by integrating user insights into the training pipeline. Extensive experiments demonstrate that VCA significantly improves user satisfaction, image-text alignment, and aesthetic quality. Compared to state-of-the-art models such as DALL-E 3, Stable Diffusion, and Imagen, VCA reduces the average number of dialogue rounds to 4.3, achieves a CLIP score of 0.92, and increases user satisfaction to 4.73/5. Additionally, we contribute a novel multi-round dialogue dataset comprising prompt-image pairs and user intent annotations, enabling further experimentation and showcasing the efficacy of our approach. The dataset page is available at https://huggingface.co/datasets/naseele/Image_prompt/tree/main .}, archive = {J_NEUCOM}, author = {Yijin Wang and Yangfan He and Jianhui Wang and Kun Li and Li Sun and Jun Yin and Miao Zhang and Xueqian Wang}, doi = {10.1016/j.neucom.2025.130415}, journal = {Neurocomputing}, month = {9}, pages = {130415}, shortjournal = {Neurocomputing}, title = {Enhancing intent understanding for ambiguous prompt: A human–machine co-adaption strategy}, volume = {646}, year = {2025}, }
• (2025). MPMixer: Multi-scale patch mixer for long-term time series forecasting. NEUCOM, 646, 130398. (www)

  @article{ , abstract = {Multi-scale-based time series forecasting (TSF) models transform 1-dimensional time series into 2-dimensional ones and make predications. They make it easier to capture implicit information in time series. However, the existing methods ignore the correlation between series with different scales. This paper proposes a model, named Multi-scale Patch Mixer (MPMixer), for Long-Term Time Series Forecasting. MPMixer is a double-ended prediction model, where one end analyzes the correlation within multi-scale series and predicts, the other one analyzes the correlation between patches from multi-scale series and makes predictions. The final forecasting results are achieved by summarizing predications generated by two ends. The model first downsamples the raw time series and obtains the multi-scale series. While analyzing and predicting based on the inter multi-scale series, the multi-scale series are patched into a 2-dimensional patch series. After that, the multi-scale patch encoder analyzes patch series and makes predictions. In the other end, analyzing and predicting based on the intra multi-scale series, the model first decomposes the multi-scale series into trends and seasons, then constructs their patch sequences and generates predictions using a linear module respectively. In the experiments, MPMixer achieves state-of-the-art on 9 challenging real-world datasets, which verifies the rationality of our model. We also analyze the impact of downsampling setting, patch size, and lookback length on performance, and conduct ablation experiments to verify the importance of each component. Code is available at https://github.com/Xoxyer/MPMixer .}, archive = {J_NEUCOM}, author = {Yijun Xu and Weijie Zhang and Huixia Lai and Shi Zhang and Ruliang Xiao}, doi = {10.1016/j.neucom.2025.130398}, journal = {Neurocomputing}, month = {9}, pages = {130398}, shortjournal = {Neurocomputing}, title = {MPMixer: Multi-scale patch mixer for long-term time series forecasting}, volume = {646}, year = {2025}, }
• (2025). CausalSR: Structural causal model-driven super-resolution with counterfactual inference. NEUCOM, 646, 130375. (www)

  @article{ , abstract = {Physical factors interacting with sensor characteristics create complex image degradation patterns. Despite advances in deep learning-based super-resolution, existing methods overlook the causal nature of degradation by adopting simplistic black-box mappings. This paper formulates super-resolution using structural causal models to reason about image degradation processes. We establish a mathematical foundation that unifies principles from causal inference, deriving necessary conditions for identifying latent degradation mechanisms and corresponding propagation. We propose a novel counterfactual learning strategy that leverages semantic guidance to reason about hypothetical degradation scenarios, leading to theoretically-grounded representations that capture invariant features across different degradation conditions. The framework incorporates an adaptive intervention mechanism with provable bounds on treatment effects, allowing precise manipulation of degradation factors while maintaining semantic consistency. Through extensive empirical validation, we demonstrate that our approach achieves significant improvements over state-of-the-art methods, particularly in challenging scenarios with compound degradations. The model provides interpretable insights through disentangled representations of degradation factors, enabling visualization of how elements like blur and noise affect reconstruction. This allows precise analysis of degradation mechanisms and their impact on restoration quality. On standard benchmarks, our method outperforms existing approaches by significant margins, while providing interpretable insights into the restoration process. The empirical results demonstrate the fundamental importance of causal reasoning in understanding image restoration systems.}, archive = {J_NEUCOM}, author = {Zhengyang Lu and Bingjie Lu and Feng Wang}, doi = {10.1016/j.neucom.2025.130375}, journal = {Neurocomputing}, month = {9}, pages = {130375}, shortjournal = {Neurocomputing}, title = {CausalSR: Structural causal model-driven super-resolution with counterfactual inference}, volume = {646}, year = {2025}, }
• (2025). Bi-level evolutionary model tree chain induction for multi-output regression. NEUCOM, 646, 130280. (www)

  @article{ , abstract = {Multi-output Regression (MOR) is a machine learning technique that aims to predict several values simultaneously. Some existing approaches addressed this problem by decomposing the MOR problem into separate single-target ones. However, in real-world applications, it is more advantageous to exploit the inter-target correlations in the prediction task. Some other approaches proposed simultaneous prediction but they are based on greedy algorithms and are prone to fall easily into local optima. In order to solve these issues, we propose a novel approach called Bi-level Evolutionary Model Tree Chain Induction (BEMTCI) which is able to deal with multi-output datasets using a bi-level evolutionary algorithm. BEMTCI evolves a population of Model Tree Chains (MTCs) where each Model Tree (MT) focuses on the prediction of one single target. The upper-level explores different orderings of the MTs of each MTC to find the best chaining order which is able to express the relationships among the output variables. A further optimization is performed in the lower-level of BEMTCI which concerns the linear models at the leaves of the MTs. The experimental study showed the effectiveness of our approach compared to the existing ones when applied on sixteen MOR datasets. The genetic operators employed in our BEMTCI ensure the variation of the population and guarantee a fair and a precise prediction due to the evaluation process. The obtained results prove the performance of our BEMTCI in solving MOR problems.}, archive = {J_NEUCOM}, author = {Safa Mahouachi and Maha Elarbi and Slim Bechikh}, doi = {10.1016/j.neucom.2025.130280}, journal = {Neurocomputing}, month = {9}, pages = {130280}, shortjournal = {Neurocomputing}, title = {Bi-level evolutionary model tree chain induction for multi-output regression}, volume = {646}, year = {2025}, }
• (2025). A locally active memristive synapse-coupled HR-FN neural network and predefined time synchronization. NEUCOM, 645, 130540. (www)

  @article{ , abstract = {In this paper, a novel locally active memristor is first proposed, and its characteristics are analyzed using standard nonlinear theoretical analysis methods. Then, the locally active memristor is used to simulate biological synapses, and a synapse-coupled HR-FN neuron network is further constructed. In addition, the equivalent circuit of the coupled neural network is realized through simulation, and the correctness of the numerical simulation is verified. A novel synchronization controller is designed using the sliding mode control method, and it is demonstrated that under the influence of this controller, the error system can reach the sliding surface. Finally, the predefined-time heterogeneous synchronization of a 4D system and the proposed neural network is achieved in the presence of internal and external disturbances. The study provides an important basis for engineering applications based on neurons and biological neuron experiments and has potential research values.}, archive = {J_NEUCOM}, author = {Hanbing Zhang and Shaohui Yan and Yuyan Zhang}, doi = {10.1016/j.neucom.2025.130540}, journal = {Neurocomputing}, month = {9}, pages = {130540}, shortjournal = {Neurocomputing}, title = {A locally active memristive synapse-coupled HR-FN neural network and predefined time synchronization}, volume = {645}, year = {2025}, }
• (2025). A scalable digital assets framework for distributed robot system’s anomaly detection based on hybrid convolutional autoencoder. NEUCOM, 645, 130516. (www)

  @article{ , abstract = {The scale of robots in industry is growing up rapidly. Using data-driven methods to create tradable digital assets for large-scale devices is beneficial for achieving knowledge sharing and capability transfer among different users. The article implements an anomaly detection framework in robot manufacturing area, which are built by hybrid convolutional autoencoder. It approximates the dynamic model between motion and torque of robot, then, terminals judge anomaly with KL divergence between really torque and estimated torque. The framework shares those weight file as robot’s digital asset to the distributed network and allow load balancing algorithm to apply to it. The article ultimately designed a scalable management system based on terminal clusters, which can adjust synchronously with the increase or decrease of device size to avoid overload and waste. The experiment shows that the proposed framework can ensure that each terminal’s CPU usage is less than 62 % in the five-months durability and anomaly test. The RMSE for torque estimation is below 2 %, the F1-score for anomaly detection is over 99 %. Compared to previous time-series-based methods, there has been a significant improvement in the collected robot dataset. The proposed method has positive significance for establishing digital twin projects in real-time distributed information systems.}, archive = {J_NEUCOM}, author = {Shijie Wang and Jianfeng Tao and Qincheng Jiang and Wei Chen and Chengjin Qin and Pencheng Xia and Chengliang Liu}, doi = {10.1016/j.neucom.2025.130516}, journal = {Neurocomputing}, month = {9}, pages = {130516}, shortjournal = {Neurocomputing}, title = {A scalable digital assets framework for distributed robot system’s anomaly detection based on hybrid convolutional autoencoder}, volume = {645}, year = {2025}, }
• (2025). PLTN: Noisy label learning in long-tailed medical images with adaptive prototypes. NEUCOM, 645, 130514. (www)

  @article{ , abstract = {Label noise poses a significant challenge in medical image classification, particularly for minority diseases in long-tailed distribution. Most Noisy Label Learning methods rely on clean sample selection to learn representative patterns. However, the imbalanced data distribution and intra-class variability cause model bias towards head classes at the inter-class level and dominant patterns at the intra-class level. Existing methods show limited performance in handling these biases simultaneously, causing inefficient clean sample selection and hindering minority class training. Thus, we propose PLTN, named P rototype-guided L ong- T ail N oisy Learning, a novel approach that selects comprehensive clean samples from each class and enhances minority class training to improve diagnostic accuracy of rare disease in long tailed medical data. First, Adaptive Multi-Prototype Learning (AMPL) module is designed to enable precise selection of clean samples, especially for minority classes. It achieves this by capturing diverse feature patterns through a dynamic prototype learning approach. The clean samples are then treated as labeled data while the noisy samples are used as unlabeled data in a semi-supervised learning paradigm. Second, we introduce Sharpness-aware Distributional Mixup (SDM) to enhance minority class training. SDM employs distributional mixup to enrich tail classes representations, simultaneously minimizing loss landscape sharpness to enhance model generalization for all classes, benefiting tail classes. Experiments on three typical medical datasets demonstrate that our method outperforms state-of-the-art methods. On HAM10000 dataset with 20 % label noise, we achieve an 80.34 % Macro-F1 score, with notable gains in tail class performance while maintaining head class performance, showing potential for real-world clinical applications.}, archive = {J_NEUCOM}, author = {Zhiqing He and Jie Xu and Peng Huang and Qin Wang and Yuanyuan Wang and Yi Guo}, doi = {10.1016/j.neucom.2025.130514}, journal = {Neurocomputing}, month = {9}, pages = {130514}, shortjournal = {Neurocomputing}, title = {PLTN: Noisy label learning in long-tailed medical images with adaptive prototypes}, volume = {645}, year = {2025}, }
• (2025). Sarcasm detection with induced sentimental cues using heuristic search based on unconstrained optimisation learning quantifying callousness on social media. NEUCOM, 645, 130499. (www)

  @article{ , abstract = {Sarcasm is an intent of speech used to mock and ridicule others. Existing sarcasm detection methods mainly focus on analysing the text using various natural language processing techniques. However, these methods neglect the subtle knowledge of context and intonations inherent in the sarcastic text. Additionally, they fail to analyse the connection between sentimental cues and intricately inhibited intent in sentences, which leads to sarcasm. To address these gaps, we introduce a mathematical optimisation method integrates with lexicon semantic models and entropy distributions for enhanced sarcasm detection. The proposed method helps to analyse the inhibited intent and sentimental cues of sentences for sarcasm detection. Additionally, the nature of sarcastic comments, either as callous or not, is demonstrated in this study for the first time. This knowledge of callousness helps to create public awareness of erroneous comments on social media. The proposed method has been tested on four benchmark datasets. The results of the proposed method show a significant improvement in performance for sarcasm detection. Furthermore, the proposed method has outstanding reliability compared to existing state-of-the-art methods in terms of accuracy, precision, recall, and F1 scores.}, archive = {J_NEUCOM}, author = {Himani Pokhriyal and Goonjan Jain}, doi = {10.1016/j.neucom.2025.130499}, journal = {Neurocomputing}, month = {9}, pages = {130499}, shortjournal = {Neurocomputing}, title = {Sarcasm detection with induced sentimental cues using heuristic search based on unconstrained optimisation learning quantifying callousness on social media}, volume = {645}, year = {2025}, }
• (2025). MSPE-fusion: A multimodal 3D object detection method with multi-sensor perception enhanced fusion. NEUCOM, 645, 130486. (www)

  @article{ , abstract = {The outdoor scene 3D object detection algorithm based on LiDAR often encounters challenges related to the missing and false detection of small distant targets and occluded objects. These issues arise from the sparsity and disorder of point clouds. To address these challenges, a multi-sensor perception-enhanced collaborative fusion 3D object detection method is proposed, termed MSPEFusion. This framework integrates perception information from LiDAR and visual data to enhance the feature representation of small distant targets and occluded objects at close range. Specifically, a multi-dimensional point cloud completion module (MPCM) is designed to utilize the perception information from images to complete sparse targets and enhance the contour information of point clouds. A depth image generation module (DMGM) is designed to leverage the coordinate transformation relationships between sensors to enhance the depth perception of image data, effectively reducing feature discrepancies across different modalities. Furthermore, to enhance semantic consistency during the multimodal perception feature fusion process, an attention-based feature alignment and fusion module (FAFM) is designed, which adaptively aggregates pixel-level image features for each voxel by learning the correlations and differences between the two modalities, enriching the semantic information of point clouds and improving the detection accuracy of small distant targets and occluded objects. Experimental results on the large-scale KITTI dataset show that our algorithm in this paper can achieve 82.87 mAP for cars and 65.80 mAP for pedestrians. Our best model can achieve 71.63 NDS on the nuScenes dataset. These results validate the effectiveness of the proposed method in addressing the issues of false positives and missed detections for small distant targets. The code will be released at https://github.com/lyanyushuainan/MSPE-Fusion .}, archive = {J_NEUCOM}, author = {Shuainan Yu and Xichao Wang and Yan Ci and Yawen Li and Min Zhou and Xiangqing Dong and Meng Cai}, doi = {10.1016/j.neucom.2025.130486}, journal = {Neurocomputing}, month = {9}, pages = {130486}, shortjournal = {Neurocomputing}, title = {MSPE-fusion: A multimodal 3D object detection method with multi-sensor perception enhanced fusion}, volume = {645}, year = {2025}, }
• (2025). Quality boost of tabular data synthesis using interpolative cumulative distribution function decoding and type-specific conditioner. NEUCOM, 645, 130484. (www)

  @article{ , abstract = {Tabular data synthesis is an important research area in terms of privacy and data utilization. To enhance the utilization of tabular data, data synthesis techniques are extensively explored. The primary goal of tabular data synthesis is to generate high-quality data that preserve original insights while reducing the risk of data breaches. In this study, we propose a novel generative adversarial network (GAN) for quality boost of tabular data synthesis. The method of transforming continuous variables and correct conditioning for capturing dependencies between variables is considered a critical factor in determining data quality. Therefore, our proposed method uses interpolative cumulative distribution function (CDF) decoding for continuous columns and type-specific conditioner. Interpolative CDF decoding addresses a limitation of the inverse CDF method that restricts the diversity of synthetic data. In addition, the type-specific conditioner conditions the interdependencies between columns by integrating both discrete and continuous conditions. The introduction of conditional dependencies enables the generator to accurately capture complex dependencies between columns, thereby enhancing the fidelity of the synthetic data. The proposed framework, encompassing the interpolation in the decoding process and the generation method for conditions, serves to render synthetic data more realistic. A comprehensive evaluation on six datasets demonstrated that the proposed method is effective in terms of data quality, usability, and privacy level of the synthesized data. The source code is available at https://github.com/rch1025/Tabular-GAN .}, archive = {J_NEUCOM}, author = {Seungchan Roh and Seunghwan Song and Kwan-Yong Park and Byoung-mo Koo and Jun-Geol Baek}, doi = {10.1016/j.neucom.2025.130484}, journal = {Neurocomputing}, month = {9}, pages = {130484}, shortjournal = {Neurocomputing}, title = {Quality boost of tabular data synthesis using interpolative cumulative distribution function decoding and type-specific conditioner}, volume = {645}, year = {2025}, }
• (2025). Adaptive dynamic programming-based optimal pursuit–evasion control for quadrotor unmanned aerial vehicles with obstacle avoidance. NEUCOM, 645, 130483. (www)

  @article{ , abstract = {This article investigates the challenging problem of optimized intelligent pursuit–evasion control for quadrotor unmanned aerial vehicles with obstacle avoidance. Firstly, a novel penalty function is developed to achieve obstacle avoidance in the designed cost function. Subsequently, a critic-only structure is employed to substitute the conventional actor-critic structure for addressing the approximate solutions of Hamilton–Jacobi–Isaacs equations. Meanwhile, two new critic neural networks are constructed to learn the optimal cost functions and control policies for the pursuit–evasion control systems. More specifically, the developed weight update laws not only enable the critic neural network weights to be updated online, but also alleviate the persistent excitation condition with a simple structure. In particular, the obstacle avoidance function is incorporated into the construction of the approximate cost function learned by the proposed critic neural networks. In addition, the proposed pursuit–evasion control systems can be guaranteed to achieve uniformly ultimately bounded stability by utilizing Lyapunov methodology. Finally, the effectiveness of the developed pursuit–evasion control scheme is fully illustrated through two simulation cases.}, archive = {J_NEUCOM}, author = {Bo Li and Ziqi Yang and Hui Liu and Bing Xiao}, doi = {10.1016/j.neucom.2025.130483}, journal = {Neurocomputing}, month = {9}, pages = {130483}, shortjournal = {Neurocomputing}, title = {Adaptive dynamic programming-based optimal pursuit–evasion control for quadrotor unmanned aerial vehicles with obstacle avoidance}, volume = {645}, year = {2025}, }
• (2025). Talent3D: Optimizing geometry and enhancing appearance for high-quality text-to-3D shape generation. NEUCOM, 645, 130475. (www)

  @article{ , abstract = {Advancements in text-to-image generation have significantly propelled the development of text-guided 3D shape generation, enabling the creation of imaginative and detailed 3D assets for a wide range of real-world applications. However, existing methods still face significant challenges in generating high-quality models. Traditional Neural Radiance Field (NeRF)-based methods, constrained by their implicit representations that diverge from conventional polygon meshes, achieve superior rendering results but often lack detailed geometric structures. Moreover, the simultaneous assignment of color information during geometric structure generation can compromise texture consistency and realism. To address these limitations, we propose a two-stage model that separately handles geometry and appearance generation. In the first stage, an iterative surface optimization algorithm is designed to adaptively adjust vertex positions and surface densities, resulting in smooth and accurate geometric structures. In the second stage, we introduce Physically Based Rendering (PBR) with a spatially adaptive Bidirectional Reflectance Distribution Function (BRDF) to enhance the realism of surface materials. Comprehensive evaluations demonstrate that the results generated by the proposed method possess complete geometric topology and exhibit more realistic appearance textures, advancing those of the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Jun Wang and Shangzhe Wu and Ajian Liu and Ziyang Meng}, doi = {10.1016/j.neucom.2025.130475}, journal = {Neurocomputing}, month = {9}, pages = {130475}, shortjournal = {Neurocomputing}, title = {Talent3D: Optimizing geometry and enhancing appearance for high-quality text-to-3D shape generation}, volume = {645}, year = {2025}, }
• (2025). Synchronization control for delayed neural networks under round-robin protocol and input saturations. NEUCOM, 645, 130474. (www)

  @article{ , abstract = {This paper considers the synchronization control for chaotic delayed neural networks under both the Round-Robin protocol and input saturations in the master–slave framework. First, a periodic function is employed to represent the actual output errors, and the modified sector condition is utilized to handle input saturations. Then, based on the Lyapunov–Krasovskii theory, a sufficient condition is proposed under which the error system has the expected properties such as boundedness, H ∞ performance, and exponential stability. Furthermore, the linearized technique is proposed to characterize the synchronization criterion by means of linear matrix inequalities. For the case without time delay, the corresponding result is also provided. Moreover, the optimization problems related to system performance indices are discussed. Finally, two simulation examples show the validity of the derived results.}, archive = {J_NEUCOM}, author = {Yonggang Chen and Kelin Zhang and Yaxue Zhao}, doi = {10.1016/j.neucom.2025.130474}, journal = {Neurocomputing}, month = {9}, pages = {130474}, shortjournal = {Neurocomputing}, title = {Synchronization control for delayed neural networks under round-robin protocol and input saturations}, volume = {645}, year = {2025}, }
• (2025). Enhancing color selectivity in foundation models for downstream color vision tasks. NEUCOM, 645, 130471. (www)

  @article{ , abstract = {The emergence of computer vision foundation models, inspired by the success of task-agnostic pretrained representations in Natural Language Processing (NLP), is revolutionizing the field. These models produce features that excel in downstream tasks even without fine-tuning. Last year, DINOv2 emerged, surpassing previous state-of-the-art general-purpose features on computer vision benchmarks, both at the image and pixel levels. In this work, we focus on what type of color information is embedded in DINOv2 features, and to assess their performance in computer vision tasks where color is a critical cue—for instance, recognizing the color of vehicles for traffic monitoring, detecting skin tones in biometric applications, or assessing product color attributes in fashion and e-commerce. Furthermore, we also propose a training-free feature transformation that increases color selectivity in DINOv2 features, i.e. their ability to respond differently to various colors in an image, boosting the performance on several classes of the color vision tasks considered.}, archive = {J_NEUCOM}, author = {Simone Bianco}, doi = {10.1016/j.neucom.2025.130471}, journal = {Neurocomputing}, month = {9}, pages = {130471}, shortjournal = {Neurocomputing}, title = {Enhancing color selectivity in foundation models for downstream color vision tasks}, volume = {645}, year = {2025}, }
• (2025). A novel noise reduction and interaction enrichment recommendation model via contrastive learning. NEUCOM, 645, 130469. (www)

  @article{ , abstract = {To address the challenge of noisy data in recommendation systems, we propose a novel model that combines adaptive denoising with structural similarity-based interaction filling to improve recommendation accuracy. Our method dynamically prunes noisy interactions using a threshold-based denoising strategy without requiring auxiliary information. To compensate for information loss and preserve data sparsity, we introduce a structural similarity-based approach that fills pruned interactions with high-reliability user–item pairs, computed via cosine similarity. We further enhance representation learning via a contrastive learning framework that separates reliable and unreliable interactions based on mutual information maximization. Experimental results on three public datasets demonstrate that our model outperforms state-of-the-art baselines, achieving up to 4.5% and 6.7% improvements in Recall@5 and NDCG@5, respectively, confirming its effectiveness and practical applicability.}, archive = {J_NEUCOM}, author = {Zhiqiang Zhang and Qi Wu and Jiangzhou Deng and Jianmei Ye and Fang Yang and Yong Wang and Kobiljon Kh. Khushvakhtzoda}, doi = {10.1016/j.neucom.2025.130469}, journal = {Neurocomputing}, month = {9}, pages = {130469}, shortjournal = {Neurocomputing}, title = {A novel noise reduction and interaction enrichment recommendation model via contrastive learning}, volume = {645}, year = {2025}, }
• (2025). Dual representation modeling and progressive contrastive learning for unsupervised video person re-identification. NEUCOM, 645, 130467. (www)

  @article{ , abstract = {Unsupervised video person re-identification (USL-VReID) aims to identify the same person across cameras without annotation. The key to this challenging task is to take advantage of the discriminative features of frames and tracklets. Most existing methods model tracklet features by mean average pooling and address the USL-VReID problem by using cluster centroid contrastive learning. However, the mean average pooling strategy and cluster centers primarily focus on commonality, overlooking divergencies. In this paper, we propose a Dual Representation Modeling and Progressive Contrastive Learning (DRMPCL) method to learn fine-grained features of the diversity distribution. Specifically, our approach explores the discriminant features of diversity distribution in both frame-level tracklet modeling and tracklet-level discrepant contrastive learning. Then a multi-level progressive learning strategy effectively optimizes the model from reliable to informative, and from commonality to divergence. Comprehensive experiments on two widely-used datasets demonstrate that DRMPCL outperforms the state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Cong Zhang and Yanzhao Su and Nian Wang and Yunwei Lan and Tao Wang and Aihua Li}, doi = {10.1016/j.neucom.2025.130467}, journal = {Neurocomputing}, month = {9}, pages = {130467}, shortjournal = {Neurocomputing}, title = {Dual representation modeling and progressive contrastive learning for unsupervised video person re-identification}, volume = {645}, year = {2025}, }
• (2025). PDFL: Progressive discriminative feature learning for long-tailed recognition. NEUCOM, 645, 130463. (www)

  @article{ , abstract = {The class imbalance (or long-tailed) problem emerges in many visual tasks such as image classification, where data is scarce and biased, resulting in a performance bottleneck. Previous studies on this problem usually require the label distribution in advance, while the distribution of labels is unknown in most practical scenarios. To resolve the reliance on label distribution, we propose a Progressive Discriminative Feature Learning (PDFL) framework to enhance the discriminability of features for long-tailed recognition. Firstly, we observe that the imbalance of a dataset is hidden in the learned parameters of a model, thereby an adaptive imbalance indicator is designed to avoid the requirement of the label distribution. Based on that, the unified model-aware margin and re-weighting functions are proposed to enlarge inter-class discrepancy. Secondly, for greater intra-class compactness, we put forward momentum variational prototype learning to introduce sample-to-sample comparisons and improve the quality of injected features in a memory bank with a selection and momentum update scheme. Thirdly, in the training process, a progressive learning strategy is devised for the gradual optimization of tail classes. Extensive experiments on three long-tailed benchmarks demonstrate the superiority of our proposed framework, surpassing an average of 5.5% accuracy towards the prevalent LDAM over three public benchmarks.}, archive = {J_NEUCOM}, author = {Weiqiu Wang and Zining Chen and Zhicheng Zhao and Fei Su}, doi = {10.1016/j.neucom.2025.130463}, journal = {Neurocomputing}, month = {9}, pages = {130463}, shortjournal = {Neurocomputing}, title = {PDFL: Progressive discriminative feature learning for long-tailed recognition}, volume = {645}, year = {2025}, }
• (2025). Optimal privacy-preserving economic dispatch of smart grids via lightweight state decomposition and designed noise. NEUCOM, 645, 130460. (www)

  @article{ , abstract = {This paper investigates privacy protection in the distributed economic dispatch problem (EDP) within smart grids. The algorithm employs a lightweight state decomposition method to separate certain variables, thereby safeguarding the initial state information. Subsequently, by incorporating random zero-sum exponential attenuation noise, the algorithm prevents attackers from eavesdropping and inferring critical generator power and cost function parameters. The proposed algorithm assumes that the eavesdropper possesses sufficient system information; this stringent assumption renders the design of the privacy algorithm more rigorous. The algorithm effectively safeguards the generator’s privacy information while ensuring the optimal solution for economic dispatch (ED). We analyze the privacy of the algorithm by ϵ , δ -data privacy. Additionally, the algorithm is applicable to B -strongly connected directed topologies, and it incorporates multiple constraints and application scenarios to enhance its versatility. Notably, the power ramp constraint is implemented by controlling the consensus variable, resulting in a faster convergence speed and a tighter power fluctuation range compared to traditional methods. Finally, the effectiveness of the algorithm is validated through several simulations.}, archive = {J_NEUCOM}, author = {Lianghao Ji and Zhiqiang Ren and Yan Xie and Shasha Yang and Xing Guo}, doi = {10.1016/j.neucom.2025.130460}, journal = {Neurocomputing}, month = {9}, pages = {130460}, shortjournal = {Neurocomputing}, title = {Optimal privacy-preserving economic dispatch of smart grids via lightweight state decomposition and designed noise}, volume = {645}, year = {2025}, }
• (2025). GIPTP: Global interaction policy trend prediction based on historical behavior for multi-agent cooperative reinforcement learning. NEUCOM, 645, 130459. (www)

  @article{ , abstract = {Policy prediction and dynamic adaptation are crucial research areas for cooperative multi-agent reinforcement learning (MARL). Recent MARL methods have predominantly relied on local observations and historical behaviors for decision-making, which often leads to short-sighted policies. However, historical behaviors not only reflect past decisions and experiences but also significantly influence future global decision-making. Existing methods ignore the potential impact of agents’ historical behaviors on current and future decision trends. Therefore, we propose the Historical Behavior-Driven Global Interaction Policy Trend Prediction (GIPTP) algorithm, which captures agents’ interaction behaviors and historical experiences to predict future decision trends. The algorithm dynamically selects interaction policies by considering the agent’s goals, environmental states, and enemy actions. First, we decompose the complete graph of global entities into several non-complete policy subgraphs, using a sparse activation function to extract these subgraphs and represent real-time interaction policies among entities. Next, we introduce an interaction graph selector, where agents generate influence weights for the policy subgraphs based on historical behavioral trajectories, and integrate the local features of neighboring agents using the attention mechanism. Finally, agents aggregate the policy subgraphs into a compact interaction policy prediction graph, enabling coordinated prediction of future global interaction policies at the entity level. Experiments conducted on the StarCraft Multi-Agent Challenge (SMAC) benchmark show that the GIPTP algorithm outperforms classical MARL algorithms, offering superior generalization capabilities in complex tasks.}, archive = {J_NEUCOM}, author = {Shudong Wang and Wenhao Ji and Haiyuan Gui and Kuijie Zhang and Luqi Wang and Shanchen Pang}, doi = {10.1016/j.neucom.2025.130459}, journal = {Neurocomputing}, month = {9}, pages = {130459}, shortjournal = {Neurocomputing}, title = {GIPTP: Global interaction policy trend prediction based on historical behavior for multi-agent cooperative reinforcement learning}, volume = {645}, year = {2025}, }
• (2025). A generalized classification of human sperm head morphology via contrastive meta-learning with auxiliary tasks. NEUCOM, 645, 130454. (www)

  @article{ , abstract = {Semen analysis is the primary method for evaluating male infertility, with sperm morphology being a critical indicator. However, existing methods for classifying human sperm head morphology (HSHM) often lack cross-domain generalizability. To address this limitation, we propose an enhanced meta-learning algorithm that learns invariant features across tasks, improving generalization by transferring knowledge to new tasks. To mitigate gradient conflicts in multi-task learning, we separate meta-training tasks into primary and auxiliary tasks. This approach, in conjunction with auxiliary tasks, enhances the model’s generalization using diverse HSHM datasets. We introduce the Contrastive Meta-Learning with Auxiliary Tasks (HSHM-CMA) algorithm, which integrates localized contrastive learning in the outer loop of meta-learning to exploit invariant sperm morphology features across domains. This improves task convergence and adaptation to new categories. In our evaluation, we assess the model’s generalization performance using three testing objectives: the same dataset with different HSHM categories, different datasets with the same HSHM categories, and different datasets with different HSHM categories. Our evaluation across these objectives demonstrates that HSHM-CMA outperforms existing meta-learning approaches, achieving accuracies of 65.83%, 81.42%, and 60.13%, respectively.}, archive = {J_NEUCOM}, author = {Yuh-Shyan Chen and Yu-Chi Chang and Jia-Yi Hong}, doi = {10.1016/j.neucom.2025.130454}, journal = {Neurocomputing}, month = {9}, pages = {130454}, shortjournal = {Neurocomputing}, title = {A generalized classification of human sperm head morphology via contrastive meta-learning with auxiliary tasks}, volume = {645}, year = {2025}, }
• (2025). Release probability distribution modulates attractor state stability and mobility in continuous-attractor neural networks. NEUCOM, 645, 130450. (www)

  @article{ , abstract = {We investigate how the distribution of synaptic release probabilities influences neural network dynamics using a continuous-attractor neural network (CANN) model with short-term synaptic depression. Motivated by experimental findings showing that astrocytic NMDA receptor blockade narrows neurotransmitter release probability distributions, we incorporate gamma-distributed release probabilities into our model. Our analysis reveals that the variation of release probabilities modulates attractor state stability within the network. Specifically, release probability variation narrowing increases the threshold for translational instability and reduces the network’s diffusion constant under noise perturbations, while the variation widening affects the network oppositely. These findings suggest a novel mechanism by which astrocytic NMDA receptors may modulate neural circuit dynamics and information processing through the regulation of synaptic release probability distributions.}, archive = {J_NEUCOM}, author = {Zihan Liu and Flavia Nathaline Chanentia and Patteera Supvithayanond and Chi Chung Alan Fung}, doi = {10.1016/j.neucom.2025.130450}, journal = {Neurocomputing}, month = {9}, pages = {130450}, shortjournal = {Neurocomputing}, title = {Release probability distribution modulates attractor state stability and mobility in continuous-attractor neural networks}, volume = {645}, year = {2025}, }
• (2025). Two-factor smoothed incomplete multi-view clustering via inconsistent guidance. NEUCOM, 645, 130438. (www)

  @article{ , abstract = {Multi-source data is often incomplete due to collection environment, technical limitations, or privacy protection, Incomplete Multi-View Clustering (IMC) has been proposed to minimize the impact of missing data on clustering. However, most existing methods only smooth the consensus representation from a single viewpoint during the clustering process and finally obtain the final result by k -means or other clustering methods. This makes it challenging to ensure the accuracy and stability of the clustering results. To solve this problem, this paper proposes a two-factor smoothed incomplete multi-view clustering via inconsistent guidance method (TSIIG) that is learned in two stages. In the first stage, the affinity graphs of ideal structures are mainly learned. In the second stage, cross-view learning, two-factor smoothing learning, and adaptive weighted multi-view learning are integrated into an optimization framework. Different from other IMC methods, TSIIG introduces an inconsistency cost function to drive the low-dimensional embedding matrices of individual views to converge to a common consensus. In addition, a two-factor smoothing correction is applied to the consensus graph from the data’s feature attributes and manifold structure to obtain a more accurate consensus graph. Experimental results on eleven incomplete multi-view datasets show that our method outperforms existing clustering methods in terms of performance to validate the method’s effectiveness.}, archive = {J_NEUCOM}, author = {Keyu Ma and Youlong Yang and Zeping Ge}, doi = {10.1016/j.neucom.2025.130438}, journal = {Neurocomputing}, month = {9}, pages = {130438}, shortjournal = {Neurocomputing}, title = {Two-factor smoothed incomplete multi-view clustering via inconsistent guidance}, volume = {645}, year = {2025}, }
• (2025). Automatic ensemble feature selection for multi-view data. NEUCOM, 645, 130437. (www)

  @article{ , abstract = {In recent years, multi-view data has been extensively utilized across various fields of machine learning. Due to its high dimensionality, numerous methods have been proposed for feature selection. However, existing approaches often focus on the relationships between views while overlooking the correlation between selected features and final learning outcomes. Additionally, these methods struggle to identify an optimal feature subset that improves learning performance. This paper presents an automatic feature selection ensemble framework for multi-view data to address these challenges. Our framework integrates feature selection with final learning results, ensuring the selected features are highly relevant to the learning task. Initially, we use the minimum Redundancy Maximum Relevance (mRMR) criterion to identify relevant features within each view, forming an initial feature set. We then evaluate this set’s generalization performance using classifier accuracy as a baseline. Next, we introduce an Automatic Optimal Feature Subset Iterative Selection method, which iteratively refines the feature subset by considering the differences and relationships between views, ensuring minimal redundancy and maximal contribution to classification performance. By comparing our framework with other state-of-the-art algorithms on nine widely used image and text multi-view datasets, we demonstrate that our framework selects features with superior learning performance, reducing the number of selected features by 13% and achieving an average classification accuracy improvement of 11% compared to other methods. Additionally, our framework also exhibits advantages in terms of training speed compared to other algorithms. Code for this paper available on: https://github.com/boredcui/AFIS .}, archive = {J_NEUCOM}, author = {Xiaojian Ding and Menghan Cui and Kaixiang Wang}, doi = {10.1016/j.neucom.2025.130437}, journal = {Neurocomputing}, month = {9}, pages = {130437}, shortjournal = {Neurocomputing}, title = {Automatic ensemble feature selection for multi-view data}, volume = {645}, year = {2025}, }
• (2025). Human-to-robot handovers based on multimodal perception. NEUCOM, 645, 130435. (www)

  @article{ , abstract = {Human–robot collaboration has always been an active research field in robotics. Understanding human intentions accurately and achieving safe and reliable handovers pose a challenge in the field of human–robot collaboration. In this paper, we present a multi-modal fusion framework for human–robot handover, consisting of three sub-tasks: human intention analysis, grasp configurations generation, and interaction trajectory generation. In the first sub-task, the intent of the giver is analyzed by fusing speech recognition, syntactic parsing,and visual recognition, resulting in the generation of operation vector. For the second sub-task, we propose a method on point cloud for generating 6-Degree-of-Freedom (6-DoF) grasp planning in the human–robot handover task. For the third sub-task, an improved Dynamic Movement Primitive (DMP) is proposed to generate human-like interaction trajectories. Based on the proposed framework, we construct a human–robot collaboration system capable of safely handing over and placing different objects upon receiving commands. Experimental results confirm the effectiveness of our proposed framework.}, archive = {J_NEUCOM}, author = {Chunfang Liu and Weifan Wang and Chenxin Li and Ruitian Pang and Yan Shang and Yitong Gao}, doi = {10.1016/j.neucom.2025.130435}, journal = {Neurocomputing}, month = {9}, pages = {130435}, shortjournal = {Neurocomputing}, title = {Human-to-robot handovers based on multimodal perception}, volume = {645}, year = {2025}, }
• (2025). More observation leads to more clarity: Multi-view collaboration network for camouflaged object detection. NEUCOM, 645, 130433. (www)

  @article{ , abstract = {Currently, Camouflaged Object Detection (COD) methods often rely on single-view feature perception, which struggles to fully capture camouflaged objects due to environmental interference such as background clutter, lighting variations, and viewpoint changes. To address this, we propose the multi-view collaboration network (MCNet), inspired by human visual strategies for complex scene analysis. MCNet incorporates multiple perspectives for enhanced feature extraction. The global perception module takes the original, far, and near views, using different large-kernel convolutions and multi-head attention mechanisms for global feature embedding. In parallel, the local perception module processes the tilted, projected, and color-jittered views, extracting fine-grained local features through multi-branch deep convolutions and dilated convolutions. To facilitate deep interaction between global and local features, we introduce the hybrid interactive module, which explores the correlation of multi-view feature information and adaptively fuses features. For feature decoding, the dynamic pyramid shrinkage module integrates dynamic gated convolutions with a pyramid shrinkage mechanism, progressively aggregating semantic features through a hierarchical shrinking strategy and group fusion strategy. Experimental results on popular COD benchmark datasets show that MCNet outperforms 18 state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Fangyan Wang and Ge Jiao and Guowen Yue}, doi = {10.1016/j.neucom.2025.130433}, journal = {Neurocomputing}, month = {9}, pages = {130433}, shortjournal = {Neurocomputing}, title = {More observation leads to more clarity: Multi-view collaboration network for camouflaged object detection}, volume = {645}, year = {2025}, }
• (2025). Boosting few-shot action recognition via time-enhanced multimodal adaptation learning. NEUCOM, 645, 130427. (www)

  @article{ , abstract = {Few-shot action recognition (FSAR) aims to classify new classes using a limited number of labeled samples, which has attracted considerable attention due to difficulties in collecting and annotating large-scale video data. Nevertheless, the majority of existing methods grapple with significant challenges. They often struggle with a lack of multimodal cues, insufficient temporal modeling and ineffective feature integration. In this paper, we innovatively propose the Time-enhanced Multimodal Adaptation Learning Network (TMAL-Net), a straightforward yet highly effective multimodal learning framework for FSAR. TMAL-Net uniquely incorporates both spatial and temporal information and leverages semantic knowledge to improve action understanding. Specifically, following the parameter-efficient transfer learning setting, we incorporate lightweight adapters into pretrained image models such as CLIP to implement dynamic modeling in videos. Moreover, hierarchical vision-text and vision-flow fusions are developed to effectively combine multimodal cues for comprehensive representation learning. Notably, the incorporation of optical flow significantly enhances the temporal modeling crucial for action understanding. Extensive experiments on multiple challenging benchmark datasets demonstrate the effectiveness of the proposed method. Notably, our method outperforms the competitor CLIP-FSAR by a large margin of 2.5% and 2.8% on the Kinetics and SSV2-small datasets under the most challenging 5-way 1-shot setting, while our method has 28.57M tunable parameters on the SSV2-small dataset, only one-third of those in the full fine-tuning setting (90.99M).}, archive = {J_NEUCOM}, author = {Dan Liu and Qing Xia and Ai Peng and Zhouli Shen and Feng Zhang and Mao Ye and Jianwei Zhang}, doi = {10.1016/j.neucom.2025.130427}, journal = {Neurocomputing}, month = {9}, pages = {130427}, shortjournal = {Neurocomputing}, title = {Boosting few-shot action recognition via time-enhanced multimodal adaptation learning}, volume = {645}, year = {2025}, }
• (2025). Model-agnostic personalized adaptation for segment anything model. NEUCOM, 645, 130424. (www)

  @article{ , abstract = {The Segment Anything Model (SAM) and its family of models have made significant strides in open-set, prompt-driven instance segmentation. However, some closed-source SAM family models often face ethical, copyright, or commercial restrictions, limiting their accessibility and further personalized adaptation. To overcome these limitations, we introduce MapSAM, a model-agnostic, personalized plugin for SAM family models. MapSAM features a lightweight threshold learner that enables nuanced post-hoc processing of confidence maps, leading to improved segmentation accuracy. By leveraging mask-focused learning, our approach determines pixel-wise and hardness-aware thresholds, allowing for more effective adaptation to diverse datasets. Furthermore, we critically examine the limitations of the commonly used Dice loss, which can overlook sample hardness when allocating penalties. We theoretically demonstrate that the Mean Squared Error (MSE) loss complements Dice loss by providing a stronger focus on sample hardness. Through extensive experiments on seven diverse datasets using multiple SAM family models, we validate the effectiveness of MapSAM in achieving superior segmentation results, particularly in challenging domains. Our findings open up new avenues for personalized, open-set instance segmentation across various application areas, leveraging any closed-source SAM family model. Code will be available at https://github.com/wjc2830/MapSAM.git .}, archive = {J_NEUCOM}, author = {Juncheng Wang and Lei Shang and Wang Lu and Xiangyang Ji and Shujun Wang}, doi = {10.1016/j.neucom.2025.130424}, journal = {Neurocomputing}, month = {9}, pages = {130424}, shortjournal = {Neurocomputing}, title = {Model-agnostic personalized adaptation for segment anything model}, volume = {645}, year = {2025}, }
• (2025). A framework for neural topic modeling with mutual information and group regularization. NEUCOM, 645, 130420. (www)

  @article{ , abstract = {Recent advances in topic modeling have leveraged pre-trained language models (PLMs) and refined topic-word associations to improve both topic discovery and document topic distributions. However, directly integrating PLMs often leads to higher inference costs, making them less suitable for low-latency applications. At the same time, effectively capturing inter-topic relationships remains a critical yet challenging task. In this paper, we propose NeuroMIG ( Neu ral T o pic Modeling with M utual I nformation and G roup Topic Regularization), a novel framework that addresses both issues. NeuroMIG (1) maximizes mutual information between document topic distributions and PLM embeddings to efficiently incorporate PLM knowledge, and (2) employs Group Topic Regularization based on optimal transport to model interactions among topics. Compatible with a wide range of topic modeling architectures, NeuroMIG significantly improves performance over baselines while preserving efficient inference, as validated by experimental results.}, archive = {J_NEUCOM}, author = {Tung Nguyen and Linh Ngo Van and Anh Nguyen Duc and Sang Dinh Viet}, doi = {10.1016/j.neucom.2025.130420}, journal = {Neurocomputing}, month = {9}, pages = {130420}, shortjournal = {Neurocomputing}, title = {A framework for neural topic modeling with mutual information and group regularization}, volume = {645}, year = {2025}, }
• (2025). SqSFill: Joint spatial and spectral learning for high-fidelity image inpainting. NEUCOM, 645, 130414. (www)

  @article{ , abstract = {Image inpainting has made significant progress due to recent advances in deep learning. However, most generative inpainting networks face challenges such as producing blurry results that lack high-frequency details or introducing inconsistent structures. To address these issues, we propose a novel transformer-based approach, SqSFill, which exploits rich information in both spatial and spectral domains. Specifically, SqSFill incorporates Rectified Frequency Feature Extractor (RecFFE) in the early layers of the network to capture fine-grained details by leveraging frequency information, guided by frequency loss. Moreover, we design a Scout Attention Block with linear complexity to replace vanilla self-attention, thereby effectively capturing long-range dependencies with lower computational cost. By integrating the RecFFE and Scout Attention Block, SqSFill is able to generate both coherent structures and sharp textures. Extensive experiments demonstrate the proposed SqSFill achieves superior results, outperforming previous state-of-the-art approaches with fewer parameters.}, archive = {J_NEUCOM}, author = {Zihao Zhang and Feifan Cai and Qin Zhou and Youdong Ding}, doi = {10.1016/j.neucom.2025.130414}, journal = {Neurocomputing}, month = {9}, pages = {130414}, shortjournal = {Neurocomputing}, title = {SqSFill: Joint spatial and spectral learning for high-fidelity image inpainting}, volume = {645}, year = {2025}, }
• (2025). Intra-class patch swap for self-distillation. NEUCOM, 645, 130408. (www)

  @article{ , abstract = {Knowledge distillation (KD) is a valuable technique for compressing large deep learning models into smaller, edge-suitable networks. However, conventional KD frameworks rely on pre-trained high-capacity teacher networks, which introduce significant challenges such as increased memory/storage requirements, additional training costs, and ambiguity in selecting an appropriate teacher for a given student model. Although a teacher-free distillation (self-distillation) has emerged as a promising alternative, many existing approaches still rely on architectural modifications or complex training procedures, which limit their generality and efficiency. To address these limitations, we propose a novel framework based on teacher-free distillation that operates using a single student network without any auxiliary components, architectural modifications, or additional learnable parameters. Our approach is built on a simple yet highly effective augmentation, called intra-class patch swap augmentation. This augmentation simulates a teacher–student dynamic within a single model by generating pairs of intra-class samples with varying confidence levels, and then applying instance-to-instance distillation to align their predictive distributions. Our method is conceptually simple, model-agnostic, and easy to implement, requiring only a single augmentation function. Extensive experiments across image classification, semantic segmentation, and object detection show that our method consistently outperforms both existing self-distillation baselines and conventional teacher-based KD approaches. These results suggest that the success of self-distillation could hinge on the design of the augmentation itself. Our codes are available at https://github.com/hchoi71/Intra-class-Patch-Swap.}, archive = {J_NEUCOM}, author = {Hongjun Choi and Eun Som Jeon and Ankita Shukla and Pavan Turaga}, doi = {10.1016/j.neucom.2025.130408}, journal = {Neurocomputing}, month = {9}, pages = {130408}, shortjournal = {Neurocomputing}, title = {Intra-class patch swap for self-distillation}, volume = {645}, year = {2025}, }
• (2025). Spiking trans-YOLO: A range-adaptive energy-efficient bridge between YOLO and transformer. NEUCOM, 645, 130407. (www)

  @article{ , abstract = {The remarkable success of Transformers in Artificial Neural Networks (ANNs) has driven growing interest in leveraging self-attention mechanisms and Transformer-based architectures for Spiking Neural Network (SNN) object detection. However, existing methods combining Transformers with YOLO lack reasonable bridging strategies, leading to significantly increased computational costs and limitations in local feature extraction. To address these challenges, we propose Spiking Trans-YOLO, introducing a Top-Attention Hybrid Feature Fusion module, which exclusively applies self-attention to high-level spiking features that are more stable and semantically meaningful. This prevents redundant computations caused by unstable low-level spikes and reduces energy consumption. Subsequently, we perform cross-scale feature fusion to compensate for Transformers’ shortcomings in local feature extraction. This approach efficiently bridges YOLO and Transformer architectures while preserving the low-power characteristics of SNNs. Additionally, The newly proposed Integer Leaky Integrate-and-Fire (I-LIF) neuron has demonstrated significant potential in SNNs by enabling integer-valued training and spike-driven inference, thereby reducing quantization errors. However, existing spiking self-attention mechanisms fail to incorporate proper scaling factors for I-LIF neurons, which may lead to gradient vanishing. To address this, we propose a Range-Adaptive Spiking Attention for intra-scale interactions. By dynamically adjusting scaling coefficients, RASA mitigates gradient vanishing issues associated with integer training, allowing I-LIF neurons to exploit the benefits of spike-based self-attention fully. The proposed method achieves 67.7% mAP@50 on the COCO dataset and 68.6% mAP@50 on the Gen1 dataset, outperforming state-of-the-art YOLO architectures and achieving superior energy efficiency compared to advanced Transformer-based architectures. Code: https://github.com/s1110/Spiking-Trans-YOLO .}, archive = {J_NEUCOM}, author = {Yushi Huo and Hongwei Ge and Guozhi Tang and Shengxuan Gao and Jiale Xu}, doi = {10.1016/j.neucom.2025.130407}, journal = {Neurocomputing}, month = {9}, pages = {130407}, shortjournal = {Neurocomputing}, title = {Spiking trans-YOLO: A range-adaptive energy-efficient bridge between YOLO and transformer}, volume = {645}, year = {2025}, }
• (2025). A contrastive learning-based heterogeneous dual-branch network for source camera identification. NEUCOM, 645, 130406. (www)

  @article{ , abstract = {Source camera identification has been a significant focus in image forensics over the past decades. However, as camera model and instance related forensic features are weak compared to image contents, identification performance is far from satisfactory for practical applications. This paper introduces a novel contrastive learning strategy, aimed at enhancing the learning of camera fingerprints by leveraging the similarity between the two branches in a heterogeneous dual-branch network. Initially, a heterogeneous dual-branch feature extraction module is designed, employing two distinct strategies: noise residual estimation and progressive direct estimation, to independently extract forensic information. Contrastive learning is then utilized to enhance shared forensic features related to camera models between the two branches while filtering out irrelevant content residuals. During training, in addition to supervised classification loss, both spatial and frequency losses are applied to ensure the features consistency between the two branches, thereby enhancing the similarity of the features learned by both branches in the spatial and frequency domains. Drawing inspiration from the peak correlation energy metric commonly used in traditional methods, a frequency domain correlation loss is proposed. Extensive experimental results on the Dresden and Vision datasets demonstrate that the proposed method outperforms state-of-the-art approaches. Furthermore, it shows improved robustness against common preprocessing attacks such as JPEG recompression and image resizing, making it more suitable for real-world applications.}, archive = {J_NEUCOM}, author = {Zijuan Han and Yang Yang and Jinkai Zhang and Yang Li and Yunxia Liu and Ngai-Fong Bonnie Law}, doi = {10.1016/j.neucom.2025.130406}, journal = {Neurocomputing}, month = {9}, pages = {130406}, shortjournal = {Neurocomputing}, title = {A contrastive learning-based heterogeneous dual-branch network for source camera identification}, volume = {645}, year = {2025}, }
• (2025). Efficient CUR decomposition for interpretable low-rank approximations and imaging applications. NEUCOM, 645, 130401. (www)

  @article{ , abstract = {Low-rank approximations based on the selected columns and rows from a given matrix are an alternative approach to singular value decomposition (SVD) and offer more interpretable outputs. They have been successfully used in pattern recognition, recommendation systems, bioinformatics, and social network analysis. One popular approach is CUR matrix decomposition, which represents the data matrix using three small matrices and can be achieved through randomized or deterministic approaches. Deterministic algorithms yield higher approximation accuracy than randomized ones. However, they are highly computationally expensive, especially with large datasets. To address these limitations, we introduce a new method for efficiently computing the CUR matrix decomposition called HyCUR. It incorporates a deterministic sampling strategy based on guided information of the singular vectors and a randomized sampling to select rows and columns for efficient CUR low-rank approximation. HyCUR overcomes the main drawbacks of deterministic methods and can estimate the approximation rank using an incremental QR factorization approach, which also approximates the SVD of the data matrix instead of the classical method. Therefore, HyCUR is suitable for applications with unknown data rank. Experimental results on synthetic and real datasets of varying sizes demonstrate that HyCUR achieves comparable accuracy to recent deterministic approaches while significantly improving computational efficiency for large low-rank datasets. We validate the proposed method in two real imaging applications: simultaneous sample and feature extraction from large datasets and hyperspectral unmixing to extract endmember signatures and estimate their fractions from hyperspectral image datasets (HSIs). The experiment results show promising performance in both applications.}, archive = {J_NEUCOM}, author = {Muhammad A.A. Abdelgawad and Ray C.C. Cheung and Hong Yan}, doi = {10.1016/j.neucom.2025.130401}, journal = {Neurocomputing}, month = {9}, pages = {130401}, shortjournal = {Neurocomputing}, title = {Efficient CUR decomposition for interpretable low-rank approximations and imaging applications}, volume = {645}, year = {2025}, }
• (2025). QueryAccelerator: A SQL converter capturing and transforming the slow plan. NEUCOM, 645, 130395. (www)

  @article{ , abstract = {For slow queries, rewriting is a prerequisite for effective query optimization in database systems. To convert a slow query into a more efficient SQL, a suitable rewrite rule application sequence is required. However, existing rewriting methods have the problems of insufficient representation of query features, too large exploration space and lack of early avoidance of error queries, which limits rewriting performance, benefits and accuracy, therefore affects the efficiency of query execution. In this paper, we propose QueryAccelerator, a slow query rewriting framework that explores suitable rule application sequences based on execution plans. We use the deep reinforcement learning to capture the optimizable space within slow query plans, and use rule categories as the exploration granularity to reduce the search space of rule application paths and avoid falling into local optimality. Additionally, to improve the accuracy of rewritten queries, we propose to reduce the rewriting benefits along paths with error nodes during the model exploration process, so as to avoid semantically inequivalent rewritten queries. Experimental results on slow queries written by real developers on LeetCode show that, compared with state-of-the-art query rewriting techniques, QueryAccelerator can not only optimize more slow queries, but also bring more performance improvements.}, archive = {J_NEUCOM}, author = {Shaocong Yang and Ning Wang and Qixiong Zeng and Anshun Zhou}, doi = {10.1016/j.neucom.2025.130395}, journal = {Neurocomputing}, month = {9}, pages = {130395}, shortjournal = {Neurocomputing}, title = {QueryAccelerator: A SQL converter capturing and transforming the slow plan}, volume = {645}, year = {2025}, }
• (2025). SPLAC: A single-step PTZ camera linear auto-calibration method. NEUCOM, 645, 130377. (www)

  @article{ , abstract = {Pan-tilt-zoom (PTZ) cameras require auto-calibration techniques to estimate intrinsic parameters. Existing methods often involve multi-step processes or higher-order nonlinear constraints, which increase computational complexity and potentially lead to error accumulation. To address these challenges, we propose SPLAC, a single-step PTZ camera linear auto-calibration method based on an imaging model. SPLAC introduces a novel linear constraint derived from geometric relationships of line pairs in the perspective projection process. By solving a single linear set of equations, the method estimates camera parameters directly from images and pan–tilt feedback, thereby simplifying the calibration process. Additionally, to mitigate the impact of feature point matching errors, which can reduce calibration accuracy or cause failure, we introduce a filtering method based on the rotation characteristics of PTZ cameras to eliminate mismatched feature points. Experiments conducted on simulated data, warped panoramic images, and real PTZ camera images demonstrate the feasibility, accuracy, and robustness of our proposed method.}, archive = {J_NEUCOM}, author = {Feng Yang and Hua Chen and Xiaogang Zhang and Shi Wang and Lei Zhang}, doi = {10.1016/j.neucom.2025.130377}, journal = {Neurocomputing}, month = {9}, pages = {130377}, shortjournal = {Neurocomputing}, title = {SPLAC: A single-step PTZ camera linear auto-calibration method}, volume = {645}, year = {2025}, }
• (2025). Further analysis of weighted integral inequalities for improved exponential stability analysis of time delay neural networks systems. NEUCOM, 645, 130370. (www)

  @article{ , abstract = {This work investigates the exponential stability of neural networks systems with time delay. By incorporating orthogonal polynomials with weighted terms, a new weighted integral inequality is presented that extends several recently established results. Furthermore, the study focuses on analyzing the exponential stability of systems with time-varying delays, which include an exponential decay factor. To achieve this, a weighted version of the reciprocally convex inequality is first derived. Utilizing these inequalities and suitable Lyapunov–Krasovskii functionals within the framework of linear matrix inequalities, a new criterion for the exponential stability of neural networks systems is established. The effectiveness of the proposed method is demonstrated through multiple numerical examples.}, archive = {J_NEUCOM}, author = {Yuanyuan Zhang and Han Xue and Kachong Lao and Chonkit Chan and Chenyang Shi and Seakweng Vong}, doi = {10.1016/j.neucom.2025.130370}, journal = {Neurocomputing}, month = {9}, pages = {130370}, shortjournal = {Neurocomputing}, title = {Further analysis of weighted integral inequalities for improved exponential stability analysis of time delay neural networks systems}, volume = {645}, year = {2025}, }
• (2025). Identification of coupling matrices for neutral multi-weighted stochastic coupled systems with markovian switching. NEUCOM, 645, 130298. (www)

  @article{ , abstract = {This paper identifies the coupling matrices of the neutral stochastic multi-weighted coupled systems with Markovian switching (SMCM). To enhance adaptability of the model, coupled delay (the point-wise delay and the distributed delay) and neutral delay are considered in this paper, which may introduce deviations between the actual data and the observed data during the identification process of coupling matrices. In certain instances, these delays can lead to identification failure. To address this issue, it is necessary to introduce adaptive feedback control (AFC), which automatically adjusts controller parameters to compensate for the effects of delays, thereby mitigating changes in model parameters and input signals. By using graph-theoretic method, the challenge associated with constructing a global Lyapunov function is addressed, and the identification criteria are obtained based on adaptive synchronization method. Numerical simulations are provided to verify the feasibility of the theoretical results.}, archive = {J_NEUCOM}, author = {Tiantian Ge and Chunmei Zhang and Jiamin Zhou}, doi = {10.1016/j.neucom.2025.130298}, journal = {Neurocomputing}, month = {9}, pages = {130298}, shortjournal = {Neurocomputing}, title = {Identification of coupling matrices for neutral multi-weighted stochastic coupled systems with markovian switching}, volume = {645}, year = {2025}, }
• (2025). Foreground natural anomaly synthesis for attention guided anomaly detection. NEUCOM, 645, 130282. (www)

  @article{ , abstract = {In industrial scenarios, defect detection is an important application, and most of the traditional defect detection methods use supervised learning, which requires us to have a large number of labeled datasets for model training. However, it is often difficult to obtain a large number of negative samples containing defects in industrial scenarios, which brings a challenge for defect detection. How to achieve the detection and positioning of defects without supervision is an urgent problem for the industry. In this paper, we propose a cluster-based synthesis method for foreground natural anomalies to synthesize pseudo-anomalies similar to those in real industrial scenarios, and construct an attention-guided self-supervised proxy task to solve the problem of detection and localization of engine surface defect anomalies. Our proposed method is named as FNAtAD. We recommend seamlessly inserting natural patches onto the engine surface rather than the background to build synthetic anomalies that are closer to the real ones. We use the clustering algorithm to divide the image according to the foreground and background, obtain the foreground mask of the image, and control the abnormal synthetic position according to the foreground mask. Compared with the self-supervised learning task constructed by traditional anomaly synthesis methods, our anomaly construction is closer to the distribution of anomalies in the dataset. Meanwhile, our proxy tasks with attention mechanism can also better learn the abnormal patterns of pictures, and can evaluate abnormalities more effectively. Experimental results show that our method achieves better results than existing anomaly synthesis methods like DRÆM and PII, on our own engine surface defect dataset FNAtAD getting 93.4% AUROC and on MVTec FNAtAD getting 97.6% AUROC.}, archive = {J_NEUCOM}, author = {Xinyuan Xiang and Meiqin Liu and Senlin Zhang and Ping Wei and Badong Chen}, doi = {10.1016/j.neucom.2025.130282}, journal = {Neurocomputing}, month = {9}, pages = {130282}, shortjournal = {Neurocomputing}, title = {Foreground natural anomaly synthesis for attention guided anomaly detection}, volume = {645}, year = {2025}, }
• (2025). Theoretical bases for the relation between excitability, variability and synchronization in sequential neural dynamics. NEUCOM, 645, 130218. (www)

  @article{ , abstract = {Functional neural rhythms arise from the excitability of individual neurons and the collective dynamics of the circuits in which they are embedded. The self-organization of neural networks can be best investigated in circuits where these two aspects can be elucidated. In this context, Central Pattern Generators (CPGs) are an ideal model system where the interplay between synaptic and ionic currents shapes robust yet flexible rhythmic sequences underlying a wide range of motor functions. CPGs often integrate both chemical and electrical synapses, forming closed-loop network topologies that autonomously generate and coordinate functional rhythms. Recent findings in CPGs show the existence of sequential dynamical invariants, which are robust relationships between specific time intervals that build the rhythmic sequence constraining their cycle-by-cycle variability. However, the precise mechanisms by which electrical and chemical synapses interact to modulate sequential neural dynamics remain largely unknown. This modeling study examines the role of electrical currents in rhythm modulation within CPGs, with a particular focus on how these currents influence neural excitability and, in turn, modulate the variability of the rhythmic pattern on a cycle-by-cycle basis. We show that changes in neuronal excitability induced by electrical currents can modulate dynamical invariants, leading to the emergence of distinct sequence interval relationships that facilitate the rhythm adaptability. Synchronization between electrically coupled neurons is thus a key factor in regulating these effects. Finally, our results show that sequential dynamical invariants sustain the balance between excitability, variability, and synchronization, underscoring their essential role in the adaptive coordination of functional neural rhythmic patterns.}, archive = {J_NEUCOM}, author = {Blanca Berbel and Roberto Latorre and Pablo Varona}, doi = {10.1016/j.neucom.2025.130218}, journal = {Neurocomputing}, month = {9}, pages = {130218}, shortjournal = {Neurocomputing}, title = {Theoretical bases for the relation between excitability, variability and synchronization in sequential neural dynamics}, volume = {645}, year = {2025}, }
• (2025). ConFusionformer: Locality-enhanced conformer through multi-resolution attention fusion for speaker verification. NEUCOM, 644, 130429. (www)

  @article{ , abstract = {Conformers are capable of capturing both global and local dependencies in a sequence. Notably, the modeling of local information is critical to learning speaker characteristics. However, applying Conformers to speaker verification (SV) has not witnessed much success due to their inferior locality modeling capability and low computational efficiency. In this paper, we propose an improved Conformer, ConFusionformer, to address these two challenges. For increasing model efficiency, the conventional Conformer block is modified by placing one feed-forward network between a self-attention module and a convolution module. The modified Conformer block has fewer model parameters, thus reducing the computation cost. The modification also enables a deeper network, boosting the SV performance. Moreover, multi-resolution attention fusion is introduced into the self-attention mechanism to improve locality modeling. Specifically, the restored map from a low-resolution attention score map produced by downsampled queries and keys is fused with the original attention score map to exploit the local information within the restored local region. The proposed ConFusionformer is shown to outperform the Conformer for SV on VoxCeleb, CNCeleb, SRE21, and SRE24, demonstrating the superiority of the ConFusionformer in speaker modeling.}, archive = {J_NEUCOM}, author = {Youzhi Tu and Man-Wai Mak and Kong-Aik Lee and Weiwei Lin}, doi = {10.1016/j.neucom.2025.130429}, journal = {Neurocomputing}, month = {9}, pages = {130429}, shortjournal = {Neurocomputing}, title = {ConFusionformer: Locality-enhanced conformer through multi-resolution attention fusion for speaker verification}, volume = {644}, year = {2025}, }
• (2025). A comprehensive survey for hyperspectral image classification: The evolution from conventional to transformers and mamba models. NEUCOM, 644, 130428. (www)

  @article{ , abstract = {Hyperspectral Image Classification (HSIC) presents significant challenges owing to the high dimensionality and intricate nature of Hyperspectral data. While traditional Machine Learning (TML) approaches have demonstrated effectiveness, they often encounter substantial obstacles in real-world applications, including the variability of optimal feature sets, subjectivity in human-driven design, inherent biases, and methodological limitations. Specifically, TML suffers from the curse of dimensionality, difficulties in feature selection and extraction, insufficient consideration of spatial information, limited robustness against noise, scalability issues, and inadequate adaptability to complex data distributions. In recent years, Deep Learning (DL) techniques have emerged as robust solutions to address these challenges. This survey offers a comprehensive overview of current trends and future prospects in HSIC, emphasizing advancements from DL models to the increasing adoption of Transformer and Mamba Model architectures. We systematically review key concepts, methodologies, and state-of-the-art approaches in DL for HSIC. Furthermore, we investigate the potential of Transformer-based models and the Mamba Model in HSIC, detailing their advantages and challenges. Emerging trends in HSIC are explored, including in-depth discussions on Explainable AI and interpretability concepts, alongside Diffusion Models for denoising, feature extraction, and fusion. Comprehensive experimental results were conducted on three Hyperspectral datasets to substantiate the efficacy of various conventional DL models. Additionally, we identify several open challenges and pertinent research questions in the field of HSIC. Finally, we outline future research directions and potential applications aimed at enhancing the accuracy and efficiency of HSIC. The Source code is available at https://github.com/mahmad000/HSIC-2024 .}, archive = {J_NEUCOM}, author = {Muhammad Ahmad and Salvatore Distefano and Adil Mehmood Khan and Manuel Mazzara and Chenyu Li and Hao Li and Jagannath Aryal and Yao Ding and Gemine Vivone and Danfeng Hong}, doi = {10.1016/j.neucom.2025.130428}, journal = {Neurocomputing}, month = {9}, pages = {130428}, shortjournal = {Neurocomputing}, title = {A comprehensive survey for hyperspectral image classification: The evolution from conventional to transformers and mamba models}, volume = {644}, year = {2025}, }
• (2025). DGOcc: Depth-aware global query-based network for monocular 3D occupancy prediction. NEUCOM, 644, 130426. (www)

  @article{ , abstract = {Monocular 3D occupancy prediction, aiming to predict the occupancy and semantics within interesting regions of 3D scenes from a single 2D image, has garnered increasing attention recently for its vital role in 3D scene understanding. While prior studies primarily focus on either effectiveness or efficiency, achieving high performance with limited computational resources is still challenging. In this paper, we present DGOcc , a D epth-aware G lobal query-based network for monocular 3D Occ upancy prediction, enabling both effectiveness and efficiency. We first explore prior depth maps to extract depth context features that provide explicit geometric information for the occupancy network. Then, in order to fully exploit the depth context features, we propose a Global Query-based (GQ) Module. The cooperation of attention mechanisms and scale-aware operations facilitates the feature interactions between images and 3D voxels. Moreover, a Hierarchical Selective Supervision (HSS) Strategy is designed to avoid upsampling all of the high-dimension 3D voxel features to full resolution, which mitigates GPU memory utilization and time cost. Extensive experiments on SemanticKITTI and SSCBench-KITTI-360 datasets demonstrate that the proposed method surpasses the current state-of-the-art performance on monocular semantic occupancy prediction while having lower GPU and time overhead.}, archive = {J_NEUCOM}, author = {Xu Zhao and Pengju Zhang and Bo Liu and Yihong Wu}, doi = {10.1016/j.neucom.2025.130426}, journal = {Neurocomputing}, month = {9}, pages = {130426}, shortjournal = {Neurocomputing}, title = {DGOcc: Depth-aware global query-based network for monocular 3D occupancy prediction}, volume = {644}, year = {2025}, }
• (2025). Adaptation and learning of spatio-temporal thresholds in spiking neural networks. NEUCOM, 644, 130423. (www)

  @article{ , abstract = {Spiking neural networks (SNNs) have attracted substantial attention in recent years due to their brain-inspired and event-driven characteristics. To mimic the behavior of biological neurons, the neuron in SNNs generates spikes to transmit information across the network once its membrane potential surpasses a certain firing threshold. Due to model complexity and computational challenges, the threshold is often set as a fixed value, which limits the rich dynamical features of neurons and is inconsistent with the dynamic nature of thresholds observed in biological systems. Additionally, treating the threshold as an optimized parameter presents challenges in achieving convergence and maintaining stability. Therefore, we introduce a spatio-temporal adjustment strategy for the firing threshold. We propose a Learnable Temporal Factor (LTF) to dynamically adapt the threshold over time and an Adaptive Learnable Spatial Factor (ALSF) to spatially extend the threshold. By coupling these factors with the neuronal dynamics, we achieve a stronger spike coding capacity by utilizing more information in the generation of spikes. Our experiments show that the proposed method yields remarkable performance on both static and neuromorphic datasets. Our code is available at github.com/gzxdu/ST-Thresholds-SNN .}, archive = {J_NEUCOM}, author = {Jiahui Fu and Shuiping Gou and Peizhao Wang and Licheng Jiao and Zhang Guo and Jisheng Li and Rong Liu}, doi = {10.1016/j.neucom.2025.130423}, journal = {Neurocomputing}, month = {9}, pages = {130423}, shortjournal = {Neurocomputing}, title = {Adaptation and learning of spatio-temporal thresholds in spiking neural networks}, volume = {644}, year = {2025}, }
• (2025). CIFFormer: A contextual information flow guided transformer for colorectal polyp segmentation. NEUCOM, 644, 130413. (www)

  @article{ , abstract = {Automatic segmentation of polyps in endoscopic images plays a critical role in the early diagnosis of colorectal cancer. In recent years, Visual Transformers, especially pyramid vision transformers, have achieved remarkable strides and become dominating methods in polyp segmentation. However, due to the high resemblance between polyps and normal tissues in terms of size, appearance, color, and other aspects, the pyramid vision transformer methods still face the challenges of the representation of fine-grained details and identifying highly disguised polyps that could be pivotal in precise segmentation of colorectal polyp. To address these challenges, we propose a novel Contextual Information Flow Guided Transformer (CIFFormer) for colorectal polyp segmentation to reconstruct the architecture of a pyramid vision transformer via a contextual information flow design. Our proposed method utilizes a pyramid-structured encoder to obtain multi-resolution feature maps. To effectively capture the target object’s features at various levels of detail, from coarse-grained global information to fine-grained local information, we design a Global-Local Feature Synergy Fusion module (GLFS). GLFS adopts a progressive fusion strategy, first fusing the features of adjacent hierarchy and then gradually fusing across the hierarchy. This allows the model to utilize the features of different semantic levels better and avoid the information loss caused by direct fusion. In addition, we also introduce a Multi-Density Global-Local Residual Module (MDGL). The multi-density residual units within MDGL improve feature propagation and information flow. By employing both local and global residual learning, the model gains a better ability to capture detailed information at both global and local scales. The experimental results demonstrate that our CIFFormer model surpasses 17 benchmark models and achieves state-of-the-art performance on five popular datasets. Furthermore, our model exhibits remarkable performance on two video datasets as well. The source code of this work is available at https://github.com/lonlin404/CIFFormer .}, archive = {J_NEUCOM}, author = {Cunlu Xu and Long Lin and Bin Wang and Jun Liu}, doi = {10.1016/j.neucom.2025.130413}, journal = {Neurocomputing}, month = {9}, pages = {130413}, shortjournal = {Neurocomputing}, title = {CIFFormer: A contextual information flow guided transformer for colorectal polyp segmentation}, volume = {644}, year = {2025}, }
• (2025). BCDAN: A balanced method for community detection in attributed networks. NEUCOM, 644, 130411. (www)

  @article{ , abstract = {Community detection in attribute networks is a fundamental problem in complex network analysis, aiming to identify well-connected communities based on both topology and node attributes. Although numerous methods for community detection in attribute networks have been proposed, they often fail to balance the use of structural and attribute information. In these approaches, attribute information typically plays a dominant role, while higher-order structural information, such as motif structures, is often neglected. To address this issue, we propose a balanced method for community detection in attribute networks (BCDAN). First, for structural information, we construct structural vectors for each node and employ a structure autoencoder to learn the nodes’ structural embeddings. Concurrently, for attribute information, we use an attribute autoencoder and a graph convolution module to jointly learn the nodes’ attribute embeddings, linking the two types of neural networks via a delivery operator. Furthermore, we introduce a self-supervised module that integrates both structural and attribute information from the target distribution to guide iterative updates of the model, thereby facilitating the community detection task. This approach enables a balanced utilization of both attribute and structural information, improving the identification of community features in attribute networks. Extensive experimental results demonstrate that BCDAN outperforms most baseline algorithms.}, archive = {J_NEUCOM}, author = {Yuchen Liu and Yabin Peng and Hongchang Chen and Jianpeng Zhang and Shaomei Li}, doi = {10.1016/j.neucom.2025.130411}, journal = {Neurocomputing}, month = {9}, pages = {130411}, shortjournal = {Neurocomputing}, title = {BCDAN: A balanced method for community detection in attributed networks}, volume = {644}, year = {2025}, }
• (2025). Reinforcement learning-based fixed-time tracking control for nonlinear systems with asymmetrical guaranteed performance. NEUCOM, 644, 130410. (www)

  @article{ , abstract = {This paper develops an optimized fixed-time tracking control method for nonlinear systems characterized by unknown dynamics and unknown virtual control coefficients (UVCCs), employing optimization as its foundational framework. Utilizing the backstepping approach, the system is decomposed into several subsystems. By employing the neural network (NN)-based reinforcement learning strategy with a critic-only architecture, the optimal virtual control of these subsystems is achieved, and the overall control of the system is optimized. Specifically, the unknown dynamics are approximated using radial basis function neural networks (RBFNNs). Additionally, novel Lyapunov functions and adaptive methods are introduced to eliminate the need for lower and upper bounds of the UVCCs. Introducing a unique global fixed-time prescribed function and barrier function fulfills asymmetrical prescribed performance requirements. Theoretical analysis demonstrates that the proposed controller ensures bounded signals in the closed-loop system, and the tracking error converges to the neighborhood of the origin within a fixed time. Finally, a numerical example validates the efficacy of the proposed method.}, archive = {J_NEUCOM}, author = {Quan Wan and Tianpeng Fan and Zhongguo Li and Zhengtao Ding}, doi = {10.1016/j.neucom.2025.130410}, journal = {Neurocomputing}, month = {9}, pages = {130410}, shortjournal = {Neurocomputing}, title = {Reinforcement learning-based fixed-time tracking control for nonlinear systems with asymmetrical guaranteed performance}, volume = {644}, year = {2025}, }
• (2025). Mapping-to-parameter nonlinear functional regression with iterative local B-spline knot placement. NEUCOM, 644, 130403. (www)

  @article{ , abstract = {Many real-world phenomena are inherently continuous, yet traditionally represented as finite-dimensional vectors or matrices of discrete data points. Functional data analysis offers a natural paradigm by modeling observations as continuous functions, preserving intrinsic continuity and structural dependencies, thereby better capturing real-world dynamics and their underlying truth. However, functional modeling within infinite-dimensional spaces presents significant challenges due to its infinite degrees of freedom and computational complexity. These difficulties have led most studies on functional regression to focus on linear models, with general nonlinear approaches remaining underdeveloped. This paper introduces the Mapping-to-Parameter model, a simple yet effective approach for nonlinear functional regression. The key idea is straightforward: transform nonlinear functional regression problems from infinite-dimensional function spaces to finite-dimensional parameter spaces, where standard machine learning techniques can be readily applied. This transformation is accomplished by uniformly approximating all input or output functions using a common set of B-spline basis functions of any chosen order and representing each function by its vector of basis coefficients. For optimal approximation, we develop a novel Iterative Local Placement algorithm that adaptively distributes knots according to localized function complexity while providing theoretical guarantees on approximation error bounds. The performance of the proposed knot placement algorithm is shown to be robust and efficient in both single-function approximation and multiple-function approximation contexts. Through several real-world applications, the effectiveness and superiority of the Mapping-to-Parameter model are demonstrated in handling both function-on-scalar regression and function-on-function regression problems, consistently outperforming state-of-the-art methods including statistical functional models, neural network models, and functional neural networks.}, archive = {J_NEUCOM}, author = {Chengdong Shi and Xiao-Jun Zeng and Ching-Hsun Tseng and Wei Zhao}, doi = {10.1016/j.neucom.2025.130403}, journal = {Neurocomputing}, month = {9}, pages = {130403}, shortjournal = {Neurocomputing}, title = {Mapping-to-parameter nonlinear functional regression with iterative local B-spline knot placement}, volume = {644}, year = {2025}, }
• (2025). Adaptive multilevel regression integration with error compensation for online soft sensing of data streams. NEUCOM, 644, 130397. (www)

  @article{ , abstract = {In industrial data stream environments, achieving reliable online quality monitoring is challenging due to the complexity and time-varying characteristic of industrial processes. The key lies in capturing the mapping relationships between variables in complex processes and continuously adapting to online data while ensuring efficiency. To this end, a novel deep learning-based soft sensor with online learning update mechanism, A daptive M ultilevel regression I ntegration with E rror C ompensation (AMIEC), is proposed. Firstly, to enhance the ability to cope with data streams, AMIEC dynamically weights and integrates multilevel regression outcomes from layers of various depths during the online period. Secondly, a hierarchical supervision strategy is introduced to facilitate effective optimization, and its validity is demonstrated through a theoretical analysis. Moreover, a real-time error compensation module is included to capture the trends in estimation errors and implement corrections. Finally, experimental evaluations are conducted on the sulfur recovery unit (SRU) benchmark and a real-world polyester esterification process. The results show that AMIEC can efficiently handle continuous data streams and achieve reliable online soft sensing, significantly outperforming existing adaptive soft sensor methods.}, archive = {J_NEUCOM}, author = {Guomin Wu and Lei Chen and Hengqian Wang and Chuang Peng and Kuangrong Hao}, doi = {10.1016/j.neucom.2025.130397}, journal = {Neurocomputing}, month = {9}, pages = {130397}, shortjournal = {Neurocomputing}, title = {Adaptive multilevel regression integration with error compensation for online soft sensing of data streams}, volume = {644}, year = {2025}, }
• (2025). Multimodal misinformation detection based on the multi-granularity consistency. NEUCOM, 644, 130393. (www)

  @article{ , abstract = {In the current era of information proliferation, social media platforms frequently present multimodal data that may propagate misinformation. To address this challenge, computer vision techniques have been increasingly applied to online misinformation detection. Existing approaches often concentrate on learning deep semantic features from text and images without sufficiently exploring coherent alignments across multiple granular levels. This could hinder the construction of a comprehensive multimodal detection framework for misinformation detection. To overcome these issues, in this paper, we propose a three-level granularity consistency detection framework that explores image-text alignments at fine-grained, medium-grained, and coarse-grained levels. Meanwhile, we delve into the hierarchical structures of two modalities as well as the intricate semantic relationships between them. Specifically, our framework models token-level consistency through a multi-head cross-attention mechanism, phrase-level consistency via a graph neural network (GNN), and global-level consistency using a contrastive language-image pre-training model (CLIP), respectively. An adaptive fusion module then integrates these consistency features across all three granularities, enabling more nuanced and robust multimodal misinformation detection. Extensive experiments conducted on multiple benchmark datasets demonstrate that our approach substantially improves the performance of multimodal misinformation detection.}, archive = {J_NEUCOM}, author = {Shibo Zhang and Hongchang Chen and Shuxin Liu and Ran Li and HaoCong Jiang and Liang Dong}, doi = {10.1016/j.neucom.2025.130393}, journal = {Neurocomputing}, month = {9}, pages = {130393}, shortjournal = {Neurocomputing}, title = {Multimodal misinformation detection based on the multi-granularity consistency}, volume = {644}, year = {2025}, }
• (2025). Sentiment-enhanced triplet region detection framework for aspect sentiment triplet extraction. NEUCOM, 644, 130392. (www)

  @article{ , abstract = {Aspect Sentiment Triplet Extraction (ASTE) is a crucial task in sentiment analysis that focuses on identifying aspect terms, corresponding opinion terms and the sentiments expressed, ultimately forming sentiment triplets. Recent tagging-based methods where triplets are represented by a two-dimensional matrix of word-pair relations, have achieved promising performance on ASTE. However, existing tagging-based methods usually treat multi-word aspect or opinion terms as a set of independent words and solely focus on the interactions between words, causing the model to potentially extract only parts of the aspect terms or opinion terms, thus producing incorrect triplets. To alleviate this limitation, we propose Sentiment-Enhanced Triplet Region Detection Framework (SE-TRDF) for ASTE. Instead of classifying the relationship between each word pair directly, SE-TRDF conducts ASTE by detecting a triplet region enclosed by aspect and opinion words firstly and then classifying the sentiment polarity of the detected region. With the goal of identifying the four vertices of the triplet region, SE-TRDF is able to obtain the correct aspect and opinion terms with multi-word. Moreover, to further enhance the capability of our SE-TRDF, a sentiment detection module (SDM) is devised which aims to identify the aspect and opinion terms within a sentence and helps to filter some wrong triplet regions detected. Extensive experimental results on four public benchmarks demonstrate the superiority of the proposed SE-TRDF compared with baselines.}, archive = {J_NEUCOM}, author = {Xiaoyong Liu and Jiahao Chen and Chunlin Xu and Zhiguo Du and Weiqi Chen and Huihui Li}, doi = {10.1016/j.neucom.2025.130392}, journal = {Neurocomputing}, month = {9}, pages = {130392}, shortjournal = {Neurocomputing}, title = {Sentiment-enhanced triplet region detection framework for aspect sentiment triplet extraction}, volume = {644}, year = {2025}, }
• (2025). Complex graph neural networks for multi-hop propagation. NEUCOM, 644, 130364. (www)

  @article{ , abstract = {Graph neural networks have made significant advancements over the years, enabling them to tackle a wide range of real-world tasks. However, many existing models still face persistent challenges, such as limitations in representing directed edges and difficulties in aggregating information from distant nodes. Aggregating distant node information typically requires stacking multiple layers, which can lead to over-smoothing and reduce the network’s effectiveness. To address these challenges, we propose a complex Laplacian matrix that effectively captures directed edge information and design an approximation method utilizing the Implicitly Restarted Lanczos (IRL) algorithm. This approach provides a low-rank approximation of the complex Laplacian matrix, facilitating fast convolution on both low-order and high-order nodes. Building on this foundation, we introduce the Complex Aggregating Graph Network (CAGN), which can aggregate both low-order and high-order node information without the need for multiple stacked layers. Extensive experiments on multiple real-world datasets demonstrate that CAGN achieves strong performance in node classification and link prediction tasks, with particularly notable results in link prediction, validating the model’s effectiveness. The code is available on https://github.com/LaZuiYa/CAGN .}, archive = {J_NEUCOM}, author = {Dong Xu and Ziye Liu and Fengming Li and Yulong Meng}, doi = {10.1016/j.neucom.2025.130364}, journal = {Neurocomputing}, month = {9}, pages = {130364}, shortjournal = {Neurocomputing}, title = {Complex graph neural networks for multi-hop propagation}, volume = {644}, year = {2025}, }
• (2025). Efficient multi-satellite trajectory planning: Multi-agent soft actor–critic reinforcement learning with mixed expert experience replay for formation reconfiguration. NEUCOM, 644, 130354. (www)

  @article{ , abstract = {This paper addresses the problem of multi-satellite trajectory planning in which the satellite formation requires different configurations to observe targets on Earth. In order to truly interpret the space trajectory planning scenario, a physics engine-based multi-satellite Trajectory Planning Scenario (MTPS) is established, which enables the real-time interaction between the agent with the environment and the learned strategy to be more realistic. To address the slow convergence of traditional reinforcement learning methods, a multi-agent soft actor–critic algorithm based on mixed expert experience net (SCP-MASAC NET) is proposed. Sequential convex programming (SCP) is integrated into multi-agent reinforcement learning (MARL) to enable multi-satellites to generate higher quality trajectories with a better convergence rate. Specifically, the trajectory data generated by the sequential convex programming method is added to the expert experience database and sampled in the update of the critic and actor network with a monotonically decreasing probability, facilitating faster convergence. The advanced performance of the proposed algorithm in MTPS are validated through comparative experiments. Compared with the traditional reinforcement learning method, the SCP-MASAC NET exhibits a faster convergence rate and illustrates a 10.38% reduction in energy consumption compared to the SCP algorithm.}, archive = {J_NEUCOM}, author = {Da Liu and Xiuyun Zhang and Qun Zong and Hongbo Li and Liqian Dou and Bailing Tian}, doi = {10.1016/j.neucom.2025.130354}, journal = {Neurocomputing}, month = {9}, pages = {130354}, shortjournal = {Neurocomputing}, title = {Efficient multi-satellite trajectory planning: Multi-agent soft actor–critic reinforcement learning with mixed expert experience replay for formation reconfiguration}, volume = {644}, year = {2025}, }
• (2025). A multi-scale vision mixture-of-experts for salient object detection with Kolmogorov–Arnold adapter. NEUCOM, 644, 130349. (www)

  @article{ , abstract = {Diverse domains and object variations make salient object detection (SOD) a challenging task in computer vision. Many previous studies have adopted multi-scale neural networks with attention mechanisms. Although they are popular, the design of their networks lacks sufficient flexibility, which hinders their generalization across objects of different scales and domains. To address the above issue, we propose a novel mixture-of-experts salient object detection (MoESOD) approach. We design a multi-scale mixture-of-experts (MMoE) module, essentially large neural networks, to improve the model’s expressive power and generalization ability without significantly increasing computational cost. By leveraging expert competition and collaboration strategies, the MMoE module effectively integrates contributions from different experts. The MMoE module not only captures multi-scale features but also effectively fuses semantic information across scales through the expert gating mechanism. Additionally, the novel Kolmogorov–Arnold adapter (KAA) is designed to enhance the model’s flexibility, allowing it to adapt easily to SOD tasks across different domains. Comprehensive experiments show that MoESOD consistently achieves higher performance than, or at least comparable performance to, state-of-the-art methods on 17 different SOD benchmarks and 1 downstream tasks. To the best of our knowledge, this is the first study to explore Kolmogorov–Arnold network within the SOD community.}, archive = {J_NEUCOM}, author = {Chaojun Cen and Fei Li and Ping Hu and Zhenbo Li}, doi = {10.1016/j.neucom.2025.130349}, journal = {Neurocomputing}, month = {9}, pages = {130349}, shortjournal = {Neurocomputing}, title = {A multi-scale vision mixture-of-experts for salient object detection with Kolmogorov–Arnold adapter}, volume = {644}, year = {2025}, }
• (2025). SU-YOLO: Spiking neural network for efficient underwater object detection. NEUCOM, 644, 130310. (www)

  @article{ , abstract = {Underwater object detection is critical for oceanic research and industrial safety inspections. However, the complex optical environment and the limited resources of underwater equipment pose significant challenges to achieving high accuracy and low power consumption. To address these issues, we propose Spiking Underwater YOLO (SU-YOLO), a Spiking Neural Network (SNN) model. Leveraging the lightweight and energy-efficient properties of SNNs, SU-YOLO incorporates a novel spike-based underwater image denoising method based solely on integer addition, which enhances the quality of feature maps with minimal computational overhead. In addition, we introduce Separated Batch Normalization (SeBN), a technique that normalizes feature maps independently across multiple time steps and is optimized for integration with residual structures to capture the temporal dynamics of SNNs more effectively. The redesigned spiking residual blocks integrate the Cross Stage Partial Network (CSPNet) with the YOLO architecture to mitigate spike degradation and enhance the model’s feature extraction capabilities. Experimental results on URPC2019 underwater dataset demonstrate that SU-YOLO achieves mAP of 78.8% with 6.97M parameters and an energy consumption of 2.98 mJ, surpassing mainstream SNN models in both detection accuracy and computational efficiency. These results underscore the potential of SNNs for engineering applications. The code is available in https://github.com/lwxfight/snn-underwater .}, archive = {J_NEUCOM}, author = {Chenyang Li and Wenxuan Liu and Guoqiang Gong and Xiaobo Ding and Xian Zhong}, doi = {10.1016/j.neucom.2025.130310}, journal = {Neurocomputing}, month = {9}, pages = {130310}, shortjournal = {Neurocomputing}, title = {SU-YOLO: Spiking neural network for efficient underwater object detection}, volume = {644}, year = {2025}, }
• (2025). AstMatch: Adversarial self-training consistency framework for semi-supervised medical image segmentation. NEUCOM, 643, 130491. (www)

  @article{ , abstract = {Semi-supervised learning (SSL) has demonstrated significant potential in medical image segmentation, primarily through consistency regularization and pseudo-labeling. However, many SSL approaches focus predominantly on low-level consistency, neglecting the significance of pseudo-label reliability. Therefore, in this work, we propose an adversarial self-training consistency framework (AstMatch). First, we design an adversarial consistency regularization (ACR) approach to enhance knowledge transfer and strengthen prediction consistency under varying intensity perturbations. Second, we incorporate a feature matching loss within adversarial training to achieve high-level consistency regularization. Furthermore, we present the pyramid channel attention (PCA) and efficient channel and spatial attention (ECSA) modules to enhance the discriminator’s effectiveness. Finally, we propose an adaptive self-training (AST) approach to ensure high-quality pseudo-labels. The proposed AstMatch has been extensively evaluated with state-of-the-art SSL methods on three publicly available datasets. Experimental results across different labeled ratios demonstrate that AstMatch outperforms other existing methods, achieving new state-of-the-art performance. Our code is publicly available at http://github.com/GuanghaoZhu663/AstMatch .}, archive = {J_NEUCOM}, author = {Guanghao Zhu and Jing Zhang and Juanxiu Liu and Xiaohui Du and Ruqian Hao and Yong Liu and Lin Liu}, doi = {10.1016/j.neucom.2025.130491}, journal = {Neurocomputing}, month = {8}, pages = {130491}, shortjournal = {Neurocomputing}, title = {AstMatch: Adversarial self-training consistency framework for semi-supervised medical image segmentation}, volume = {643}, year = {2025}, }
• (2025). A review of zeroing neural network: Theory, algorithm and application. NEUCOM, 643, 130425. (www)

  @article{ , abstract = {As a powerful method for solving complex computational equations, neural networks have attracted widespread attention due to their unique advantages. However, due to the existence of time-varying problems in practical applications, traditional Gradient neural network (GNN) models may not be able to satisfy the requirements for accurately solving such problems, leading to the emergence of a dynamic system solution method - Zeroing neural network (ZNN), a dynamic system solver designed specifically for solving various time-varying mathematical problems and real-time control applications. ZNN eliminates errors through the use of dynamic differential equations, fundamentally overcoming the limitations of GNN in effectively converging for time-varying problems. Additionally, considering different application demands in practical scenarios and interference from noise in realistic environments, various robust ZNN models with different convergence properties have emerged. This paper will summarize the development of ZNN models in recent years from theoretical foundation, algorithm improvement and practical application aspects, and finally prospect the future research directions of ZNN models to provide researchers with a systematic reference.}, archive = {J_NEUCOM}, author = {Shaoting Cao and Jie Jin and Daobing Zhang and Chaoyang Chen}, doi = {10.1016/j.neucom.2025.130425}, journal = {Neurocomputing}, month = {8}, pages = {130425}, shortjournal = {Neurocomputing}, title = {A review of zeroing neural network: Theory, algorithm and application}, volume = {643}, year = {2025}, }
• (2025). A state-of-the-art survey on neural computation-enhanced Dempster–Shafer theory for safety accidents: Applications, challenges, and future directions. NEUCOM, 643, 130419. (www)

  @article{ , abstract = {Frequent workplace accidents associated with rapid industrialization has become the global focus. Such accidents not only pose a serious threat to the safety and health of employees but can also incur substantial economic losses and reputational damage for companies. Therefore, determining ways to effectively reduce workplace accidents and their possibilities has become an urgent problem that needs to be solved. In the field of safety management, numerous scholars and practitioners are committed to researching and applying various theories and methods to improve workplace safety. Recently, the Dempster–Shafer theory (DST) has garnered attention an important uncertainty reasoning method in the field of safety management. By integrating information from multiple sources, the theory can make effective reasoning using uncertain or incomplete information and provide strong support for safety management decision-making. Recently, neural computation techniques have been explored to enhance the reasoning capabilities of DST, enabling more efficient fusion and analysis of complex and high-dimensional safety data. In view of this, this study systematically reviews relevant papers on DST in the field of safety, accident, and emergency management and forms a systematic literature review (SLR). To support and guide the completion of this work, this paper proposes three research questions based on the 4R crisis management theoretical framework. On the same theoretical basis, the selected literature is sorted and analyzed according to the four dimensions of R e d u c t i o n , R e a d i n e s s , R e s p o n s e and R e c o v e r y , covering all aspects of these dimensions; moreover, the results are obtained through SLR. The final results show that DST enhanced with neural computation plays vital roles in the field of safety, accident, and emergency management. This coupling reduces the limitations of DST, expanding its scope in the process. However, several limitations remain to be solved. Accordingly, this paper analyzes ways to solve the extended theory, practical application dimension, and related defects of DST in the field of safety, accident, and emergency management and draws some conclusions.}, archive = {J_NEUCOM}, author = {Liguo Fei and Tao Li and Weiping Ding}, doi = {10.1016/j.neucom.2025.130419}, journal = {Neurocomputing}, month = {8}, pages = {130419}, shortjournal = {Neurocomputing}, title = {A state-of-the-art survey on neural computation-enhanced Dempster–Shafer theory for safety accidents: Applications, challenges, and future directions}, volume = {643}, year = {2025}, }
• (2025). Emotion recognition via affective EEG signals: State of the art. NEUCOM, 643, 130418. (www)

  @article{ , abstract = {With advancements in brain–computer interface technology, research on emotion recognition based on electroencephalogram (EEG) signals has gained significant attention. This review systematically explores signal acquisition, feature extraction, classification methods, and applications related to emotion recognition. We begin by reviewing the acquisition of affective EEG signals, including emotion models, emotion induction methods, signal acquisition techniques, and popular public datasets. Second, we provide a detailed discussion of feature extraction methods for emotional EEG signals, including time-domain, frequency-domain, time–frequency domain, and spatial domain features, as well as feature fusion techniques. The classification methods section highlights recent developments in machine learning, deep learning, and multimodal learning, exploring their applications in emotion recognition tasks. Additionally, we assess practical applications of emotion recognition technologies in areas such as cognitive workload, fatigue estimation, neuropsychiatric condition assessment, and affective care. Finally, the article summarizes the major challenges currently faced and the future development opportunities. By synthesizing existing research, we provide valuable insights and guidance for further studies on EEG-based emotion recognition and its applications in various fields such as education, transportation, and healthcare.}, archive = {J_NEUCOM}, author = {Wei Meng and Fazheng Hou and Mengyuan Zhao and Jingjing Kong and Jiahao Wu and Jie Zuo and Quan Liu}, doi = {10.1016/j.neucom.2025.130418}, journal = {Neurocomputing}, month = {8}, pages = {130418}, shortjournal = {Neurocomputing}, title = {Emotion recognition via affective EEG signals: State of the art}, volume = {643}, year = {2025}, }
• (2025). Vision transformers on the edge: A comprehensive survey of model compression and acceleration strategies. NEUCOM, 643, 130417. (www)

  @article{ , abstract = {In recent years, vision transformers (ViTs) have emerged as powerful and promising techniques for computer vision tasks such as image classification, object detection, and segmentation. Unlike convolutional neural networks (CNNs), which rely on hierarchical feature extraction, ViTs treat images as sequences of patches and leverage self-attention mechanisms. However, their high computational complexity and memory demands pose significant challenges for deployment on resource-constrained edge devices. To address these limitations, extensive research has focused on model compression techniques and hardware-aware acceleration strategies. Nonetheless, a comprehensive review that systematically categorizes these techniques and their trade-offs in accuracy, efficiency, and hardware adaptability for edge deployment remains lacking. This survey bridges this gap by providing a structured analysis of model compression techniques, software tools for inference on edge, and hardware acceleration strategies for ViTs. We discuss their impact on accuracy, efficiency, and hardware adaptability, highlighting key challenges and emerging research directions to advance ViT deployment on edge platforms, including graphics processing units (GPUs), application-specific integrated circuit (ASICs), and field-programmable gate arrays (FPGAs). The goal is to inspire further research with a contemporary guide on optimizing ViTs for efficient deployment on edge devices.}, archive = {J_NEUCOM}, author = {Shaibal Saha and Lanyu Xu}, doi = {10.1016/j.neucom.2025.130417}, journal = {Neurocomputing}, month = {8}, pages = {130417}, shortjournal = {Neurocomputing}, title = {Vision transformers on the edge: A comprehensive survey of model compression and acceleration strategies}, volume = {643}, year = {2025}, }
• (2025). A dual-branch deep interaction network for multi-channel speech enhancement. NEUCOM, 643, 130412. (www)

  @article{ , abstract = {Multi-channel speech enhancement removes noise and reverberation interference from noisy speech signals captured by microphone arrays. In this paper, we propose a dual-branch deep interaction network (DBDINet) for multi-channel speech enhancement, which complements the important features of both time domain and time–frequency domain in the speech signal. We design a waveform and complex spectrum interaction module (WCIM) to interact deeply with the information of two domains and propose an efficient Conformer (eConformer) as a transition layer of the network to improve network efficiency. We conducted extensive experiments on the synthetic AISHELL-1 dataset and the CHiME-3 dataset. The experimental results show that the proposed method achieves competitive performance on several metrics while maintaining lower computational complexity with faster inference speed than existing advanced methods.}, archive = {J_NEUCOM}, author = {Xiaoyu Lian and Nan Xia and Gaole Dai and Hongqin Yang}, doi = {10.1016/j.neucom.2025.130412}, journal = {Neurocomputing}, month = {8}, pages = {130412}, shortjournal = {Neurocomputing}, title = {A dual-branch deep interaction network for multi-channel speech enhancement}, volume = {643}, year = {2025}, }
• (2025). Population-based meta-heuristic optimization algorithm booster: An evolutionary and learning competition scheme. NEUCOM, 643, 130405. (www)

  @article{ , abstract = {In a Population-Based Meta-Heuristic Optimization Algorithm (PMOA), individuals in the population will constantly generate new promising individuals, to form new populations. Although the population continuously changes, the variations in each individual are traceable in most algorithms. An individual in the population comes from the individual in the previous population. The direction of the evolution of populations can be identified on top of this historical inheritance relationship, which improves the efficiency of PMOAs and solves optimization problems more effectively. Since Recurrent Neural Networks (RNNs) are able to capture the temporal dependencies in sequences, we are motivated to propose a novel but simple Evolutionary and Learning Competition Scheme (ELCS), also referred to as the PMOA Booster, in which individuals keep changing for the better fitness based on the heuristic rules of the PMOA while an RNN is used to learn the process that each individual changes in order to guide the generation of promising individuals. The ELCS automatically selects the RNN or PMOA which generates more individuals with the better fitness. We test the proposed scheme using the benchmark of IEEE Congress on Evolutionary Computation 2022 competition (CEC 2022). The results show that this scheme is able to boost the performance of both the classical and state-of-the-art PMOAs and outperforms its counterparts. Also, the ELCS produces promising results in two real-world industrial scenarios. We believe that the effectiveness of the proposed ELCS is due to the adaptive competition between the RNN and the PMOA.}, archive = {J_NEUCOM}, author = {Jun Wang and Junyu Dong and Huiyu Zhou and Xinghui Dong}, doi = {10.1016/j.neucom.2025.130405}, journal = {Neurocomputing}, month = {8}, pages = {130405}, shortjournal = {Neurocomputing}, title = {Population-based meta-heuristic optimization algorithm booster: An evolutionary and learning competition scheme}, volume = {643}, year = {2025}, }
• (2025). Incorporating edge sharpening and covariance attention for named entity recognition. NEUCOM, 643, 130402. (www)

  @article{ , abstract = {Named Entity Recognition (NER) is a key application in the field of Artificial Intelligence and Natural Language Processing, which automatically identifies and categorizes entities in text by intelligent algorithms. In NER, all spans of a sentence can be organized into a two-dimensional representation. The semantic plane has the advantage to represent the semantic structure of a sentence and to learn the interaction between spans. One of an important phenomenon of this representation is that neighboring elements of the semantic plane are spans denoted to overlapped subsequences in a sentence. Because they share the same contextual features and semantic dependencies, it is difficult to distinguish true entities from the backgrounds. Therefore, refining span representations and building the semantic dependency between spans is helpful for the entity recognition task. In this paper, we propose an Edge Sharpening and Covariance Attention (ES&CA) model to support recognizing named entities from the semantic plane representation. The edge sharpening (ES) module adopts a differential convolution to sharpen the semantic gradients in the semantic plane, which has the ability to gather semantic information from neighborhoods. In the covariance attention (CA) module, the covariance between spans are applied to weight the attention of spans relevant to task-relevant learning objective. Establishing semantic relationships across spans is a highly successful approach. The ES&CA model is assessed on five public datasets for nested and flattened named entity recognition. The evaluation results demonstrate the effectiveness of our strategy in distinguishing entity spans from the backgrounds, hence significantly enhancing the final performance.}, archive = {J_NEUCOM}, author = {Caiwei Yang and Yanping Chen and Shuai Yu and Ruizhang Huang and Yongbin Qin}, doi = {10.1016/j.neucom.2025.130402}, journal = {Neurocomputing}, month = {8}, pages = {130402}, shortjournal = {Neurocomputing}, title = {Incorporating edge sharpening and covariance attention for named entity recognition}, volume = {643}, year = {2025}, }
• (2025). Spatio-temporal prediction using graph neural networks: A survey. NEUCOM, 643, 130400. (www)

  @article{ , abstract = {The analysis of spatial time series is increasingly relevant as spatio-temporal data are becoming widespread due to the ever-growing diffusion of data acquisition devices. Spatio-temporal prediction is crucial for grasping insights on spatio-temporal dynamics in diverse domains. In many cases, spatio-temporal data can be effectively represented using graphs, thus making Graph Neural Networks the most sounding deep learning architecture for the modelling of spatio-temporal series. The aim of the work is to provide a self-consistent and thorough overview on Graph Neural Networks for spatio-temporal prediction, giving a taxonomy of the diverse approaches proposed in the literature. Moreover, attention is paid to the description of the most used benchmarks and metrics in different real-world spatio-temporal domains and to the discussion of the main drawbacks of spatio-temporal Graph Neural Networks. Furthermore, unlike other similar works on deep learning, statistical methods for spatio-temporal modelling are briefly surveyed in this work. Finally, insights on future developments of Graph Neural Networks for spatio-temporal prediction are suggested.}, archive = {J_NEUCOM}, author = {Vincenzo Capone and Angelo Casolaro and Francesco Camastra}, doi = {10.1016/j.neucom.2025.130400}, journal = {Neurocomputing}, month = {8}, pages = {130400}, shortjournal = {Neurocomputing}, title = {Spatio-temporal prediction using graph neural networks: A survey}, volume = {643}, year = {2025}, }
• (2025). Universality conditions of unified classical and quantum reservoir computing. NEUCOM, 643, 130391. (www)

  @article{ , abstract = {Reservoir computing is a versatile paradigm in computational neuroscience and machine learning, that exploits a recurrent neural network to efficiently process time-dependent information. The power of many neural network architectures resides in their universality approximation property. As widely known, classes of reservoir computers serve as universal approximators of functionals with fading memory. The construction of such universal classes often appears context-specific, but, in fact, they follow the same principles. Here we present a unified theoretical framework and we propose a ready-made setting to secure universality, based on the minimal sufficient conditions for a class of reservoir computers to be universal, namely the fading memory and the polynomial algebra structure of the set of their associated functionals. We test the result in the arising context of quantum reservoir computing. Guided by such a unified theorem we suggest why spatial multiplexing serves as a computational resource when dealing with quantum registers, as empirically observed in specific implementations on quantum hardware. The analysis sheds light on a unified view of classical and quantum reservoir computing.}, archive = {J_NEUCOM}, author = {Francesco Monzani and Enrico Prati}, doi = {10.1016/j.neucom.2025.130391}, journal = {Neurocomputing}, month = {8}, pages = {130391}, shortjournal = {Neurocomputing}, title = {Universality conditions of unified classical and quantum reservoir computing}, volume = {643}, year = {2025}, }
• (2025). Efficient incremental model reduction approach for time-varying spatially distributed processes. NEUCOM, 643, 130390. (www)

  @article{ , abstract = {The accurate and efficient modeling of complex spatially distributed processes (SDPs) represents a significant challenge for researchers in the field. A variety of modeling approaches have been developed with the objective of addressing issues related to spatiotemporal modeling. The conventional offline modeling approaches necessitate the availability of a spatiotemporal data set that encompasses the entirety of the system’s dynamic features. It is not possible to fully satisfy this prerequisite in the context of real-world applications, particularly in the case of time-varying SDPs. Furthermore, the process of updating the model by repeatedly applying the offline algorithm is inherently time-consuming. In this paper, we put forth an efficient incremental model reduction approach for time-varying SDPs. Following the initialization phase, the modified Gram–Schmidt orthogonalization method is employed to extract the feature subspace in a sequential manner. The time-space synthesis is then utilized to reconstruct the spatiotemporal dynamics. The efficacy of the model is evaluated on a representative diffusion-reaction process, and the comparative experimental outcomes demonstrate that the presented algorithm is an efficient and effective approach for the online learning of time-varying SDPs.}, archive = {J_NEUCOM}, author = {Ai Ling and Hao Ru and Xueqin Chen and Kok Lay Teo}, doi = {10.1016/j.neucom.2025.130390}, journal = {Neurocomputing}, month = {8}, pages = {130390}, shortjournal = {Neurocomputing}, title = {Efficient incremental model reduction approach for time-varying spatially distributed processes}, volume = {643}, year = {2025}, }
• (2025). MSLoRA: Meta-learned scaling for adaptive fine-tuning of LoRA. NEUCOM, 643, 130374. (www)

  @article{ , abstract = {Low rank adaptation (LoRA) methods have demonstrated strong capabilities in efficiently fine-tuning large models. However, existing LoRA-based approaches typically require manually setting the scaling factor, a process that involves extensive search efforts to find optimal values. To address this challenge, we first develop data-driven heuristic methods that automatically determine layer-wise scaling factors through either activation pattern analysis during forward propagation or gradient behavior monitoring during backward updates. However,their practical performance remains unsatisfactory in applications. Building upon these theoretical foundations, we present MSLoRA, a novel framework that reformulates scaling factor determination as a dynamic optimization problem in parameter-efficient fine-tuning. Our approach innovatively models scaling factors as self-adaptive meta-parameters whose optimal values emerge organically through the interplay between transformer architecture hierarchies and task-specific learning objectives. Extensive experiments conducted across both natural language understanding and generative tasks reveal that MSLoRA consistently outperforms baseline models. This highlights the effectiveness of MSLoRA’s dynamic, layer-specific adjustment mechanism in capturing the complex nature of task-specific activation patterns, making it a more robust and scalable solution for parameter-efficient fine-tuning of large models.}, archive = {J_NEUCOM}, author = {Dan Luo and Kangfeng Zheng and Chunhua Wu and Xiujuan Wang}, doi = {10.1016/j.neucom.2025.130374}, journal = {Neurocomputing}, month = {8}, pages = {130374}, shortjournal = {Neurocomputing}, title = {MSLoRA: Meta-learned scaling for adaptive fine-tuning of LoRA}, volume = {643}, year = {2025}, }
• (2025). Beneficial and flowing: Omni efficient feature aggregation network for image super-resolution. NEUCOM, 643, 130369. (www)

  @article{ , abstract = {Image super-resolution (SR) is a classic low-level vision task that reconstructs high-resolution (HR) images from low-resolution (LR) ones. Recent Transformer-based SR methods have achieved remarkable performance through modeling long-range dependencies. However, there are two critical challenges in existing approaches: (1) Invalid feature interaction and homogeneous aggregation scheme; (2) Network feature propagation blocking. These challenges imply that the potential of Self-Attention (SA) and Transformer architecture is still not fully exploited. To this end, we propose a novel Transformer model, Omni Efficient Aggregation Transformer (OEAT), boosting SR performance by mining and aggregating efficient information across the omni-dimensions and omni-stages. Specifically, we first design an Omni Efficient Aggregation Self-Attention (OEASA) with a local–global-channel feature interaction scheme, to aggregate heterogeneous features from multi-scales and multi-dimensions while facilitating information flow. Specifically, we design a global semantic SA based on content self-similarity in the spatial dimension and an adaptive sparse channel SA in the channel dimension, efficiently gathering the most useful feature for accurate reconstruction. Furthermore, we design a simple yet effective Omni Feature Fusion (OFF) to enhance global groups feature fusion and inter-layer adaptive feature aggregation, thus introducing more critical information for an accurate reconstruction. Extensive experiments demonstrate that our OEAT outperforms recent state-of-the-art methods both quantitatively and qualitatively.}, archive = {J_NEUCOM}, author = {Zhicun Zhang and Yu Han and Xiaoqi Xi and Linlin Zhu and Chunhui Wang and Siyu Tan and Lei Li and Bin Yan}, doi = {10.1016/j.neucom.2025.130369}, journal = {Neurocomputing}, month = {8}, pages = {130369}, shortjournal = {Neurocomputing}, title = {Beneficial and flowing: Omni efficient feature aggregation network for image super-resolution}, volume = {643}, year = {2025}, }
• (2025). StMHCG: High-confidence multi-view clustering for identification of spatial domains from spatially resolved transcriptomics. NEUCOM, 643, 130362. (www)

  @article{ , abstract = {Recent advancements within the field of spatially resolved transcriptomics (SRT) have vastly augmented the repertoire of opportunities for delineating the intricate landscapes of gene expression across tissue spatial dimensions. Currently proposed methods primarily focus on examining the consistency information between spatial location and gene expression combinations, yet they overlook potential complementary information. In response to this, we propose a multi-view clustering method for SRT data, named stMHCG, which is designed to address the challenge of inadequate spatial expression information learning. Firstly, by utilizing a spatial expression augmentation module, we conduct an in-depth exploration of the transcriptome signal differences between individual spots and their physically adjacent spots, capturing subtle variations that may be overlooked in traditional spatial view constructions and generating enriched, complementary augmented views. Furthermore, to enhance the guiding role of consistent information in the clustering process, we have designed a high-confidence clustering guidance module. This module dynamically calculates the target distribution for each view and adjusts the clustering results accordingly, thereby enabling a finer-grained segmentation of the spatial domain. We validated stMHCG across multiple tissue types and technology platforms by comparing it with existing typical methods. Experimental results demonstrate that stMHCG exhibits excellent performance in downstream analysis tasks such as spatial domain identification, trajectory inference, and data denoising of SRT data.}, archive = {J_NEUCOM}, author = {Xiaohan Zhang and Junliang Shang and Yan Zhao and Baojuan Qin and Qianqian Ren and Feng Li and Jin-Xing Liu}, doi = {10.1016/j.neucom.2025.130362}, journal = {Neurocomputing}, month = {8}, pages = {130362}, shortjournal = {Neurocomputing}, title = {StMHCG: High-confidence multi-view clustering for identification of spatial domains from spatially resolved transcriptomics}, volume = {643}, year = {2025}, }
• (2025). Text similarity based on two independent channels: Siamese convolutional neural networks and siamese recurrent neural networks. NEUCOM, 643, 130355. (www)

  @article{ , abstract = {In the present-day context, a large amount of information exists in text. It is hard to extract meaningful and potential information from the text. From the current research, text similarity provides a method applied in many practical scenarios. Traditional text similarity algorithms are easy to implement, but more than these algorithms are needed to extract text features. At present, most text similarity algorithms are based on deep learning. However, these algorithms often struggle with adequately extracting both local and context text features, and they typically do not differentiate between the effectiveness of these two types of feature extraction. To address these shortcomings, this paper proposes Two Independent Channels: Siamese Convolutional Neural Networks and Siamese Recurrent Neural Networks (TIC-SCNN-SRNN). This approach is designed for binary classification, where ‘1’ indicates similarity and ‘0’ indicates dissimilarity. In detail, this paper proposes the Siamese Convolutional Neural Networks (SCNN) model to address the issue of insufficient extraction of local text features. Additionally, it introduces the Siamese Recurrent Neural Networks (SRNN) model to tackle the problem of insufficient extraction of context text features. Due to the issue of not distinguishing the effects of local and context text features extractions, this paper conducts independent weight learning on these two models to research which is more effective for text similarity tasks. In order to verify the effectiveness of the model, this paper experiments on SciTail, TwitterPPD, and QQP datasets. The experimental results show that SCNN and SRNN influence the text similarity tasks, but SRNN is more effective than SCNN. Furthermore, to verify the advantages of the TIC-SCNN-SRNN model, this paper tests the performance of the state-of-the-art CNN&RNN models. The test results show that the TIC-SCNN-SRNN model performs the best, indicating that the model proposed in this paper is more effective for the text similarity tasks.}, archive = {J_NEUCOM}, author = {Zhengfang He}, doi = {10.1016/j.neucom.2025.130355}, journal = {Neurocomputing}, month = {8}, pages = {130355}, shortjournal = {Neurocomputing}, title = {Text similarity based on two independent channels: Siamese convolutional neural networks and siamese recurrent neural networks}, volume = {643}, year = {2025}, }
• (2025). Multi-scale spatio-temporal memory network for semi-supervised video object segmentation. NEUCOM, 642, 130487. (www)

  @article{ , abstract = {Recently, video object segmentation has received great attention in the computer vision community. Most of the existing methods exhibit poor segmentation performance when dealing with complex scenes with local information confusion such as foreground-foreground similarity and foreground-background similarity. To tackle this problem, this study proposes a semi-supervised video target segmentation (semi-VOS) framework multi-scale memory (MSM) based on spatio-temporal memory, aiming to solve the problem of insufficient local attention by constructing the correlation of space and time in a given video. The spatio-temporal memory network is used as the basic framework to display and store the target appearance feature calculated from the historical frames in the external memory, resulting in employing the historical frame feature information over a long period of time. Considering the difficulty of computing local correlations, a filtering mechanism is designed to remove the global noise in the memory reading stage. An atrous spatial pyramid pooling module is added before decoding to prevent the local information loss induced by the downsampling operation. Extensive experiments are conducted on video object segmentation benchmarks including DAVIS-16 validation, DAVIS-17 validation and test, and YouTube-2018 validation datasets. The experimental results demonstrate the feasibility and effectiveness of the proposed framework on various complex scenarios compared with previous methods.}, archive = {J_NEUCOM}, author = {Hui Wang and Yuqian Zhao and Fan Zhang and Lingli Yu and Chunhua Yang}, doi = {10.1016/j.neucom.2025.130487}, journal = {Neurocomputing}, month = {8}, pages = {130487}, shortjournal = {Neurocomputing}, title = {Multi-scale spatio-temporal memory network for semi-supervised video object segmentation}, volume = {642}, year = {2025}, }
• (2025). Automatic sign language to text translation using MediaPipe and transformer architectures. NEUCOM, 642, 130421. (www)

  @article{ , abstract = {This study presents a transformer-based architecture for translating Sign Language to spoken language text using embeddings of body keypoints, with the mediation of glosses. To the best of our knowledge, this work is the first to successfully leverage body keypoints for Sign Language-to-text translation, achieving comparable performance to baseline models without reducing translation quality. Our approach introduces extensive augmentation techniques for body keypoints, and convolutional keypoint embeddings, and integrates Connectionist Temporal Classification Loss and position encoding for Sign2Gloss translation. For the Gloss2Text stage, we employ fine-tuning of BART, a state-of-the-art transformer model. Evaluation on the Phoenix14T dataset demonstrates that our integrated Sign2Gloss2Text model achieves competitive performance, with BLEU-4 scores that show marginal differences compared to baseline models using pixel embeddings. On the How2Sign dataset, which lacks gloss annotations, direct Sign2Text translation posed challenges, as reflected in lower BLEU-4 scores, highlighting the limitations of gloss-free approaches. This work addresses the narrow domain of the datasets and the unidirectional nature of the translation process while demonstrating the potential of body keypoints for Sign Language Translation. Future work will focus on enhancing the model’s ability to capture nuanced and complex contexts, thereby advancing accessibility and assistive technologies for bridging communication between individuals with hearing impairments and the hearing community.}, archive = {J_NEUCOM}, author = {Wesley F. Maia and António M. Lopes and Sergio A. David}, doi = {10.1016/j.neucom.2025.130421}, journal = {Neurocomputing}, month = {8}, pages = {130421}, shortjournal = {Neurocomputing}, title = {Automatic sign language to text translation using MediaPipe and transformer architectures}, volume = {642}, year = {2025}, }
• (2025). Robust fMRI time-varying functional connectivity analysis using multivariate swarm decomposition. NEUCOM, 642, 130404. (www)

  @article{ , abstract = {Time-varying functional connectivity (TVFC) measured with functional MRI (fMRI) captures dynamic changes in statistical dependencies among regional time series, which can be studied with instantaneous phase synchronization analyses. Phase extraction requires narrow-banded resting-state fMRI (rs-fMRI) data typically extracted with conventional band-pass filtering or advanced mode decomposition techniques. However, filtering methods often struggle to eliminate noise effectively, require prior knowledge of cutoff frequencies, and fail to account for non-stationarity in the data. Likewise, existing mode decomposition techniques strongly depend on input parameters and are less reliable for multivariate analyses. Here, we introduce multivariate swarm decomposition (MSwD), a bio-inspired signal decomposition technique that combines the iterative nature of empirical methods with a robust mathematical foundation. Using synthetic signals and real rs-fMRI data from the Human Connectome Project and the Autism Brain Imaging Data Exchange I, we showed that MSwD-based PS (MSwD-PS) outperforms four state-of-the-art decomposition techniques in several key areas: (1) being more robust to input parameters and better at detecting true synchronizations, showing a 3–65% lower normalized root mean square error in simulated data and being 15.8-73.1% less prone to identifying short biologically implausible transitions between brain states, and (2) showing a reduced likelihood of false positives, being less affected by spurious synchronizations. Likewise, MSwD-informed functional connectivity analysis improved subject fingerprinting and autism spectrum disorder classification using graph neural networks. Overall, MSwD-PS can reduce the risk of false positives in TVFC, which could be extremely useful for processing rs-fMRI data with unknown ground truth in diverse clinical populations.}, archive = {J_NEUCOM}, author = {Charalampos Lamprou and Georgios Apostolidis and Aamna Alshehhi and Leontios J. Hadjileontiadis and Mohamed L. Seghier}, doi = {10.1016/j.neucom.2025.130404}, journal = {Neurocomputing}, month = {8}, pages = {130404}, shortjournal = {Neurocomputing}, title = {Robust fMRI time-varying functional connectivity analysis using multivariate swarm decomposition}, volume = {642}, year = {2025}, }
• (2025). Modeling of distributed parameter systems based on independent partial derivative-physics-informed neural network. NEUCOM, 642, 130396. (www)

  @article{ , abstract = {Physics-informed neural networks (PINNs) have attracted considerable interest due to their capacity to incorporate established physical principles, thus facilitating efficient training with a limited amount of observed data. This approach ensures that the results of supervised learning are in accordance with the governing physical laws of the system, thereby enhancing the interpretability of models. However, existing research predominantly employs a single network architecture to simultaneously learn the information of the partial differential equations (PDEs). Such architectures typically require large network sizes to learn the system characteristics, which leads to prolonged training times and inefficiencies. This paper introduces a novel modeling framework, termed the independent partial derivative-physics-informed neural network (IPD-PINN). In contrast to conventional methodologies, IPD-PINN employs a network comprising multiple subnetworks, with each subnetwork dedicated to the learning of individual partial derivatives in the PDEs. The modular network structure allows for flexible and scalable adjustments to the size of each subnetwork, accommodating modeling tasks of varying complexities. Consequently, the entire network is capable of reducing the required training time while maintaining precise modeling accuracy. The efficacy of the proposed IPD-PINN method is demonstrated through two simulation case studies, which highlight its potential to improve the efficiency and accuracy of IPD-PINN modeling.}, archive = {J_NEUCOM}, author = {Lijie Wang and Yangshu Lin and Xinrong Yan and Yuhao Shao and Zuhua Xu and Chao Yang and Haidong Fan and Yurong Xie and Chenghang Zheng}, doi = {10.1016/j.neucom.2025.130396}, journal = {Neurocomputing}, month = {8}, pages = {130396}, shortjournal = {Neurocomputing}, title = {Modeling of distributed parameter systems based on independent partial derivative-physics-informed neural network}, volume = {642}, year = {2025}, }
• (2025). Over-squashing in graph neural networks: A comprehensive survey. NEUCOM, 642, 130389. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) revolutionize machine learning for graph-structured data, effectively capturing complex relationships. They disseminate information through interconnected nodes, but long-range interactions face challenges known as ”over-squashing”. This survey delves into the challenge of over-squashing in GNNs, where long-range information dissemination is hindered, impacting tasks reliant on intricate long-distance interactions. It comprehensively explores the causes, consequences, and mitigation strategies for over-squashing. Various methodologies are reviewed, including graph rewiring, novel normalization, spectral analysis, and curvature-based strategies, with a focus on their trade-offs and effectiveness. The survey also discusses the interplay between over-squashing and other GNN limitations, such as over-smoothing, and provides a taxonomy of models designed to address these issues in node and graph-level tasks. Benchmark datasets for performance evaluation are also detailed, making this survey a valuable resource for researchers and practitioners in the GNN field.}, archive = {J_NEUCOM}, author = {S. Akansha}, doi = {10.1016/j.neucom.2025.130389}, journal = {Neurocomputing}, month = {8}, pages = {130389}, shortjournal = {Neurocomputing}, title = {Over-squashing in graph neural networks: A comprehensive survey}, volume = {642}, year = {2025}, }
• (2025). MDiffSR: Mutual information and diffusion model in image super-resolution. NEUCOM, 642, 130372. (www)

  @article{ , abstract = {In recent years, diffusion model has shown powerful functions and great potential in image super-resolution (SR) tasks due to its superiority in imaging capability and stability. Although it is increasingly used in image SR tasks and has achieved many gratifying results, it still has challenges such as insufficient utilization of original image information, resulting in blurred image details. In addition, the inference process for diffusion models can be time-consuming and difficult to train. To solve these problems, we propose the MDiffSR model, which improves the speed of image detail generation and training. Specifically, we introduce mutual information (MI) loss and contrast the idea of learning to improve the ability of model to utilize raw image information. In addition, in order to enhance the generation of image details, we integrate multi-head attention mechanism into the model, which enhances the ability of the model to capture global image features. Experimental results on multiple datasets show that our model outperforms existing methods both visually and quantitatively.}, archive = {J_NEUCOM}, author = {Mingze Jiang and Jinfu Fan and Linqing Huang and Zhencun Jiang and Qingkai Bu}, doi = {10.1016/j.neucom.2025.130372}, journal = {Neurocomputing}, month = {8}, pages = {130372}, shortjournal = {Neurocomputing}, title = {MDiffSR: Mutual information and diffusion model in image super-resolution}, volume = {642}, year = {2025}, }
• (2025). A novel adaptive weighted fusion network based on pixel level feature importance for two-stage 6D pose estimation. NEUCOM, 642, 130371. (www)

  @article{ , abstract = {In intelligent industry, accurate recognition and localization of objects in an image is the basis for robots to perform autonomous and intelligent operations. With the rapid development and application of deep learning data fusion technology in pose estimation, the existing 6D pose estimation methods have made many achievements. However, most of the existing methods are not accurate enough to cope with scenes with cluttered backgrounds, inconspicuous textures, and occluded objects. In addition, the existing methods ignore the effect of the accuracy of instance segmentation on the accuracy of pose estimation. To address above issues, this paper proposes a two-stage 6D pose estimation method based on adaptive pixel-importance weighted fusion network with lightweight instance segmentation, named TAPWFusion. In the instance segmentation stage, a lightweight instance segmentation network based on multiscale attention and boundary constraints, named CVi-BC-YOLO, is proposed to improve segmentation accuracy and efficiency. In the pose estimation stage, to eliminate the interference of lighting and occlusion, and enhance the accuracy of the pose estimation, we propose an adaptive pixel-importance weighted fusion network, named APWFusion, which adaptively evaluates the importance of RGB color and the geometrical information of the point cloud. Experiments on LineMOD, YCB-Video and T-LESS datasets prove the advanced and effective nature of our proposed method.}, archive = {J_NEUCOM}, author = {Haitao Xiao and Linkun Ma and Qinyao Li and Shuo Ma and Hongxuan Guo and Wenjie Wang and Harutoshi Ogai}, doi = {10.1016/j.neucom.2025.130371}, journal = {Neurocomputing}, month = {8}, pages = {130371}, shortjournal = {Neurocomputing}, title = {A novel adaptive weighted fusion network based on pixel level feature importance for two-stage 6D pose estimation}, volume = {642}, year = {2025}, }
• (2025). Bridging modal gaps: A cross-modal feature complementation and feature projection network for visible-infrared person re-identification. NEUCOM, 642, 130367. (www)

  @article{ , abstract = {Visible-infrared person re-identification (VI-ReID) presents a significant challenge due to the substantial modal differences between infrared (IR) and visible (VIS) images, primarily resulting from their distinct color distributions and textural characteristics. One effective strategy for reducing this modal gap is to utilize feature projection to create a shared embedded space for the modal features. However, a key research question remains: how to effectively align cross-modal features during projection while minimizing the loss of information. To address this challenge, this paper proposed a Cross-Modal Feature Complementation and Feature Projection Network (FCFPN). Specifically, a modal complementation strategy was introduced to bridge the discrepancies between cross-modal features and facilitate their alignment. Additionally, a cross-modal feature projection mechanism was employed to embed modality-correlated features into the shared feature space, thereby mitigating feature loss caused by modality differences. Furthermore, multi-channel and multi-level features were extracted from the shared space to enhance the overall feature representation. Extensive experimental results demonstrated that the proposed FCFPN model effectively mitigated the modal discrepancy, achieving 84.7% Rank-1 accuracy and 86.9% mAP in the indoor test mode of the SYSU-MM01 dataset, and 93.0% Rank-1 accuracy and 87.3% mAP in the VIS-to-IR test mode of the RegDB dataset, thereby outperforming several state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Lei Shi and Yumao Ma and Yongcai Tao and Haowen Liu and Lin Wei and Yucheng Shi and Yufei Gao}, doi = {10.1016/j.neucom.2025.130367}, journal = {Neurocomputing}, month = {8}, pages = {130367}, shortjournal = {Neurocomputing}, title = {Bridging modal gaps: A cross-modal feature complementation and feature projection network for visible-infrared person re-identification}, volume = {642}, year = {2025}, }
• (2025). Generative neural architecture search. NEUCOM, 642, 130360. (www)

  @article{ , abstract = {Neural architecture search (NAS) is an important approach for automatic neural architecture design and has been applied to many tasks, such as image classification and object detection. However, most of the conventional NAS algorithms mainly focus on reducing the prohibitive computational cost, while choosing commonly used reinforcement learning (RL), evolutionary algorithm (EA) or gradient-based methods as their search strategy. In this paper, we propose a novel search strategy for NAS, called Generative NAS (GNAS). Specifically, we assume that high-performing convolutional neural networks adhere to a latent distribution, and design a generator to learn this distribution for generating neural architectures. Furthermore, in order to update the generator for better learning the latent distribution, we use the policy gradient and the performance of the generated CNNs on the validation datasets as a reward signal. To evaluate GNAS, we have conducted extensive experiments on the CIFAR-10, SVHN, MNIST, Fashion-MNIST and ImageNet datasets. The results demonstrate the effectiveness of GNAS compared to previous NAS strategies.}, archive = {J_NEUCOM}, author = {Xiaotong Zhai and Shu Li and Guoqiang Zhong and Tao Li and Fuchang Zhang and Rachid Hedjam}, doi = {10.1016/j.neucom.2025.130360}, journal = {Neurocomputing}, month = {8}, pages = {130360}, shortjournal = {Neurocomputing}, title = {Generative neural architecture search}, volume = {642}, year = {2025}, }
• (2025). Discovering causality for efficient cooperation in multi-agent environments. NEUCOM, 642, 130358. (www)

  @article{ , abstract = {In cooperative Multi-Agent Reinforcement Learning (MARL) agents are required to learn behaviours as a team to achieve a common goal. However, while learning a task, some agents may end up learning sub-optimal policies, not contributing to the objective of the team. Such agents are called lazy agents due to their non-cooperative behaviours that may arise from failing to understand whether they caused the rewards. As a consequence, we observe that the emergence of cooperative behaviours is not necessarily a byproduct of being able to solve a task as a team. In this paper, we investigate the applications of causality in MARL and how it can be applied in MARL to penalise these lazy agents. We observe that causality estimations can be used to improve the credit assignment to the agents and show how it can be leveraged to improve independent learning in MARL. Furthermore, we investigate how Amortised Causal Discovery can be used to automate causality detection within MARL environments. The results demonstrate that causality relations between individual observations and the team reward can be used to detect and punish lazy agents, making them develop more intelligent behaviours. This results in improvements not only in the overall performances of the team but also in their individual capabilities. In addition, results show that Amortised Causal Discovery can be used efficiently to find causal relations in MARL.}, archive = {J_NEUCOM}, author = {Rafael Pina and Varuna De Silva and Corentin Artaud}, doi = {10.1016/j.neucom.2025.130358}, journal = {Neurocomputing}, month = {8}, pages = {130358}, shortjournal = {Neurocomputing}, title = {Discovering causality for efficient cooperation in multi-agent environments}, volume = {642}, year = {2025}, }
• (2025). ProIn: Learning to predict trajectory based on progressive interactions for autonomous driving. NEUCOM, 642, 130346. (www)

  @article{ , abstract = {Accurate motion prediction of pedestrians, cyclists and other vehicles in the environment (all called agents) is very important for autonomous driving. Most existing work captures map information through a one-stage interaction with the map, to provide map constraints for social interaction and multi-modal differentiation. However, these methods need to encode all necessary map rules into the focal agent’s feature, so as to retain all possible paths of intentions while adapting to possible social interactions. In this work, a progressive interaction network based on graph neural networks is proposed that allows the agent feature to progressively focus on relevant maps, in order to better learn the agents’ features capturing the relevant map constraints. The network progressively encodes the complex influence of map constraints into the agent’s feature through graph convolutions in the following three stages: after historical trajectory encoding, after social interaction, and after multi-modal differentiation. Furthermore, a weight allocation mechanism is proposed for multi-modal training, so that each mode has the opportunity to learn from a single-mode ground truth. Experimental results demonstrate the superior accuracy of the proposed method compared to existing methods for trajectory prediction. Especially, it achieves minFDE of 1.12 and 1.31 on the Argoverse 1 and Argoverse 2 test datasets, which improve over the state of the art HeteroGCN by 0.04 and 0.03 respectively.}, archive = {J_NEUCOM}, author = {Yinke Dong and Haifeng Yuan and Hongkun Liu and Wei Jing and Fangzhen Li and Hongmin Liu and Bin Fan}, doi = {10.1016/j.neucom.2025.130346}, journal = {Neurocomputing}, month = {8}, pages = {130346}, shortjournal = {Neurocomputing}, title = {ProIn: Learning to predict trajectory based on progressive interactions for autonomous driving}, volume = {642}, year = {2025}, }
• (2025). Hierarchical language description knowledge base for LLM-based human pose estimation. NEUCOM, 642, 130336. (www)

  @article{ , abstract = {Language plays an important role in human communication and knowledge representation, and human cognition involves numerous top-down and bottom-up processes that rely on different levels of knowledge guidance. To align with human cognition, hierarchical language descriptions can be used in human pose estimation as different levels of knowledge guidance, which is lacking in existing studies. We propose HLanD-Pose, the Hierarchical Language Description Knowledge Base for Human Pose Estimation using the Large Language Model (LLM). It describes human posture from the whole to the components and models the poses within a scene to construct a hierarchical knowledge base. When the relevant knowledge is activated by visual information, the matched hierarchical language description of the current human pose can serve as a guide for performing the keypoint localization task. With the powerful reasoning and language comprehension abilities of large language models, human poses in images can be effectively understood, which helps to recognize and accurately locate the target keypoints. Experiments show the remarkable performance of our method on standard keypoint localization benchmarks. Moreover, the designed hierarchical language description and external knowledge base enhance the model’s superior ability to understand the human body in scene-specific datasets, demonstrating strong generalizable capability in cross-dataset keypoint localization.}, archive = {J_NEUCOM}, author = {Wenjie Chen and Xuemei Xie}, doi = {10.1016/j.neucom.2025.130336}, journal = {Neurocomputing}, month = {8}, pages = {130336}, shortjournal = {Neurocomputing}, title = {Hierarchical language description knowledge base for LLM-based human pose estimation}, volume = {642}, year = {2025}, }
• (2025). Neural network-based fully distributed dynamic event-triggered formation-containment control for nonlinear multi-agent systems. NEUCOM, 642, 130324. (www)

  @article{ , abstract = {Achieving robust and fully distributed time-varying formation-containment control for multi-agent systems (MASs) necessitates a trade-off among control accuracy, communication efficiency, and fully distributed implementation, particularly in the presence of unknown system dynamics. To address this issue, this paper proposes a novel adaptive neural network (NN)-based fully distributed dynamic event-triggered control scheme, which ensures robust formation-containment while minimizing communication and computational overhead. Theoretically, a fully distributed event-triggering condition is developed, which eliminates the need for global knowledge and continuous monitoring. Moreover, an aperiodic NN-based approximator is designed to compensate for unknown dynamics, updating NN weights only when triggered, significantly reducing computational load. To sum up, the proposed time-varying formation-containment controller governs both control inputs and NN weights to ensure convergence of approximation and formation errors, achieving practical distributed control. Unlike conventional continuous mechanisms, the proposed scheme triggers controller updates and state transmissions only when a dynamic condition is satisfied, avoiding Zeno behavior and reducing unnecessary resource consumption. Finally, simulation results confirm the effectiveness of the scheme, demonstrating precise control and reduced resource utilization.}, archive = {J_NEUCOM}, author = {Long You and Kaiyu Qin and Jiangfeng Yue and Weihao Li and Boxian Lin and Mengji Shi}, doi = {10.1016/j.neucom.2025.130324}, journal = {Neurocomputing}, month = {8}, pages = {130324}, shortjournal = {Neurocomputing}, title = {Neural network-based fully distributed dynamic event-triggered formation-containment control for nonlinear multi-agent systems}, volume = {642}, year = {2025}, }
• (2025). Learning-observer-based fixed-time optimal tracking control for robotic manipulators with full-state constraints. NEUCOM, 642, 130186. (www)

  @article{ , abstract = {This paper investigates the fixed-time optimal tracking control problem for an n-link robotic manipulator with full-state constraints and external disturbances. Compared to existing n-link robotic manipulators, this paper adopts the idea of an extended state observer (ESO), integrates neural network (NN) learning capabilities with fixed-time control technology, and establishes a fixed-time learning observer to estimate the states and disturbances of the system. To address the full-state constraints, a nonlinear state-dependent function (NSDF), which purely depends on the constrained states is introduced. In contrast to traditional schemes for handling state constraints, this approach removes the necessity for the virtual controller α i to satisfy feasibility conditions, that is, − χ i 1 ≤ α i ≤ χ i 2 ( i = 1 , … , n − 1 ) should always be satisfied, where χ i 1 and χ i 2 are the boundaries of constraints. Subsequently, a reinforcement learning algorithm with a critic-actor framework is utilized to achieve the fixed-time optimal control objective within the framework of backstepping technology. Furthermore, the proposed fixed-time optimal control method ensures that all signals converge within a fixed time without violating the constraint boundaries. Finally, simulation results are provided to verify the effectiveness of the designed control strategy.}, archive = {J_NEUCOM}, author = {Huizhen Yuan and Liang Cao and Wenshuai Lin and Wenbin Xiao and Xiaomeng Li}, doi = {10.1016/j.neucom.2025.130186}, journal = {Neurocomputing}, month = {8}, pages = {130186}, shortjournal = {Neurocomputing}, title = {Learning-observer-based fixed-time optimal tracking control for robotic manipulators with full-state constraints}, volume = {642}, year = {2025}, }
• (2025). One-shot handwriting imitation via self-supervised cross spatial transformer networks. NEUCOM, 642, 129969. (www)

  @article{ , abstract = {Generating stylized characters has been a focus of research for a long time. However, current font-led algorithms in the field of handwriting generation often perform suboptimally due to severe geometric distortion among handwriting samples. Additionally, the non-Euclidean characteristic of hidden embedding is identified as another key issue in our research. To address shape distortion, we propose a Self-Supervised Style Learning (SSL) method, aiming to increase the skeleton similarity between the output image and target. Furthermore, to enhance the Euclidean spatial distributions of the generator, this paper introduces a joint spatial transformer by refining the fusion process of content and style information using an attention mechanism. Moreover, the training process of the current GAN is reformed to meet the requirements of SSL. As a result, the proposed method enriches the calligraphy diversity of stylized handwriting generation compared to existing font generation algorithms. Evaluation of the proposed method using a public Chinese handwriting database demonstrates that our approach generates handwriting images with superior quality, stroke correctness, and calligraphy similarity when compared to other prevailing font imitation methods.}, archive = {J_NEUCOM}, author = {Bocheng Zhao and Guanwen Feng and Wenxing Zhang and Yunan Li and Qiguang Miao and Xiangzeng Liu and Ruyi Liu}, doi = {10.1016/j.neucom.2025.129969}, journal = {Neurocomputing}, month = {8}, pages = {129969}, shortjournal = {Neurocomputing}, title = {One-shot handwriting imitation via self-supervised cross spatial transformer networks}, volume = {642}, year = {2025}, }
• (2025). Attention-enhanced separable residual with dilation net for medical image segmentation. NEUCOM, 641, 130434. (www)

  @article{ , abstract = {In the process of medical image segmentation, due to the complexity of the organizational structure, many networks suffer from spatial information loss and semantic mismatch due to the limited fusion of features at different scales. In response, this paper proposes the Attention-enhanced Separable Residual with Dilation Net (ASRDNet), a network that incorporates multi-level residual convolution, attention mechanism and cross-layer upsampling connection. This network adopts three complementary designs to enhance the model performance: firstly, the ASRD block in the encoder adaptively weights the information of different channels to enhance the capture of global information; secondly, the skip connection introduces multi-level residual convolution to avoid the loss of the underlying details; and lastly, the decoder optimizes the fusion of multi-scale features through cross-layer upsampling connections, thus improving the segmentation accuracy comprehensively. The test results on five datasets, ETIS-LaribPolypDB, CVC-ColonDB, Kvasir-SEG, DSB-2018, and TCGA-LGG show that ASRDNet achieves 0.9313, 0.9337, 0.9218, 0.9129, and 0.9230 in terms of the Dice metrics, which is excellent among the current existing models. The visualization results of the segmentation show that ASRDNet is robust and maintains clear continuity at the boundary.}, archive = {J_NEUCOM}, author = {Leyi Xiao and Yinuo Liu and Chaodong Fan}, doi = {10.1016/j.neucom.2025.130434}, journal = {Neurocomputing}, month = {8}, pages = {130434}, shortjournal = {Neurocomputing}, title = {Attention-enhanced separable residual with dilation net for medical image segmentation}, volume = {641}, year = {2025}, }
• (2025). FSCMF: A dual-branch frequency-spatial joint perception cross-modality network for visible and infrared image fusion. NEUCOM, 641, 130376. (www)

  @article{ , abstract = {Existing image fusion methods face limitations in deep feature modeling and high-frequency information enhancement, leading to detail loss and reduced target saliency in complex scenarios. To address these issues, this paper proposes a Dual-Branch Frequency-Spatial Joint Perception Cross-Modality Network (FSCMF), which integrates local details, global context, and frequency-domain information through a dual-branch architecture to enhance multimodal feature complementarity. Specifically, FSCMF combines CNN and Transformer in a dual-branch design, where the CNN branch focuses on extracting local structures and texture details, while the Transformer branch captures long-range dependencies to improve global consistency. To further optimize feature representation, we introduce the Frequency-Spatial Adaptive Attention Module (FSAA), in which the frequency domain branch enhances high-frequency components to improve edge sharpness, while the spatial domain branch adaptively refines salient region features, ensuring a dynamic balance between global and local information. Additionally, we propose the Weighted Cross-Spectral Feature Fusion Module (WCSFF) to enhance cross-modality feature interaction through adaptive weighting, thereby improving detail integrity and semantic consistency in the fused image. A maximum frequency loss function is further incorporated to ensure the preservation of critical frequency components. Extensive experiments on three public datasets — MSRS, M 3 FD, and LLVIP — demonstrate that FSCMF outperforms existing methods in both qualitative and quantitative evaluations, producing fusion results with higher visual consistency and better information retention. Furthermore, additional experiments on object detection and semantic segmentation validate FSCMF’s potential in high-level computer vision tasks, highlighting its broad application value. The code of FSCMF is available at https://github.com/boshizhang123/FSCMF .}, archive = {J_NEUCOM}, author = {Xiaoyang Zhang and Chengpei Xu and Guodong Fan and Zhen Hua and Jinjiang Li and Jingchun Zhou}, doi = {10.1016/j.neucom.2025.130376}, journal = {Neurocomputing}, month = {8}, pages = {130376}, shortjournal = {Neurocomputing}, title = {FSCMF: A dual-branch frequency-spatial joint perception cross-modality network for visible and infrared image fusion}, volume = {641}, year = {2025}, }
• (2025). Recursive wavelet transform network for robust copy-move forgery detection. NEUCOM, 641, 130373. (www)

  @article{ , abstract = {In RWTN-Net, the Frequency Rotational-Invariant Feature Extractor (FRFE) firstly performs multi-stage wavelet transform and Sorted Convolution to extract multi-scale rotational invariant low-frequency and high-frequency features, which are robust to geometric transformations. Then, an Adaptive Multi-Scale Attention Fusion (AMAF) is designed to fuse features of different scales with an adaptive attention. The channel weights of low-resolution features are used to guide the weights allocation of high-resolution features, thereby enhancing the network’s understanding of geometric details and semantic information. Moreover, the Local Average Self-Correlation Calculation (LASCC) adopts a diagonal-guided sparse sampling strategy to select key feature points along the diagonal of each patch in the feature map for correlation calculation, which effectively improves the computational efficiency. Finally, a localization module is deployed to combine the matching maps of different receptive fields in an accumulative manner, and an adaptive U-net is further employed to obtain accurate localization results. Experimental results on public datasets demonstrate the effectiveness of the proposed RWTN-Net. The source code is available at https://github.com/studyimg/RWTN-Net .}, archive = {J_NEUCOM}, author = {Yakun Niu and Xinjie Wu and Cheng Liu}, doi = {10.1016/j.neucom.2025.130373}, journal = {Neurocomputing}, month = {8}, pages = {130373}, shortjournal = {Neurocomputing}, title = {Recursive wavelet transform network for robust copy-move forgery detection}, volume = {641}, year = {2025}, }
• (2025). 3C: Confidence-guided clustering and contrastive learning for unsupervised person re-identification. NEUCOM, 641, 130368. (www)

  @article{ , abstract = {Unsupervised person re-identification (Re-ID) aims to learn a feature network with cross-camera retrieval capability in unlabelled datasets. Although pseudo-label based methods have achieved great progress in Re-ID, their effectiveness in complex scenarios is hindered by three unresolved challenges: (i) Noisy pseudo-labels due to disturbed clustering. (ii) Camera-biased clusters that lack diversity. (iii) Unreliable hard samples that amplify noise in contrastive learning. To address these issues, a Confidence-guided Clustering and Contrastive learning (3C) framework is proposed in this paper. The 3C framework presents three confidence degrees: (i) in the clustering stage, the confidence of the discrepancy between samples and clusters is proposed to implement a harmonic discrepancy clustering algorithm (HDC); (ii) in the forward-propagation training stage, the confidence of the camera diversity of a cluster is evaluated using a novel camera information entropy (CIE), and clusters with high CIE will play the leading role in model training; and (iii) in the back-propagation training stage, the confidence of the hard sample in each cluster is designed and further used in a confidence integrated harmonic discrepancy (CHD), to select the informative sample for updating the memory for contrastive learning. Extensive experiments on three popular Re-ID benchmarks demonstrated the superiority of the proposed framework. In particular, the 3C framework achieved state-of-the-art results: 86.7%/94.7%, 45.3%/73.1% and 47.1%/90.6% in term of mAP/Rank-1 accuracy on Market-1501, MSMT17 and VeRi-776, respectively. The code is available at https://github.com/stone5265/3C-reid .}, archive = {J_NEUCOM}, author = {Mingxiao Zheng and Yanpeng Qu and Dongxuan Li and Changjing Shang and Longzhi Yang and Qiang Shen}, doi = {10.1016/j.neucom.2025.130368}, journal = {Neurocomputing}, month = {8}, pages = {130368}, shortjournal = {Neurocomputing}, title = {3C: Confidence-guided clustering and contrastive learning for unsupervised person re-identification}, volume = {641}, year = {2025}, }
• (2025). Deep neurocomputational fusion for ASD diagnosis using multi-domain EEG analysis. NEUCOM, 641, 130353. (www)

  @article{ , abstract = {Autism spectrum disorder (ASD) presents significant challenges in early detection due to its heterogeneous nature and the subtlety of neurophysiological variations. Electroencephalography (EEG) has emerged as a promising tool for ASD diagnosis, offering insights into intricate neurophysiological signal dynamics. However, existing computational EEG analysis techniques often struggle to effectively capture complex ASD-related neural activity patterns. To address these limitations, this study proposes an Encoder-Ensemble Fusion Model (EEFM), designed to enhance EEG-based ASD classification through advanced neurocomputational modeling. EEFM integrates multi-domain EEG features with a sophisticated fusion network, initially extracting a diverse range of time and frequency-domain characteristics. This model enables the identification of nuanced and heterogeneous ASD-specific neural activity, leveraging ensemble learning and autoencoder architectures to uncover complex relationships within the data. Specifically, an LSTM (Long Short-Term Memory)-autoencoder models temporal and spatial dependencies, followed by an XGBoost Regressor for nonlinear EEG signal interpretation and a linear regression model for decision boundary refinement. The proposed model has been evaluated on a dataset focusing on sex-specific EEG variations in ASD and Typically Developing (TD) subjects. By incorporating EEG-based neurophysiological biomarkers, EEFM achieved 93% accuracy in distinguishing ASD from TD individuals. This model enhances sensitivity to subtle ASD-related abnormalities while addressing challenges in data heterogeneity, scarcity, and generalizability, contributing to advancements in neurocomputing methodologies for EEG-based ASD diagnosis.}, archive = {J_NEUCOM}, author = {Abdur Rasool and Saba Aslam and Yongjie Xu and Yishan Wang and Yi Pan and Weiyang Chen}, doi = {10.1016/j.neucom.2025.130353}, journal = {Neurocomputing}, month = {8}, pages = {130353}, shortjournal = {Neurocomputing}, title = {Deep neurocomputational fusion for ASD diagnosis using multi-domain EEG analysis}, volume = {641}, year = {2025}, }
• (2025). Deep scene flow learning from point cloud with transformer. NEUCOM, 641, 130350. (www)

  @article{ , abstract = {Scene flow estimation is the task of obtaining 3D motion from a dynamic scene, which can serve a variety of applications such as robotics and autonomous navigation. Most scene flow estimation methods depend on subsampling the point cloud into groups and extracting features for each group independently. Due to the sparseness of point clouds, extracting features for each group separately produces different features that may all belong to the same object. This difference makes the global correlation prone to unacceptable flow. Local correlation restricts the algorithm to capturing limited movements and fails when fast movement or large deformation of an object occurs. Encoding the features globally is achieved through transforming point clouds into structured shapes, e.g., voxels, which lose some spatial information. Furthermore, calculating global correlation on high-resolution features consumes a lot of memory, which degrades the model’s performance. Therefore, we propose a transformer-based scene flow architecture that can perform global feature encoding and matching with low memory. We utilize self-attention for modeling global relations at one feature scale and adopt the shifted-window strategy to lower memory consumption. Additionally, we propose a cross-attention layer to perform global feature matching and estimate the global scene flow. We further propose a learnable attention-based upsampling layer to upsample the estimated flow to a higher resolution based on a single feature scale, eliminating the need to model global relations at all feature scales. Experimental results show that our model achieves competitive results on the FlyingThings3D dataset and the KITTI benchmark.}, archive = {J_NEUCOM}, author = {Xuezhi Xiang and Yiming Chen and Rokia Abdein and Lei Zhang and Xiantong Zhen}, doi = {10.1016/j.neucom.2025.130350}, journal = {Neurocomputing}, month = {8}, pages = {130350}, shortjournal = {Neurocomputing}, title = {Deep scene flow learning from point cloud with transformer}, volume = {641}, year = {2025}, }
• (2025). Broad learning systems: An overview of recent advances, applications, challenges and future directions. NEUCOM, 641, 130337. (www)

  @article{ , abstract = {Broad learning system (BLS) as a novel training framework derived from the random vector functional link neural network (RVFLNN). Unlike RVFLNN, which directly applies raw data to network learning, BLS first transforms input data into feature nodes through feature mapping. These feature nodes are then nonlinearly transformed into enhanced nodes. Both feature nodes and enhanced nodes are concatenated and connected to the output layer, and the corresponding output weights are derived via pseudo-inverse. Due to its shallow architecture and the need to train only the output weights, BLS achieves highly efficient learning capabilities.Moreover, when adding new nodes, BLS does not require retraining from scratch; it only adjusts the weights associated with the new nodes. Different from deep learning, BLS aims to expand the broad rather than the depth of the neural network to solve complex problems. It not only overcomes the drawback of the long-training process in deep learning but also enables the rapid incremental construction of network models. Therefore, BLS has received extensive attention from both the academic and industrial communities. In view of this, this paper conducts a systematic review of BLS. First, we outline the research background of the BLS. Then, we elaborate on the relevant concepts and definitions of BLS. Furthermore, recent advances in BLS are introduced. In addition, we present the extensive applications of BLS in various fields. Finally, several possible development directions for BLS are proposed.}, archive = {J_NEUCOM}, author = {Yonghe Chu and Yanlong Guo and Weiping Ding and Heling Cao and Peng Ping}, doi = {10.1016/j.neucom.2025.130337}, journal = {Neurocomputing}, month = {8}, pages = {130337}, shortjournal = {Neurocomputing}, title = {Broad learning systems: An overview of recent advances, applications, challenges and future directions}, volume = {641}, year = {2025}, }
• (2025). Multi-population estimation of distribution algorithm for multilevel thresholding in image segmentation. NEUCOM, 641, 130325. (www)

  @article{ , abstract = {Optimization algorithms are crucial for solving complex optimization problems in fields such as image processing. Image segmentation is complicated and requires accurately identifying and outlining regions within an image. Metaheuristic algorithms (MAs) are highly versatile solutions to this challenge. This work presents the Multi-population Estimation of Distribution Algorithm (M-EDA), which estimates distributions with multiple populations. The algorithm is initialized with low discrepancy sequences and random numbers. Two sampling methods are proposed: the first uses the normal distribution with 70% of individuals generated, while the second uses the Cauchy distribution with 30% of individuals. The M-EDA is used for multilevel thresholding, achieving good results with the Berkeley Segmentation Dataset 500 (BSDS500). Additionally, a procedure is developed to detect defects in footwear images, where the M-EDA helps to perform the detection. The study shows that the M-EDA is the best algorithm for optimizing the Minimum Cross-Entropy (MCE), winning at the four thresholding levels studied. Six statistics are used to map the algorithms’ results to different image histogram structures (minimum, first quartile, mean, median, third quartile, and maximum). Furthermore, the results are analyzed using Friedman’s non-parametric statistical test, supporting the significant difference between the statistical results obtained. The same study is applied to eleven images from the BSD500, which succeeds in all cases. Finally, the results obtained in footwear images with scratch defects are visually and numerically competitive because they segment the regions where the defects are found. The source code of the M-EDA can be found at https://github.com/JARF1095/M-EDA .}, archive = {J_NEUCOM}, author = {Jorge Ramos-Frutos and Diego Oliva and Israel Miguel-Andrés and Angel Casas-Ordaz and Oscar Ramos-Soto and Itzel Aranguren and Saúl Zapotecas-Martínez}, doi = {10.1016/j.neucom.2025.130325}, journal = {Neurocomputing}, month = {8}, pages = {130325}, shortjournal = {Neurocomputing}, title = {Multi-population estimation of distribution algorithm for multilevel thresholding in image segmentation}, volume = {641}, year = {2025}, }
• (2025). Spatial-adaptive mixup for domain adaptation in nighttime semantic segmentation. NEUCOM, 641, 130315. (www)

  @article{ , abstract = {In the field of autonomous driving, semantic segmentation model should adapt to different time periods throughout the day. However, nighttime semantic segmentation poses a significant challenge due to the low illumination and the lack of annotated data. In this paper, we propose a novel domain adaptation framework consisting of a bidirectional style transfer step (BST) and multi-target mean teacher step (MTMT), to improve the prediction accuracy. BST sets twilight scenes as the intermediate domain and migrates the style from other domains towards the intermediate one, which helps to promote style alignment. In MTMT, we design a new spatial-adaptive mixup method for dynamic objects (SA-DMix) that employs label based geometric matching method to adaptively adjust the mixing position. It can boost the robustness of dynamic object segmentation and ensure the coherence of contextual information. MTMT further conducts multi-level adversarial learning, including feature-level and prediction-level, to achieve domain alignment and enhance adaptability. Extensive experiments are conducted on the Dark Zurich, ACDC, and Nighttime Driving dataset and the result demonstrates that our method can bring significant improvement in nighttime semantic segmentation.}, archive = {J_NEUCOM}, author = {Zhuoming Gu and Wei Huang and Mengfan Xu and Dan Zeng and Rui Huang}, doi = {10.1016/j.neucom.2025.130315}, journal = {Neurocomputing}, month = {8}, pages = {130315}, shortjournal = {Neurocomputing}, title = {Spatial-adaptive mixup for domain adaptation in nighttime semantic segmentation}, volume = {641}, year = {2025}, }
• (2025). Uncertainty-guided diffusion model for 3D human pose estimation. NEUCOM, 641, 130306. (www)

  @article{ , abstract = {3D human pose estimation (3D HPE) focuses on reconstructing 3D human poses from 2D visual data, including images and video sequences. Recent advancements in diffusion models have significantly propelled the development of 3D HPE. Through comprehensive analysis of estimated 3D pose data, we noticed that the precision of joint estimation shows large variation, revealing that the denoising process in current diffusion models inadequately addresses joints with high uncertainty. Motivated by this critical observation, we present a novel Uncertainty-Guided Diffusion Model for 3D Human Pose Estimation (UDM-HPE), which incorporates probabilistic guidance to refine the denoising process. Specifically, we treat uncertainty attribution as an optimization task and leverage it to identify joints that significantly contribute to output uncertainty. The core innovation of our approach lies in three key technical contributions: (1) a joint-aware uncertainty attribution mechanism specifically designed for 3D joint estimation; (2) a Gumbel-sigmoid reparameterized binary uncertainty mask learning strategy; and (3) an effective integration framework that incorporates the joint-aware uncertainty mask into the diffusion model to guide the denoising process. Our extensive experimental evaluation on standard single-person pose estimation benchmarks demonstrates that UDM-HPE achieves superior performance in both accuracy and robustness compared to state-of-the-art methods. The proposed method shows particular effectiveness and robustness in handling highly flexible poses.}, archive = {J_NEUCOM}, author = {Zhihua Liu and Yuru Wang}, doi = {10.1016/j.neucom.2025.130306}, journal = {Neurocomputing}, month = {8}, pages = {130306}, shortjournal = {Neurocomputing}, title = {Uncertainty-guided diffusion model for 3D human pose estimation}, volume = {641}, year = {2025}, }
• (2025). Identity aware 3D face reconstruction from in-the-wild images. NEUCOM, 641, 130299. (www)

  @article{ , abstract = {Monocular 3D face reconstruction is a challenging task that aims to recover the 3D shape and texture of a human face from a single 2D image. Existing methods based on deep learning have achieved impressive results in terms of geometric details and photorealism. However, they often suffer from the problem of coupling between identity and expression coefficients, which means that the coefficients representing identity and expression are not independent and may interfere with each other. This issue can lead to inaccurate identity estimation and unnatural expression synthesis. To address this issue, we propose a novel deep 3D face reconstruction method called IAFR (Identity-Aware 3D Face Reconstruction), which explicitly disentangles the identity and expression coefficients by introducing a separate identity encoder and an identity-aware loss function. Our method can effectively improve the accuracy of identity coefficients and ensure that they remain invariant across different expressions. Moreover, we present a novel training strategy that leverages temporal consistency to improve the stability of 3DMM coefficients in videos. We conduct extensive experiments on various datasets and demonstrate that our method not only ensures the decoupling of identity coefficients but also outperforms other state-of-the-art methods in terms of reconstruction accuracy and visual quality.}, archive = {J_NEUCOM}, author = {Ruigang Hu and Xuekuan Wang and Cairong Zhao}, doi = {10.1016/j.neucom.2025.130299}, journal = {Neurocomputing}, month = {8}, pages = {130299}, shortjournal = {Neurocomputing}, title = {Identity aware 3D face reconstruction from in-the-wild images}, volume = {641}, year = {2025}, }
• (2025). Dual consistent constraint via disentangled consistency and complementarity for multi-view clustering. NEUCOM, 641, 130225. (www)

  @article{ , abstract = {Multi-view clustering can explore common semantics from multiple views and has received increasing attention in recent years. However, current methods focus on learning consistency in representation, neglecting the contribution of each view’s complementarity aspect in representation learning. This limit poses a significant challenge in multi-view representation learning. This paper proposes a novel multi-view clustering framework that introduces a disentangled variational autoencoder that separates multi-view into shared and private information, i.e., consistency and complementarity information. We first learn informative and consistent representations by maximizing mutual information across different views through contrastive learning. This process will ignore complementary information. Then, we employ consistency inference constraints to explicitly utilize complementary information when attempting to seek the consistency of shared information across all views. Specifically, we perform a within-reconstruction using the private and shared information of each view and a cross-reconstruction using the shared information of all views. The dual consistency constraints are not only effective in improving the representation quality of data but also easy to extend to other scenarios, especially in complex multi-view scenes. This could be the first attempt to employ dual consistent constraint in a unified MVC theoretical framework. During the training procedure, the consistency and complementarity features are jointly optimized. Extensive experiments show that our method outperforms baseline methods.}, archive = {J_NEUCOM}, author = {Bo Li and Jing Yun}, doi = {10.1016/j.neucom.2025.130225}, journal = {Neurocomputing}, month = {8}, pages = {130225}, shortjournal = {Neurocomputing}, title = {Dual consistent constraint via disentangled consistency and complementarity for multi-view clustering}, volume = {641}, year = {2025}, }
• (2025). Prompt-driven graph distillation: Enabling single-layer MLPs to outperform deep graph neural networks in graph-based tasks. NEUCOM, 641, 130106. (www)

  @article{ , abstract = {Graph distillation endeavors to transfer knowledge from large, complex teacher models, such as Graph Neural Networks (GNNs), to smaller, more efficient student models such as Multi-Layer Perceptrons (MLPs). Our work is motivated by a critical observation: as the complexity and depth of the teacher GNN increase, the performance of the distilled student MLP models tends to decline significantly. This issue is not limited to a specific distillation method but is a prevalent challenge across various GNN to MLP approaches. To address this gap, we introduce a novel prompt-driven graph distillation framework that enhances the student model’s input space by appending or integrating learned prompts with the original features. These prompts, derived from the teacher’s input, hidden and output knowledge, provide additional context that assists the MLP student in distilling information, thereby circumventing the need to directly capture any teacher’s knowledge into the simplest single-layer MLP. Our empirical experiment on benchmark graph datasets reveals a counter-intuitive phenomenon: the more capable the teacher, the greater the student’s ability to utilize prompts to outperform the teacher. This finding highlights the strength of our approach: a well-prompted student can indeed surpass its teacher, at achieving the best performance with an accuracy of 87.2% on the Cora standard split with a single-layer MLP, while also maintaining efficiency and robustness. In addition to testing on benchmark graph datasets, we applied our framework to the TUH EEG Epilepsy Corpus (TUEP), condensing complex GNN models to simple MLPs while achieving good performance and efficiency in epilepsy classification.}, archive = {J_NEUCOM}, author = {Shihan Guan and Laurent Albera and Lei Song and Yueying Li and Youyong Kong and Huazhong Shu and Régine Le Bouquin Jeannès}, doi = {10.1016/j.neucom.2025.130106}, journal = {Neurocomputing}, month = {8}, pages = {130106}, shortjournal = {Neurocomputing}, title = {Prompt-driven graph distillation: Enabling single-layer MLPs to outperform deep graph neural networks in graph-based tasks}, volume = {641}, year = {2025}, }
• (2025). SOM-associated-SNN: Enhancing audio classification with spiking neural networks through single-modality clustering and associative learning. NEUCOM, 640, 130416. (www)

  @article{ , abstract = {Audio classification remains challenging due to the complexity and variability of audio signals. This study addresses these challenges by presenting SOM-Associated-SNN, a novel three-layer Spiking Neural Network (SNN) designed for efficient audio classification. Unlike traditional SNNs, our model integrates unsupervised clustering and associative learning to enhance recognition capabilities significantly. The architecture comprises an input layer for audio feature encoding, a hidden layer utilizing Spike-Timing-Dependent Plasticity (STDP) for pattern learning, and an output layer for classification. By eliminating the need for backpropagation, SOM-Associated-SNN reduces computational complexity and more closely aligns with biological neural networks, offering an energy-efficient solution for audio signal processing. The network employs unsupervised clustering via Self-Organizing Maps (SOM) to self-organize data and form clusters based on inherent structures, achieving accurate classification without direct labeling. Associative learning, guided by the teacher layer, fine-tunes synaptic weights based on input-output correlations, enhancing generalization and robustness to audio variations. Experiments on both the Spoken-MNIST and Spiking Heidelberg Digits (SHD) datasets demonstrate significant performance improvements. On Spoken-MNIST, a basic two-layer STDP network achieves 69 % accuracy, which rises to 83.9 % with SOM clustering and further to 93.3 % with associative signals from the teacher layer. On the SHD dataset, our model achieves an accuracy of 88.1 %, showcasing its effectiveness in handling more complex audio classification tasks. The energy-efficient design of SOM-Associated-SNN makes it a promising candidate for speech recognition applications.}, archive = {J_NEUCOM}, author = {Xin Liu and Lingfei Mo and Mengting Tang}, doi = {10.1016/j.neucom.2025.130416}, journal = {Neurocomputing}, month = {8}, pages = {130416}, shortjournal = {Neurocomputing}, title = {SOM-associated-SNN: Enhancing audio classification with spiking neural networks through single-modality clustering and associative learning}, volume = {640}, year = {2025}, }
• (2025). RRAECL: A riemannian manifold graph representation and enhanced contrastive learning framework for label-efficient alzheimer’s disease diagnosis. NEUCOM, 640, 130409. (www)

  @article{ , abstract = {Objective In this study, we focus on the representation and small sample training problem of graph convolutional networks for the early diagnosis of Alzheimer’s disease (AD) based on functional magnetic resonance imaging (fMRI) data. Effective representation to quantify functional connectivity within the brain can help identify brain activity patterns and functional abnormalities associated with AD. Moreover, utilizing feature similarity from unlabeled data to solve the small sample size problem can reduce the demand for labeled data, thereby lowering annotation costs. Methods We propose a Riemannian manifold graph representation and enhanced contrastive learning framework (RRAECL) for label-efficient AD diagnosis . First, we propose a Riemannian manifold graph representation for functional brain networks that can retain more connectivity information and accurately portray the relationships between brain regions. Second, to diagnose AD using Riemannian manifold graphs, we propose a Riemannian region of interest (ROI)-based attention graph convolutional network (RRAGCN) that utilizes an ROI-based attention module to highlight the brain regions crucial for AD diagnosis and weaken the interference from irrelevant regions. Third, we propose an enhanced contrastive learning framework that aligns different sub-segments of the fMRI series in the embedding space to realize an auxiliary characterization for AD diagnosis. Results Experimental results from two databases demonstrate that the proposed method is effective and outperforms other state-of-the-art (SOTA) methods. Conclusion and significance RRAECL demonstrates outstanding performance with only a small amount of labeled fMRI data and is a superior method for label-efficient AD diagnosis.}, archive = {J_NEUCOM}, author = {Xuanqi Wang and Yihao Guo and Feng Chen and Tao Liu and Xiaofeng Xie}, doi = {10.1016/j.neucom.2025.130409}, journal = {Neurocomputing}, month = {8}, pages = {130409}, shortjournal = {Neurocomputing}, title = {RRAECL: A riemannian manifold graph representation and enhanced contrastive learning framework for label-efficient alzheimer’s disease diagnosis}, volume = {640}, year = {2025}, }
• (2025). Auto-weighted exponential tensor nuclear norm minimization for multi-attribute low-dimensional subspace clustering. NEUCOM, 640, 130394. (www)

  @article{ , abstract = {Subspace clustering (SC) is a commonly used clustering method for handling high-dimensional data. However, most existing SC algorithms ignore the multi-attribute information in the original data. In addition, they do not fully utilize the inherent low-dimensional consensus information implied in the original data, resulting in the loss of important data features during the dimensionality reduction process. To solve these issues, we propose the auto-weighted exponential tensor nuclear norm minimization for multi-attribute low-dimensional subspace clustering (MALDSC). Specifically, firstly, we design a triple matrix factorization method to find the inherent low-dimensional consensus information and obtain the corresponding multi-attribute features. Secondly, we utilize the self-expressive property of the multi-attribute features to obtain self-expressive matrices. Thirdly, to harness the complete structural information embedded within the self-expressive matrix, we tack them into a tensor, which is regulated by the auto-weighted exponential tensor nuclear norm (AWETNN), serving as a more effective substitute for the tensor rank. The AWETNN takes full account of the physical differences among singular values through a non-convex penalty function, thus more accurately representing the high-order correlation between multiple attributes. Finally, the augmented Lagrange multiplier method (ALM) is utilized to unify the above three steps into one framework. The experimental results obtained from multiple datasets demonstrate that the MALDSC algorithm outperforms the state-of-the-art algorithms in terms of performance. The code is publicly available at https://github.com/TongWuahpu/MALDSC .}, archive = {J_NEUCOM}, author = {Tong Wu and Gui-Fu Lu and Bing Hu}, doi = {10.1016/j.neucom.2025.130394}, journal = {Neurocomputing}, month = {8}, pages = {130394}, shortjournal = {Neurocomputing}, title = {Auto-weighted exponential tensor nuclear norm minimization for multi-attribute low-dimensional subspace clustering}, volume = {640}, year = {2025}, }
• (2025). AIFE-net: Transformer-based adaptive multi-scale feature extraction encoder for object detection. NEUCOM, 640, 130366. (www)

  @article{ , abstract = {In object detection, the Feature Pyramid Network (FPN) obtains a large amount of multi-scale information due to its feature fusion advantage, which improves the detection performance of targets at different scales in complex scenes. However, there are still deficiencies in feature capture, anti-jamming ability, and information transmission efficiency, which makes it challenging for the model to express the features of the target area in complex environments. To this end, this paper proposes the Adaptive Multi-scale Feature Extraction Encoder (AMFEFormer) based on the Multi-scale Contextual Information Extraction Module (MCIEM) and the Interactive Extraction Transformer Module (IETrans), which compensates for the shortcomings of the FPNs by efficiently and interactively transmitting the FPN's input features from different layers to realise adaptive multi-scale information acquisition. Based on the AMFEFormer, an Adaptive Interactive Feature Extraction (AIFE) detection network was further constructed to distinguish the target and background regions better. In terms of the local scale, MCIEM utilises large kernel expansion convolution and attention to achieve dynamic extraction of multi-scale contextual feature information from different channels under a larger sensory field, capturing semantic and spatial contextual dependencies, and enhancing target area localisation and pixel discrimination. At the global scale, IETrans adopts dual cross-attention dynamic interaction to achieve information fusion and adaptive information regulation of global and local features, establish dependencies between long-range pixels, and suppress irrelevant information interference. The experimental results validate the effectiveness of our method, and AIFE-Net achieves accuracies of 82.78 % and 50.9 % on the PASCAL VOC2007 + 2012 and MS-COCO2017 datasets, respectively, and outperforms many mainstream object detectors.}, archive = {J_NEUCOM}, author = {Zhenyi Zhang and Jianlong Zhang and Juan Wei and Lina Han and Tianping Li}, doi = {10.1016/j.neucom.2025.130366}, journal = {Neurocomputing}, month = {8}, pages = {130366}, shortjournal = {Neurocomputing}, title = {AIFE-net: Transformer-based adaptive multi-scale feature extraction encoder for object detection}, volume = {640}, year = {2025}, }
• (2025). LAA: Local awareness attention for point cloud self-supervised representation learning. NEUCOM, 640, 130365. (www)

  @article{ , abstract = {Local awareness is essential for point cloud representation learning. In recent times, due to the increase in the amount of point cloud data and the success of the self-supervised learning paradigm in other domains, there has been an increase in the number of studies investigating point cloud self-supervised representation learning. However, the majority of current methods for implementing local awareness are incompatible with the paradigm of point cloud self-supervised pre-training, which makes it difficult for pre-trained models to benefit from it. Consequently, previous point cloud pre-training models have predominantly resulted in a global effective receptive field, with less focus on local awareness. A Gaussian-distributed, larger, more natural effective receptive field without artifacts will result in a superior representation of point cloud features. To address this issue, this paper proposes Local Awareness Attention (LAA) , a plug-and-play module that enables local geometric perception while at the same time capturing global features. LAA consists of two branches. The first obtains local geometric information through the attention of each query and its neighborhood. The remaining branch learns global features through self-attention. The LAA module then fuses the features captured by the two branches through a single softmax, resulting in a competitive mechanism that achieves adaptive and multi-scale self-attention. Extensive experiments in indoor environments demonstrate that our LAA obtains stable effect enhancement in multiple transformer-based point cloud self-supervised pretraining networks, specifically outperforming multiple baselines by 0.1%–0.2% in ModelNet40 and by 0.2%–0.3% in ScanObjectNN.}, archive = {J_NEUCOM}, author = {Jiawei Yu and Hongqiang Wu and Wen Shangguan and Yanchang Niu and Biqing Huang}, doi = {10.1016/j.neucom.2025.130365}, journal = {Neurocomputing}, month = {8}, pages = {130365}, shortjournal = {Neurocomputing}, title = {LAA: Local awareness attention for point cloud self-supervised representation learning}, volume = {640}, year = {2025}, }
• (2025). Optimizing photographic composition with deep reinforcement learning. NEUCOM, 640, 130363. (www)

  @article{ , abstract = {Image composition has always been a primary consideration in professional photography, with good composition significantly enhancing the visual appeal of an image. However, achieving rapid composition can be a challenge for most casual users. To address this issue, this paper proposes a view adjustment method to improve image composition. We introduce a DRL (Deep Reinforcement Learning) based approach for adjusting and optimizing image composition. This method employs D3QN (Dueling Double Deep Q-Network) as the main network architecture, leveraging the agent's progressive exploration to find the optimal policy for composition adjustments. Additionally, we have developed a composition feature extraction module that extracts compositional information from images according to the professional photography composition guidelines, thereby further enhancing the network's predictive performance. We also created a composition defect dataset for model training and testing, expanding on existing image cropping datasets. This dataset includes samples of displacement adjustment, scaling adjustment, and their combinations, aiding in the model's training. Through various experiments, our model demonstrated excellent performance, achieving an IoU (Intersection over Union) score of 0.785 and a Disp (Displacement) score of 0.050. The experimental results indicate that our proposed view adjustment method effectively improves image composition and can guide users in real-time to take well-composed photographs. In summary, this research aims to assist casual users in more easily obtaining professional-level composition guidance, utilizing intelligent technical means so that everyone can capture visually appealing moments.}, archive = {J_NEUCOM}, author = {Nan Sheng and Shuai Yang and Huaning Liu and Kai Wang and Yongzhen Ke and Fan Qin}, doi = {10.1016/j.neucom.2025.130363}, journal = {Neurocomputing}, month = {8}, pages = {130363}, shortjournal = {Neurocomputing}, title = {Optimizing photographic composition with deep reinforcement learning}, volume = {640}, year = {2025}, }
• (2025). FREE-net: A dual-modality emotion recognition network for fusing raw and enhanced data. NEUCOM, 640, 130361. (www)

  @article{ , abstract = {Speech Emotion Recognition (SER) holds a significant position in the fields of natural language processing and affective computing. Although existing methods are able to recognize and classify emotions in speech to some extent, the captured speech information is often incomplete, and there are still limitations in feature fusion, temporal modeling, and data augmentation, which restricts recognition accuracy. This paper proposes a dual-modal emotion recognition network (FREE-Net) that achieves comprehensive modeling of speech-text dual-modal features by fusing raw and enhanced data. The Raw Modality Emotion Module (RMEM) and Enhanced Modality Emotion Module (EMEM) are designed to capture fine-grained emotional information, while dynamic interaction is achieved through the Dialogue-Driven Spectral Encoder (DDSE) and the Text-Driven Attention Encoder (TDAE) to improve recognition accuracy. Experimental results demonstrate that FREE-Net outperforms existing methods on the IEMOCAP and MELD datasets, effectively enhancing the robustness and accuracy of emotion recognition.}, archive = {J_NEUCOM}, author = {Wei Wei and Bingkun Zhang and Yibing Wang}, doi = {10.1016/j.neucom.2025.130361}, journal = {Neurocomputing}, month = {8}, pages = {130361}, shortjournal = {Neurocomputing}, title = {FREE-net: A dual-modality emotion recognition network for fusing raw and enhanced data}, volume = {640}, year = {2025}, }
• (2025). Cross-lingual prompt and knowledge enhancement-based named entity recognition for low-resource electronic medical records. NEUCOM, 640, 130357. (www)

  @article{ , abstract = {Pre-trained language models (PLMs) have rapidly evolved and show impressive performance in downstream tasks involving high-resource languages. However, these methods are insufficient in domain-specific tasks (e.g., medical entity recognition) involving low-resource languages (e.g., Mongolian). i. Mongolian is a low-resource language with complex word formation rules and a high degree of adhesion. ii. The clinical annotation data of Mongolian medicine are very scarce and difficult to annotate, leading to double-low-resource scenarios. Recently, prompt learning has shown good performance with PLMs, especially in small-sample scenarios. However, the impressive performance of prompt learning relies on the a priori knowledge of PLMs, and training a PLM for a vertical domain in such double-low-resource scenarios is difficult. Additionally, prompt design is a complex process requiring extensive expertise and considerable time for optimisation and refinement. In this paper, we propose a cross-lingual prompt and knowledge enhancement model based on named entity recognition for use with low-resource electronic medical records (CLPKB). First, word embeddings for source language prompts and data are mapped to target language vector space through a cross-lingual mapping matrix to reduce the spatial distance between the source and target languages. Then, cross-lingual soft prompts are used to transfer rich semantic features from the source to the target language to reduce the data dependence of the model. Meanwhile , external knowledge bases are utilised to enhance the contextual representations of the cross-lingual prompts to capture more domain-specific knowledge. The results show that our model improved by an average of 5.87% over the state-of-the-art baseline models.}, archive = {J_NEUCOM}, author = {Yurong Wang and Min Lin and Qitu Hu and Shuangcheng Bai and Yanling Li and Longjie Bao}, doi = {10.1016/j.neucom.2025.130357}, journal = {Neurocomputing}, month = {8}, pages = {130357}, shortjournal = {Neurocomputing}, title = {Cross-lingual prompt and knowledge enhancement-based named entity recognition for low-resource electronic medical records}, volume = {640}, year = {2025}, }
• (2025). Single image dehazing via fine-scale perception network. NEUCOM, 640, 130356. (www)

  @article{ , abstract = {Single-image dehazing is a classical but challenging problem in the field of computer vision, which can largely improve the overall performance of a machine vision system. Normally, there are two methodologies being used in dehazing, which are the prior-based methodology and the learning-based methodology. The former requires manual setting of thresholds and fails in blank areas such as the sky, while the latter makes it difficult to perceive details in hazy images. To address the above problems, a fine-scale perceptual network (FSPN) for image dehazing is proposed. Specifically, FSPN consists of a fine-scale feature encoder block (FFEB) with a multi-branch structure and a feature upsampling block (FUB), which are used to extract and model the detailed features of haze images and predict the transmission map. In the FFEB module, high-frequency information and local information in the hazing image are extracted through multi-branch convolution and maximum pooling and then fused through a multi-layer perceptron. In the FUB module, the transmission map is predicted in the decoding process through a combination of interpolation and convolution. Finally, the atmospheric illumination is determined by a priori methods and used as part of the parameters of the neural network to achieve dehazing through the atmospheric scattering model. The mainstream quantitative evaluation materials are used to analyze the performance of dehazing. The comprehensive experimental results indicate that the proposed method outperforms other state-of-the-art methods in many cases.}, archive = {J_NEUCOM}, author = {Huajie Wen and Jianhao Shen and Cheng Chi and Lin Lin and Yuanzhi Deng and Liting Chen and Qiaohui Feng and Gang Xu}, doi = {10.1016/j.neucom.2025.130356}, journal = {Neurocomputing}, month = {8}, pages = {130356}, shortjournal = {Neurocomputing}, title = {Single image dehazing via fine-scale perception network}, volume = {640}, year = {2025}, }
• (2025). ABRobOcular: Adversarial benchmarking and robustness analysis of datasets and tools for ocular-based user recognition. NEUCOM, 640, 130352. (www)

  @article{ , abstract = {Ocular biometrics, leveraging the unique traits of the eye region, has gained significant adoption for identity verification and healthcare applications, including integration into products by tech giants like Meta and Apple. Despite their widespread implementation, the vulnerability of ocular biometrics to sophisticated adversarial attacks remains inadequately addressed in the existing literature, and most of the security evaluations focus on traditional physical spoofing rather than digital manipulation. Adversarial attacks involve using digitally altered data to deceive deep learning models into producing incorrect outputs. We investigate white-box and black-box-based adversarial attacks against ocular-based user recognition algorithms, selected based on their prevalence in the literature, computational feasibility, and demonstrated effectiveness in related biometric domains. This research is paramount to address gaps in the adversarial robustness of ocular biometrics in high-security applications. We benchmark the impact of 5 popular white-box adversarial attacks (FGSM, BIM, MIM, PGD, and CW) alongside their counterpart 9 defenses and 4 black-box based attacks (Patch Occlusion, Monocle Blending, Transfer-Based Attacks, and GAN-Based attacks) alongside 4 defense strategies, including our proposed Strip and Patch Augmentations, across multiple datasets. Our experiments reveal that white-box attacks can obtain an average attack success rate of about 40%–50% for the ocular verification system. While, black-box Monocle Blending attacks demonstrate vulnerabilities with success rates ranging from 44.89% to 99.58%. Our proposed unified defenses, particularly Strip Augmentation, show impressive effectiveness with attack success rate reduction of 75–97% against Patch Occlusion attacks and up to 92% against white-box attacks like BIM. To the best of our knowledge, this is the first comprehensive study of its kind. To facilitate further research, we are releasing the ocular image datasets with adversarial attacks, as well as the evaluation code, under the ABRobOcular library, providing a valuable resource for future work in the field. Our code is available at VCBSL/Ocular Biometrics .}, archive = {J_NEUCOM}, author = {Bharath Krishnamurthy and Ajita Rattani}, doi = {10.1016/j.neucom.2025.130352}, journal = {Neurocomputing}, month = {8}, pages = {130352}, shortjournal = {Neurocomputing}, title = {ABRobOcular: Adversarial benchmarking and robustness analysis of datasets and tools for ocular-based user recognition}, volume = {640}, year = {2025}, }
• (2025). MFF-net: A multi-view feature fusion network for generalized forgery image detection. NEUCOM, 640, 130351. (www)

  @article{ , abstract = {With the rapid advancement of artificial intelligence, generative models now produce images nearly indistinguishable from real-world scenarios, challenging the authenticity of media content. Existing forgery detection methods, however, are often model-specific and lack generalization. To address this, we propose a novel detection framework that, for the first time, integrates spatial, frequency, and texture features, leveraging three distinct perspectives simultaneously. The framework is composed of three modules for feature extraction and an adaptive fusion module for integrating multi-view features. Specifically, the multi-level spatial module processes both the macrostructure and microdetails of images through global and local feature branches, enabling the extraction of multi-level spatial features. The dual-frequency module employs two distinct frequency analysis techniques to capture frequency domain features, while the GramTexture module utilizes an improved ResNet18 architecture to extract shallow texture features. A dynamic adaptive feature fusion method, integrating multi-head attention mechanisms with adaptive weight allocation, is employed to effectively combine multi-view features. Extensive comparative experiments demonstrate the proposed framework’s superior performance, achieving exceptional generalization and robustness on 18 diverse test datasets. Notably, using only 10% of the ProGAN training dataset, our method achieves an average improvement of 13.70% in mean average precision (mAP) over baseline methods. These results highlight the framework’s effectiveness in handling previously unseen datasets. The implementation is publicly available at the following GitHub repository link .}, archive = {J_NEUCOM}, author = {Ying Lin and Tenglong Mao and Ziyi Chen and Hong Lu and Zhaoyu Chen and Yan Kang}, doi = {10.1016/j.neucom.2025.130351}, journal = {Neurocomputing}, month = {8}, pages = {130351}, shortjournal = {Neurocomputing}, title = {MFF-net: A multi-view feature fusion network for generalized forgery image detection}, volume = {640}, year = {2025}, }
• (2025). Acceleration of grokking in learning arithmetic operations via Kolmogorov–Arnold representation. NEUCOM, 640, 130347. (www)

  @article{ , abstract = {We propose novel methodologies aimed at accelerating the grokking phenomenon, which refers to the rapid increment of test accuracy after a long period of overfitting as reported by Power et al. (2022). Focusing on the grokking phenomenon that arises in learning arithmetic operations via the transformer model, we begin with a discussion on data augmentation in the case of commutative arithmetic operations. To further accelerate, we elucidate arithmetic operations through the lens of the Kolmogorov–Arnold (KA) representation theorem, revealing its correspondence to the transformer architecture: embedding, decoder block, and classifier. Observing the shared structure between KA representations associated with arithmetic operations, we suggest various transfer learning mechanisms that expedite grokking. This interpretation is substantiated through a series of rigorous experiments. Our study demonstrates that a simple data augmentation makes the grokking phenomenon nearly two times faster, while the transfer approach speeds it up by almost five times. In addition, our approach is successful in learning two nonstandard arithmetic tasks: composition of operations and solving a system of equations. Furthermore, we reveal that the model is capable of learning arithmetic operations using a limited number of tokens under embedding transfer, which is supported by a set of experiments as well.}, archive = {J_NEUCOM}, author = {Yeachan Park and Minseok Kim and Yeoneung Kim}, doi = {10.1016/j.neucom.2025.130347}, journal = {Neurocomputing}, month = {8}, pages = {130347}, shortjournal = {Neurocomputing}, title = {Acceleration of grokking in learning arithmetic operations via Kolmogorov–Arnold representation}, volume = {640}, year = {2025}, }
• (2025). Efficient hybrid linear self-attention based visual object tracking with LoRA. NEUCOM, 640, 130335. (www)

  @article{ , abstract = {The Transformer has had a significant impact on the field of object tracking in recent years. A considerable number of researchers have proposed Transformer based trackers, which have demonstrated remarkable performance. However, the substantial computational burden associated with these Transformer based trackers has resulted in a poor inference speed, with some trackers failing to meet the real-time requirements. To solve this problem, we propose an Efficient Hybrid Linear Self-Attention CNN-Transformer based tracker (ECTTrack) for lightweight tracking. Our ECTTrack employs CNN to extract local features of the target, and utilizes the efficient hybrid linear self-attention to model the correlation between the template and the search region, to ensure that the search region extracts more target-specific features and improves the discriminative ability of the model. We also propose a Low-Rank Adaptation (LoRA) based Feature Adaptation Module (LFAM) to solve the task conflict problem, which can adjust the learned features according to the tasks. It will not increase the inference latency, so as to ensure the inference speed of the whole tracker. We experimentally validate the performance of the ECTTrack on seven challenging datasets, comprising LaSOT, LaSOT e x t , TrackingNet, GOT-10k, TNL2K, UAV123, and NFS. The ECTTrack achieves excellent results, and its inference speed is very fast, up to 147 fps(frames per second) on GPUs and 34fps on CPU. The source code and model are released at: https://github.com/mrlonely426/ECTTrack .}, archive = {J_NEUCOM}, author = {Liang Xu and Zhiqing Guo and Liejun Wang}, doi = {10.1016/j.neucom.2025.130335}, journal = {Neurocomputing}, month = {8}, pages = {130335}, shortjournal = {Neurocomputing}, title = {Efficient hybrid linear self-attention based visual object tracking with LoRA}, volume = {640}, year = {2025}, }
• (2025). A comprehensive study of audio profiling: Methods, applications, challenges, and future directions. NEUCOM, 640, 130334. (www)

  @article{ , abstract = {Audio profiling is at the forefront of a technological breakthrough, offering rich insights into human behavior, emotions, physical attributes, and environmental contexts through detailed analysis of voice data. As we embrace an era where the integration of smart technologies equipped with the ability to capture sound is becoming ubiquitous, the capacity to accurately infer personal traits such as age, gender, height, weight, emotional state, personality, and even environmental contexts through voice analysis opens up vast opportunities across law enforcement, healthcare, social and commercial services, and entertainment. This emerging field promises to enhance our interaction with technology by not only understanding who we are but also by interpreting the world around us. However, the remarkable landscape is fraught with challenges, including data imbalances, the complexity of predictive models, and significant privacy concerns regarding the handling of sensitive paralinguistic information. This survey explores deep into the current landscape of audio profiling, examining the techniques and datasets in use, and showcasing its diverse applications while highlighting the need for advanced methodologies, enriched dataset development, and robust privacy preservation techniques.}, archive = {J_NEUCOM}, author = {Anil Pudasaini and Muna Al-Hawawreh and Mohamed Reda Bouadjenek and Hakim Hacid and Sunil Aryal}, doi = {10.1016/j.neucom.2025.130334}, journal = {Neurocomputing}, month = {8}, pages = {130334}, shortjournal = {Neurocomputing}, title = {A comprehensive study of audio profiling: Methods, applications, challenges, and future directions}, volume = {640}, year = {2025}, }
• (2025). Consistent model selection for estimating functional interactions among stochastic neurons with variable-length memory. NEUCOM, 640, 130333. (www)

  @article{ , abstract = {We address the problem of identifying functional interactions among stochastic neurons with variable-length memory from their spiking activity, where “variable-length memory” implies that the influence of past spikes can extend over time periods whose length itself may change, reflecting adaptive or context-dependent history effects. The neuronal network is modeled by a stochastic system of interacting point processes with variable-length memory, meaning that each neuron’s firing probability depends on its own and other neurons’ historical spikes, with the length of this history not being fixed. Each chain describes the activity of a single neuron, indicating whether it spikes at a given time. One neuron’s influence on another can be either excitatory or inhibitory. To identify the existence and nature of an interaction between a neuron and its postsynaptic counterpart, we propose a model selection procedure based on the observation of the spike activity of a finite set of neurons over a finite time. The proposed procedure is also based on the maximum likelihood estimator for the synaptic weight matrix of the network neuronal model. In this sense, we prove the consistency of the maximum likelihood estimator followed by a proof of the consistency of the neighborhood interaction estimation procedure ensuring that, with enough data, the method accurately recovers both the values of the synaptic weights and the presence or absence of connections. The effectiveness of the proposed model selection procedure is demonstrated using simulated data, which validates the underlying theory showing that, under controlled conditions, the estimated connections match the true simulated network, thereby confirming the accuracy and robustness of the approach. The method is also applied to analyze spike train data recorded from hippocampal neurons in rats during a visual attention task, where a computational model reconstructs the spiking activity and the results reveal interesting and biologically relevant information.}, archive = {J_NEUCOM}, author = {Ricardo F. Ferreira and Matheus E. Pacola and Vitor G. Schiavone and Rodrigo F.O. Pena}, doi = {10.1016/j.neucom.2025.130333}, journal = {Neurocomputing}, month = {8}, pages = {130333}, shortjournal = {Neurocomputing}, title = {Consistent model selection for estimating functional interactions among stochastic neurons with variable-length memory}, volume = {640}, year = {2025}, }
• (2025). Evolution-guided Q-learning for tracking control of unknown dynamic systems. NEUCOM, 640, 130332. (www)

  @article{ , abstract = {In this article, a novel evolution-guided accelerated Q-learning (EGAQL) algorithm is proposed to solve optimal tracking control problems for unknown nonlinear systems. Traditional Q-learning methods in tracking control, which rely on gradient-based strategies, often struggle with limited accuracy in policy improvement and tracking performance. To address these challenges, a novel evolutionary algorithm is applied to Q-learning, enabling dynamic adjustment of search intensity. By integrating evolutionary computation into the policy improvement process, EGAQL eliminates the dependence on gradient information and derives more precise solutions. Furthermore, to improve the accuracy of the reference control policy, evolutionary algorithms are also employed to optimize the model neural network. The algorithm is proven to guarantee convergence and monotonic improvement. Finally, a simulation experiment validates the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Zeqiang Yuan and Ding Wang and Jiangyu Wang and Mingming Zhao and Junfei Qiao}, doi = {10.1016/j.neucom.2025.130332}, journal = {Neurocomputing}, month = {8}, pages = {130332}, shortjournal = {Neurocomputing}, title = {Evolution-guided Q-learning for tracking control of unknown dynamic systems}, volume = {640}, year = {2025}, }
• (2025). Hybrid dwell-time-based h∞ control of switched nonlinear systems and its application to switched neural networks. NEUCOM, 640, 130331. (www)

  @article{ , abstract = {The hybrid dwell-time-based H ∞ control problem of switched nonlinear systems without internal stability requirements is investigated in this paper. Under the hybrid mode-dependent dwell time (HMDT), a set of mode-dependent time-scheduled controllers (TSCs) are proposed here. Meanwhile, switched neural networks (SNNs) as typical switched nonlinear systems are further introduced to refine the theoretical derivation, and time-varying switching-signal-based multiple discontinuous Lyapunov functions (TSMDLFs) are constructed to ensure the exponential stability with the l 2 -gain condition of the SNNs. Finally, examples of practical applications are presented to demonstrate the validity of our theoretical results.}, archive = {J_NEUCOM}, author = {Jinling Wang and Jun-Guo Lu and Jiarong Li and Haijun Jiang and Cheng Hu}, doi = {10.1016/j.neucom.2025.130331}, journal = {Neurocomputing}, month = {8}, pages = {130331}, shortjournal = {Neurocomputing}, title = {Hybrid dwell-time-based h∞ control of switched nonlinear systems and its application to switched neural networks}, volume = {640}, year = {2025}, }
• (2025). Gradient neural network models for approximate tucker decomposition of time-dependent tensors. NEUCOM, 640, 130330. (www)

  @article{ , abstract = {In this paper, we introduce a gradient neural network model for approximating the Tucker decomposition of time-dependent tensors with a specified multilinear rank bound. The behavior of this model is governed by a system of ordinary differential equations. First, based on the Karush–Kuhn–Tucker (KKT) conditions of the optimization problem for the approximate Tucker decomposition of a constant tensor, a gradient neural network model is proposed, demonstrating local stability according to Lyapunov stability theory. Next, we design a gradient neural network model specifically tailored for time-dependent tensors. Finally, using the Taylor-type difference rules, we discretize the proposed neural network models, ensuring zero stability for the implementation of digital hardware. We illustrate the effectiveness and characteristics of our approach through numerical examples.}, archive = {J_NEUCOM}, author = {Maolin Che and Yimin Wei and Hong Yan}, doi = {10.1016/j.neucom.2025.130330}, journal = {Neurocomputing}, month = {8}, pages = {130330}, shortjournal = {Neurocomputing}, title = {Gradient neural network models for approximate tucker decomposition of time-dependent tensors}, volume = {640}, year = {2025}, }
• (2025). ISCDFuse: Interval sampling correlation driven visual state space models for multimodal image fusion. NEUCOM, 640, 130329. (www)

  @article{ , abstract = {Multimodal image fusion aims to retain functional highlights and detailed textures from different modalities. To address the shortcomings of existing methods regarding time complexity and cross-modal global information extraction efficiency, we propose a cross-domain distance learning image fusion framework based on Visual State Space Models (VSSMs) - the Interval Sampling Correlation-Driven Fusion Network (ISCDFuse). ISCDFuse employs a dual-branch feature extractor structure, comprising a Cross-domain Feature Association Encoder (CFAE), a High- and Low-frequency feature Extraction (HLoE) module, and a Vmamba-based Decoder (VD) for feature fusion and image generation. The CFAE traverses two modal space domains using an interval sampling cross-scan module, converting non-causal visual images into ordered patch sequences, thereby enhancing the correlation of global features across different modalities. The low-frequency feature extractors in the HLoE and VD modules both utilize a residual visual mamba structure, incorporating a multi-directional skip scanning approach that samples the image at bi-stride, enhancing deep semantic feature extraction and effectively modeling long-distance spatial dependencies. The high-frequency feature extractor employs the Invertible Neural Networks (INN) block to extract nuanced texture details. Extensive experiments have demonstrated that ISCDFuse delivers excellent fusion performance and fast speed across visible-infrared image fusion and medical image fusion. Notably, in unified benchmark tests, ISCDFuse significantly proves the practical value in downstream multimodal image processing, such as visible-infrared object detection.}, archive = {J_NEUCOM}, author = {Lian Zhang and Lingxue Wang and Yuzhen Wu and Mingkun Chen and Dezhi Zheng and Yi Cai}, doi = {10.1016/j.neucom.2025.130329}, journal = {Neurocomputing}, month = {8}, pages = {130329}, shortjournal = {Neurocomputing}, title = {ISCDFuse: Interval sampling correlation driven visual state space models for multimodal image fusion}, volume = {640}, year = {2025}, }
• (2025). PoseDucer: Implicit relation inducement for invisible keypoint reconstruction in real-world occluded scenes. NEUCOM, 640, 130328. (www)

  @article{ , abstract = {Accurately locating invisible keypoints in complex and occluded scenes is a critical challenge for human pose estimation (HPE). This challenge not only limits the effectiveness of existing HPE methods but also underscores the necessity for more robust and precise methodologies. To address these limitations and significantly enhance the performance and applicability of HPE systems in real-world scenarios, we propose PoseDucer, an innovative two-stage model. In the first stage, PoseDucer extracts rich visual features based on an off-the-shelf backbone network, while the second stage explores reliable spatial structural relationships. Specifically, the second stage consists of two branches. Each branch implicitly models human structural relations from a different perspective. The network is guided to adaptively reconstruct invisible keypoints using high-confidence visible keypoints and part-related structural cues. Extensive experiments on the COCO, CrowdPose, and OCHuman datasets demonstrated that PoseDucer achieved state-of-the-art performance. On the particularly challenging OCHuman dataset, it achieved an AP of 68.3%, significantly surpassing that of the existing methods. Qualitative analysis of heavily occluded examples further confirmed that PoseDucer reconstructed invisible keypoints more logically and accurately than the current methods. The visualization results also revealed the model’s interpretability in correcting the inaccurate initial estimates of these keypoints.}, archive = {J_NEUCOM}, author = {Junneng Feng and Rong Zhang and Yirui Wang and Shangce Gao and Lijun Guo}, doi = {10.1016/j.neucom.2025.130328}, journal = {Neurocomputing}, month = {8}, pages = {130328}, shortjournal = {Neurocomputing}, title = {PoseDucer: Implicit relation inducement for invisible keypoint reconstruction in real-world occluded scenes}, volume = {640}, year = {2025}, }
• (2025). Enhancing tiny object detection using guided object inference slicing (GOIS): An efficient dynamic adaptive framework for fine-tuned and non-fine-tuned deep learning models. NEUCOM, 640, 130327. (www)

  @article{ , abstract = {Tiny Object Detection (TOD) in UAV and standard imaging remains challenging due to extreme scale variations, occlusion, and cluttered backgrounds. This paper presents the Dynamic Adaptive Guided Object Inference Slicing (GOIS) framework, a two-stage adaptive slicing strategy that dynamically reallocates computational resources to Regions of Interest (ROIs) , enhancing detection precision and recall. Unlike static and semi-adaptive slicing methods like SAHI and ASAHI, evaluated with models such as FNet, TOOD, and TPH-YOLO, GOIS leverages VisDrone and xView datasets to optimize hierarchical slicing and dynamic Non-Maximum Suppression (NMS), improving tiny object detection while reducing boundary artifacts and false positives. Comprehensive experiments using MS COCO-pretrained Ultralytics models under fine-tuning and non-fine-tuning conditions validate its effectiveness. Evaluations across YOLO11, RT-DETR-L, YOLOv8s-WorldV2, YOLOv10, YOLOv8, and YOLOv5 demonstrate that GOIS consistently outperforms Full-Image Inference (FI-Det), achieving up to 3-4 × improvements in small-object recall. On the VisDrone2019 dataset, GOIS-Det improved mAP@0.50:0.95 from 0.12 (FI-Det) to 0.33 (+175%) on YOLO11 and from 0.18 to 0.38 (+111.10%) on YOLOv5n. Fine-tuning further enhanced AP-Small by 278.66% and AR-Small by 279.22%, confirming GOIS’s adaptability across diverse deployment scenarios. Additionally, GOIS reduced false positives by 40%–60%, improving real-world detection reliability. Ablation studies validate GOIS’s hierarchical slicing and parameter optimization, with 640-pixel coarse slices and 256-pixel fine slices achieving an optimal balance between accuracy and efficiency. As the first open-source TOD slicing framework on Hugging Face Apps and Google Colab, GOIS delivers real-time inference, open-source code, and live demonstrations, establishing itself as a breakthrough in object detection. The code and results are publicly available at https://github.com/MMUZAMMUL/GOIS with a live demoe at https://youtu.be/ukWUfXBFZ5I .}, archive = {J_NEUCOM}, author = {Muhammad Muzammul and Xuewei Li and Xi Li}, doi = {10.1016/j.neucom.2025.130327}, journal = {Neurocomputing}, month = {8}, pages = {130327}, shortjournal = {Neurocomputing}, title = {Enhancing tiny object detection using guided object inference slicing (GOIS): An efficient dynamic adaptive framework for fine-tuned and non-fine-tuned deep learning models}, volume = {640}, year = {2025}, }
• (2025). PrototypeFormer: Learning to explore prototype relationships for few-shot image classification. NEUCOM, 640, 130326. (www)

  @article{ , abstract = {Few-shot image classification has received considerable attention for overcoming the challenge of limited classification performance with limited samples in novel classes. Most existing works employ sophisticated learning strategies and feature learning modules to alleviate this challenge. In this paper, we propose a novel method called PrototypeFormer, exploring the relationships among category prototypes in the few-shot scenario. Specifically, we utilize a transformer architecture to build a prototype extraction module, aiming to extract class representations that are more discriminative for few-shot classification. Moreover, during the model training process, we propose a contrastive learning-based optimization approach to optimize prototype features in few-shot learning scenarios. Despite its simplicity, our method performs remarkably well, with no bells and whistles. We have experimented with our approach on several popular few-shot image classification benchmark datasets, which shows that our method outperforms most of current state-of-the-art methods. In particular, our method achieves 90.88% and 97.07% on 5-way 1-shot and 5-way 5-shot tasks of miniImageNet, which surpasses the state-of-the-art results with accuracy of 6.84% and 0.57%, respectively.}, archive = {J_NEUCOM}, author = {Meijuan Su and Feihong He and Gang Li and Fanzhang Li}, doi = {10.1016/j.neucom.2025.130326}, journal = {Neurocomputing}, month = {8}, pages = {130326}, shortjournal = {Neurocomputing}, title = {PrototypeFormer: Learning to explore prototype relationships for few-shot image classification}, volume = {640}, year = {2025}, }
• (2025). Semantization of memories in a hippocampal–cortical spiking neural network. NEUCOM, 640, 130323. (www)

  @article{ , abstract = {The human brain consolidates episodic memories into more semantic representations during sleep, enabling the continuous integration of new knowledge. This transformation is supported by hippocampal replays, which triggers a reactivation and reshaping of synaptic connections in the neocortex. In this study, we developed a plastic spiking neural network of Leaky Integrate-and-Fire (LIF) neurons to simulate the interaction between the hippocampus, perceptual neocortical areas, and neocortical regions responsible for processing semantic and contextual information. The model operates in two learning phases: first, the network encodes new experiences as episodic memories; then, during a sleep-like phase, hippocampal reactivation propagates to the neocortex. Crucially, the model leverages the apical mechanisms recently proposed by the experimentally grounded Dendritic Integration Theory (DIT) to modulate the activity of neural assemblies during both learning and replay. We evaluated the model on continual learning tasks, including split and rotational MNIST, and further demonstrated its practical applicability by training and testing it on sensory data from a soft pneumatic gripper in a dynamic robotic environment.}, archive = {J_NEUCOM}, author = {Federico D’Alba and Nilay Kushawaha and Lorenzo Fruzzetti and Pier Stanislao Paolucci and Egidio Falotico}, doi = {10.1016/j.neucom.2025.130323}, journal = {Neurocomputing}, month = {8}, pages = {130323}, shortjournal = {Neurocomputing}, title = {Semantization of memories in a hippocampal–cortical spiking neural network}, volume = {640}, year = {2025}, }
• (2025). DFOSfusion: Infrared and visible image fusion based on multi-scale dynamic feature optimization. NEUCOM, 640, 130321. (www)

  @article{ , abstract = {With the rapid advancement of deep learning technology, significant progress has been made in image fusion, which is crucial for applications like surveillance, remote sensing, and medical imaging. However, most existing methods fail to fully extract and optimize multi-scale features, often lacking fine detail preservation, global consistency, and effective redundancy minimization, thus limiting overall visual quality. In this paper, we introduce DFOSfusion, an innovative image fusion framework specifically designed to address these challenges. DFOSfusion combines Multi-Scale Dense Blocks (MSD) with a Dynamic Feature Optimization System (DFOS), where the MSD module captures multi-scale local details, and the DFOS dynamically enhances global features via cross-layer spatial and channel optimization, effectively minimizing redundancy and promoting feature consistency. An adaptive texture-aware loss function further ensures that texture and structural details are consistently preserved during the fusion process. Extensive experiments reveal that DFOSfusion not only delivers superior visual quality but also consistently outperforms state-of-the-art methods, excelling in feature optimization and image coherence across multiple challenging benchmark datasets.}, archive = {J_NEUCOM}, author = {Mianyao Ruan and Gang Liu and Mengliang Xing and Gang Xiao and Hanlin Xu and Xingfei Zhang}, doi = {10.1016/j.neucom.2025.130321}, journal = {Neurocomputing}, month = {8}, pages = {130321}, shortjournal = {Neurocomputing}, title = {DFOSfusion: Infrared and visible image fusion based on multi-scale dynamic feature optimization}, volume = {640}, year = {2025}, }
• (2025). Local global information aggregation graph convolution for skeleton-based action recognition. NEUCOM, 640, 130320. (www)

  @article{ , abstract = {In the field of human action recognition based on skeletal data, how to capture the complex features of skeletal sequences in both spatial and temporal dimensions is a significant challenge. Existing Graph Convolutional Network (GCN) methods typically focus on extracting spatial features using various approaches, while neglecting the temporal characteristics inherent in human actions. In this paper, we propose the L(ocal) G(lobal)-GCN, a novel approach that emphasizes the under-explored temporal dimension by segmenting time to capture both local and global information. We also introduce a relative position embedding method tailored for skeletal data to enhance the representation capability of graph convolution. Additionally, we propose a partitioning method that reduces computational complexity while retaining all information, leveraging the sequential nature of skeletal data. Our LG-GCN achieves and surpasses the state-of-the-art accuracy across three widely employed datasets: NTU RGB+D with 97.3% (X-View), NTU RGB+D 120 with 91.1% (X-Set), and NW-UCLA with 96.6%.}, archive = {J_NEUCOM}, author = {Shichong Xie and Shengze Li and Peng Chen and Bing Wang and Jun Zhang}, doi = {10.1016/j.neucom.2025.130320}, journal = {Neurocomputing}, month = {8}, pages = {130320}, shortjournal = {Neurocomputing}, title = {Local global information aggregation graph convolution for skeleton-based action recognition}, volume = {640}, year = {2025}, }
• (2025). Hierarchical hybrid prior-knowledge guided cervical cell classification. NEUCOM, 640, 130319. (www)

  @article{ , abstract = {Cervical cancer poses a significant threat to women’s health globally, making early diagnosis and treatment crucial. Pathological diagnosis is considered the gold standard for confirming cervical cancer, with cytological diagnosis, especially the Bethesda System (TBS) grading, being a critical step in cervical cancer screening. However, traditional manual cytological identification methods are not only tedious but also error-prone, particularly when cell category boundaries are unclear and sample volumes are large. To address these challenges, this study introduces a novel cervical cell classification method, namely the Hierarchical Hybrid Prior-Knowledge Guided Classification Model. This model utilizes techniques such as nucleus deep feature sensing, multi-level label classification supervision, cell morphology encoding, and uncertainty-based selective feature fusion, aiming to enhance the accuracy and efficiency of cervical cell classification. Comparative experiments with traditional methods demonstrate significant performance improvements in metrics such as accuracy and F1, validating the effectiveness of the proposed model. Additionally, ablation studies further confirm the contributions of individual modules to the overall performance of the model. This research not only achieves breakthroughs in the field of cervical cell classification but also provides valuable references for other cytological classification tasks.}, archive = {J_NEUCOM}, author = {Fan Yang and Bo Yang and Qiming He and Lili Ji and Yihui Yang and Tian Guan and Hongping Tang}, doi = {10.1016/j.neucom.2025.130319}, journal = {Neurocomputing}, month = {8}, pages = {130319}, shortjournal = {Neurocomputing}, title = {Hierarchical hybrid prior-knowledge guided cervical cell classification}, volume = {640}, year = {2025}, }
• (2025). Improving multi-modal brain tumor segmentation via pre-training and knowledge distillation based post-training. NEUCOM, 640, 130318. (www)

  @article{ , abstract = {Deep learning-based brain tumor segmentation (BTS) models for multi-modal MRI images have seen significant advancements in recent years. However, a common problem in practice is the unavailability of some modalities due to varying scanning protocols and patient conditions, making segmentation from incomplete MRI modalities a challenging issue. Previous methods have attempted to address this by fusing accessible multi-modal features, leveraging attention mechanisms, and synthesizing missing modalities using generative models. However, these methods ignore the intrinsic problems of medical image segmentation, such as the limited availability of training samples, particularly for cases with tumors. Furthermore, these methods require training and deploying a specific model for each subset of missing modalities. To address these issues, we propose a novel approach that enhances the BTS model from two perspectives. Firstly, we introduce a pre-training stage that generates a diverse pre-training dataset covering a wide range of different combinations of tumor shapes and brain anatomy. Secondly, we propose a post-training stage that enables the model to reconstruct missing modalities in the prediction results when only partial modalities are available. To achieve the pre-training stage, we conceptually decouple the MRI image into two parts: ‘anatomy’ and ‘tumor’. We pre-train the BTS model using synthesized data generated from the anatomy and tumor parts across different training samples. For the post-training stage, we introduce a knowledge distillation-based process that enables the model to adapt to partial-modality inputs. This process removes one modality intentionally from the input and encourages the model to produce outputs identical to those when all modalities are present. This allows the model to reconstruct missing information through the post-training process. Extensive experiments demonstrate that our proposed method significantly improves the performance over the baseline and achieves new state-of-the-art results on three brain tumor segmentation datasets: BRATS2020, BRATS2018, and BRATS2015. The code is available at https://github.com/ZhongAobo/Asymmetry-BTS .}, archive = {J_NEUCOM}, author = {Weide Liu and Jingwen Hou and Xiaoyang Zhong and Huijing Zhan and Jun Cheng and Yuming Fang and Guanghui Yue}, doi = {10.1016/j.neucom.2025.130318}, journal = {Neurocomputing}, month = {8}, pages = {130318}, shortjournal = {Neurocomputing}, title = {Improving multi-modal brain tumor segmentation via pre-training and knowledge distillation based post-training}, volume = {640}, year = {2025}, }
• (2025). Multi-view constraint disentangled GAT for recommendation. NEUCOM, 640, 130317. (www)

  @article{ , abstract = {Disentangled Graph Convolutional Networks (disentangled GCNs) can explicitly learn embeddings according to users’ intent, which helps increase both the accuracy and interpretability of recommendations. However, traditional disentangled GCNs primarily rely on first-order direct interaction views, neglecting high-order views and type information, leading to mutual confusion among heterogeneous embeddings and inaccurate representations. Moreover, the aggregation methods used in these models fail to distinguish the varying importance of different nodes and their corresponding intents, which degrades the quality of the learned representations. To address the above problems, we propose the Multi-view Constraint Disentangled Graph Attention Network (MC-DGAT). Specifically, MC-DGAT integrates intent disentanglement and multi-view constraint mechanisms into the graph attention network. The model incorporates high-order interaction views, enabling the capture of multi-hop relationships among homogeneous nodes, which strengthens the model's ability to discern implicit high-order connections and enhances robustness and generalization. Additionally, attention weights are dynamically assigned to reflect the varying importance of different neighboring nodes and intents, leading to more accurate node embeddings and enhanced recommendation accuracy. The experiments on three real-world datasets demonstrate the superiority of the proposed approach. Our code is available at https://github.com/lustrelake/MC-DGAT .}, archive = {J_NEUCOM}, author = {Junwei Du and Yaoze Liu and Haojie Li and Xu Yu and Zhuo Chen and Feng Jiang}, doi = {10.1016/j.neucom.2025.130317}, journal = {Neurocomputing}, month = {8}, pages = {130317}, shortjournal = {Neurocomputing}, title = {Multi-view constraint disentangled GAT for recommendation}, volume = {640}, year = {2025}, }
• (2025). Relevance-aware prompt-tuning method for multimodal social entity and relation extraction. NEUCOM, 640, 130316. (www)

  @article{ , abstract = {Multimodal Named Entity Recognition (MNER) and Multimodal Relation Extraction (MRE) aim to identify specific entities from given text–image pairs and classify the semantic relationships between them, and they have significant applications in social media platform analysis. However, the images and text in social media data are not always aligned, which makes the existing multimodal entity and relation extraction methods still mainly rely on text information. And those mismatched images can even introduce modality noise, leading to negative impacts on the model and preventing them from achieving better performance. To solve this issue, we propose a Relevance-Aware Prompt-tuning (RAP) method with dynamic router mechanism for multi-modal entity and relation extraction. Our method can adaptively learn effective multimodal features from various types of information as prompt vectors and utilize prompt-tuning for entity and relation extraction. Additionally, when integrating information from different modalities, we take into account the intermodal relevance to reduce the negative impact of mismatched visual information on the model, which allows our model to overcome modality noise and achieve better performance. Extensive experiments on three benchmark datasets of tweets demonstrated the effectiveness and superiority of our proposed approach, and achieved approximately 2% increase in F1 values on the three datasets, respectively.}, archive = {J_NEUCOM}, author = {Zhenbin Chen and Zhixin Li and Mingqi Liu and Canlong Zhang and Huifang Ma}, doi = {10.1016/j.neucom.2025.130316}, journal = {Neurocomputing}, month = {8}, pages = {130316}, shortjournal = {Neurocomputing}, title = {Relevance-aware prompt-tuning method for multimodal social entity and relation extraction}, volume = {640}, year = {2025}, }
• (2025). IKFD-depth: Enhancing self-supervised monocular depth estimation with incremental knowledge distillation and feature decoupling self-distillation. NEUCOM, 640, 130314. (www)

  @article{ , abstract = {Monocular depth estimation is a critical research area in the field of environmental perception. Self-supervised monocular depth estimation faces challenges such as structural inconsistencies in depth and blurred boundaries due to the lack of precise supervisory signals. To address these issues, we propose a monocular depth estimation architecture based on incremental knowledge distillation and feature decoupling self-distillation, called IKFD-Depth. First, we introduce an incremental knowledge distillation (IKD) method to alleviate the problem of depth information inconsistency. Incremental knowledge is extracted by mining implicit depth information, which decouples the teacher’s depth map pseudo-labels into explicit depth values and implicit high-dimensional probability distributions that integrate semantic and spatial information. This implicit information serves as a supplementary signal, providing better supervision for the student network to learn depth consistency. Additionally, we propose a Self-Decoupled Enhancement Module (SDEM), which generates reliable structural features as proxy supervision and employs a Feature Decoupling self-Distillation strategy to extract precise semantic and boundary information from the proxy features. This module effectively assists the depth network in recovering complete scene structures, enabling self-learning-based enhancement and mitigating the performance limitations of conventional teacher networks. Extensive experiments demonstrate that IKFD-Depth achieves the best performance among lightweight models on KITTI and Make3D. On the KITTI dataset, the compact IKFD-Depth (2.8M parameters) achieves average improvements of approximately 4% over Lite-Mono (3.1M parameters) in terms of Abs Rel and Sq Rel.}, archive = {J_NEUCOM}, author = {Chaochao Yang and Yuanyao Lu and Yongsheng Qiu and Yuantao Wang}, doi = {10.1016/j.neucom.2025.130314}, journal = {Neurocomputing}, month = {8}, pages = {130314}, shortjournal = {Neurocomputing}, title = {IKFD-depth: Enhancing self-supervised monocular depth estimation with incremental knowledge distillation and feature decoupling self-distillation}, volume = {640}, year = {2025}, }
• (2025). Optimizing information capacity in modular neural networks through excitatory and inhibitory connectivity. NEUCOM, 640, 130313. (www)

  @article{ , abstract = {Modularity is a key principle in complex networks like the brain, enhancing scalability, flexibility, and robustness. This study investigates how modularity affects information capacity in neural networks, focusing on the balance between excitatory and inhibitory connectivity. Using a computational model, we analyzed modular architectures with varying connection probabilities within and between modules, while preserving global excitation–inhibition balance. We assessed information capacity across different dynamical states induced by external inputs. Results show that global long-range excitation leads to periodic states with low information content. In contrast, mixed excitatory and inhibitory inter-module connections produce correlated activity, constraining global information scaling. However, purely inhibitory inter-module connections promote uncorrelated, intermittent activity, maximizing information capacity locally and network-wide. We also explored a Hopfield-like network with global inhibition, where excitatory intra-module connections foster partial correlations within assemblies while preserving inter-assembly independence. These findings underscore the critical role of distinct excitatory and inhibitory connectivity patterns and modular organization in optimizing neural information processing. This study offers insights into the mechanisms driving neural dynamics and their contribution to efficient information processing in the brain, highlighting the interplay between structure and function in complex networks.}, archive = {J_NEUCOM}, author = {Mozhgan Khanjanianpak and Alireza Valizadeh}, doi = {10.1016/j.neucom.2025.130313}, journal = {Neurocomputing}, month = {8}, pages = {130313}, shortjournal = {Neurocomputing}, title = {Optimizing information capacity in modular neural networks through excitatory and inhibitory connectivity}, volume = {640}, year = {2025}, }
• (2025). Pycol: A python package for dataset complexity measures. NEUCOM, 640, 130311. (www)

  @article{ , abstract = {Class overlap presents a significant challenge to machine learning algorithms, especially when class imbalance is present. These factors contribute substantially to the complexity of classification tasks, particularly in real-world scenarios. As a result, measuring overlap is crucial, yet it remains difficult to quantify due to its intricate nature, since it can manifest and be measured in multiple ways. To help mitigate this, recent research has conceptualized a new taxonomy of class overlap measures, divided into multiple families, which allows researchers to obtain a more complete overview of the complexity of the datasets. In line with recent research, we introduce a new Python package for class overlap measurement named pycol. This package implements 29 overlap measures, divided into four overlap families specifically designed to capture class overlap in imbalanced real-world scenarios. This makes pycol an essential tool for researchers dealing with complex classification problems, providing robust solutions to quantify the joint-effect of class overlap and class imbalance effectively.}, archive = {J_NEUCOM}, author = {Diogo Apóstolo and Miriam Seoane Santos and Ana Carolina Lorena and Pedro Henriques Abreu}, doi = {10.1016/j.neucom.2025.130311}, journal = {Neurocomputing}, month = {8}, pages = {130311}, shortjournal = {Neurocomputing}, title = {Pycol: A python package for dataset complexity measures}, volume = {640}, year = {2025}, }
• (2025). SFSWTS: A spatial-frequency shifted windows and time self-attention network for EEG emotion recognition. NEUCOM, 640, 130309. (www)

  @article{ , abstract = {Research into EEG-based emotion recognition has garnered significant interest within the neuroscience and brain-computer interface communities. While previous research has made notable progress in processing EEG data and conducting emotion analysis, effectively utilizing EEG features and integrating information from the time, spatial, and frequency domains of EEG signals for emotion analysis remains a challenging task. This study introduces a novel neural network, the Spatial-frequency Shifted Windows Time Self-attention Neural Network (SFSWTS), which efficiently leverages valuable time, spatial, and frequency domain information present in EEG signals. The SFSWTS network is designed to effectively fuse these diverse domains for enhanced EEG-based emotion analysis. The approach involves feature extraction from EEG signals, encompassing characteristics from the time domain, spatial domain, and frequency domain. The combined spatial and frequency domain information is treated as a 5-channel image, on which a hierarchical vision transformer using shifted windows (Swin Transformer) is applied. The shifted windows strategy enables both local and global attention mechanisms, effectively extracting features from the joint spatial-frequency information. Furthermore, the time slices processed by the hierarchical vision transformer with shifted windows are concatenated with time-domain pairs, and multi-head self-attention is applied. The proposed SFSWTS model was evaluated on two publicly available datasets, SEED and SEED-IV, demonstrating impressive average classification accuracies of 99.46% and 95.41%, respectively, surpassing state-of-the-art methods. The code and models are publicly available at https://github.com/MWZ668/EEGSFSWTS .}, archive = {J_NEUCOM}, author = {Weizhi Ma and Ying Li and Tianhao Li and Haowei Yang and Zhengping Li and Lijun Wang and Junyu Xuan}, doi = {10.1016/j.neucom.2025.130309}, journal = {Neurocomputing}, month = {8}, pages = {130309}, shortjournal = {Neurocomputing}, title = {SFSWTS: A spatial-frequency shifted windows and time self-attention network for EEG emotion recognition}, volume = {640}, year = {2025}, }
• (2025). EDCCN: A benchmark encoder-decoder framework for accurate crowd counting. NEUCOM, 640, 130304. (www)

  @article{ , abstract = {The increasing urban population has led to challenges in managing crowds in public places, especially in preventing tragic incidents like stampedes. Real-time accurate crowd counting (CC) in AI effectively manages crowd dynamics but faces significant obstacles such as background clutter, perspective variations, and occlusion. This study acknowledges the stated challenges in examining the effectiveness of convolutional arrangements by addressing the encoder–decoder crowd counter network (EDCCN). The model supports an integrated feature extraction process (segmented, edge-oriented, and texture), which makes it capable of calculating precise crowd counts in complex, dense situations. Its novel encoder–decoder arrangement explores low- and high-level crowd features in input images to address occlusion and uneven crowd distribution challenges in samples with different backgrounds. The EDCCN proposes two strategies to enhance people estimation accuracy: first, feature propagation guided without density maps to reduce background interference, and second, a complementary attention mechanism for improved information sharing among convolution layers. The EDCCN model harnesses intra- and inter-depth information representation through a non-increasing-order kernel arrangement, achieving state-of-the-art accuracy in people counting compared to existing methods across free (namely Mall, BRT, SmartCity, Indiana) and surveillance (JHU-Crowd, Venice, STech-B) datasets.}, archive = {J_NEUCOM}, author = {Ankit Tomar and Rahul Nijhawan and Deepika Koundal}, doi = {10.1016/j.neucom.2025.130304}, journal = {Neurocomputing}, month = {8}, pages = {130304}, shortjournal = {Neurocomputing}, title = {EDCCN: A benchmark encoder-decoder framework for accurate crowd counting}, volume = {640}, year = {2025}, }
• (2025). AI methods for personality traits recognition: A systematic review. NEUCOM, 640, 130301. (www)

  @article{ , abstract = {Recently, predicting and detecting individuals' personality traits using Artificial Intelligence (AI) has become an important research area. Researchers in these studies aim to predict a person’s personality through text, voice, video, and social media platforms. Unfortunately, due to this research area's interdisciplinary and specialized nature, many young researchers face confusion and uncertainty when selecting appropriate topics and AI tools. Many researchers also lack the expertise to identify suitable Machine Learning (ML) methods. Additionally, the available datasets exhibit multiple and sometimes incompatible characteristics. To date, several studies have addressed the issue of human personality traits using ML methods from different perspectives and disciplines. However, only a limited number of these studies have provided specialized taxonomies for ML. This systematic review examines current research on personality trait detection based on AI methods, specifically analyzing studies related to supervised learning, unsupervised learning, and Deep Learning (DL). Initially, 11 research questions are presented as requirements for the study. Then, through a statistical approach, each question is addressed. These questions include the types of ML algorithms, performance metrics, research objectives/innovations, psychological theories employed, dataset features, models, trending topics, and challenges/limitations. Finally, open issues and future trends are discussed. In this survey, we aim to provide a roadmap for future research on personality trait detection.}, archive = {J_NEUCOM}, author = {Seyed Mostafa Hashemi Motlagh and Mohammad Hossein Rezvani and Mohsen Khounsiavash}, doi = {10.1016/j.neucom.2025.130301}, journal = {Neurocomputing}, month = {8}, pages = {130301}, shortjournal = {Neurocomputing}, title = {AI methods for personality traits recognition: A systematic review}, volume = {640}, year = {2025}, }
• (2025). Class incremental learning with probability dampening and cascaded gated classifier. NEUCOM, 640, 130295. (www)

  @article{ , abstract = {Humans are capable of acquiring new knowledge and transferring learned knowledge into different domains, incurring a small forgetting. The same ability, called Continual Learning, is challenging to achieve when operating with neural networks due to the forgetting affecting past learned tasks when learning new ones. This forgetting can be mitigated by replaying stored samples from past tasks, but a large memory size may be needed for long sequences of tasks; moreover, this could lead to overfitting on saved samples. In this paper, we propose a novel regularisation approach and a novel incremental classifier called, respectively, Margin Dampening and Cascaded Gates. The first combines a constraining loss and a knowledge distillation approach to preserve past learned knowledge while allowing the model to learn new patterns effectively. The latter is a gated incremental classifier, helping the model modify past predictions without directly interfering with them. This is achieved by modifying the output of the model with auxiliary scaling functions. We empirically show that our approach performs well on multiple benchmarks against well-established baselines, and we also study each component of our proposal and how the combinations of such components affect the final results.}, archive = {J_NEUCOM}, author = {Jary Pomponi and Alessio Devoto and Simone Scardapane}, doi = {10.1016/j.neucom.2025.130295}, journal = {Neurocomputing}, month = {8}, pages = {130295}, shortjournal = {Neurocomputing}, title = {Class incremental learning with probability dampening and cascaded gated classifier}, volume = {640}, year = {2025}, }
• (2025). Adaptive neural design for permanent magnet synchronous motor with asymmetric constraints and input dead-zone. NEUCOM, 640, 130294. (www)

  @article{ , abstract = {Permanent magnet synchronous motors (PMSMs) with unknown parameters, stochastic disturbance and input nonlinearity are challenging, especially under safety considerations. This paper develops a novel adaptive Nussbaum design scheme for the PMSMs subject to input dead-zones and unknown control gains. The neural network intelligent control technique is combined with the Nussbaum design to compensate for the system uncertainty and unknown control direction. Also, the input dead-zone is taken into consideration within the Nussbaum design framework. To further consider the asymmetric and time-varying constraints, the barrier Lyapunov function is employed to deal with the safety constraints on the electric angular velocity. Theoretical discussions on the stability of the tracking error for PMSMs is investigated using the adaptive Nussbaum design. Simulation studies verify the presented design schemes.}, archive = {J_NEUCOM}, author = {Tianhao Fei and Yongliang Yang and Chunyu Zheng and Hu Zhang and Guofeng Yuan}, doi = {10.1016/j.neucom.2025.130294}, journal = {Neurocomputing}, month = {8}, pages = {130294}, shortjournal = {Neurocomputing}, title = {Adaptive neural design for permanent magnet synchronous motor with asymmetric constraints and input dead-zone}, volume = {640}, year = {2025}, }
• (2025). Non-fragile asynchronous h∞ estimation for piecewise-homogeneous markovian jumping neural networks with partly available transition rates: A dynamic event-triggered scheme. NEUCOM, 640, 130292. (www)

  @article{ , abstract = {This article tackles the non-fragile asynchronous H ∞ estimation issue for piecewise-homogeneous Markovian jumping neural networks, in which the concerned transition rates are variable during some intervals but invariant during one interval. Besides, due to existence of random factors, in the invariant part, transition rate information is immeasurable, which leads to piecewise-homogeneous Markov process with partly available transition rates. To reduce communication burden and improve resource utilization, an adoptable dynamic event-triggered strategy is constructed, where Zeno phenomenon can be directly avoided. By utilizing stochastic analysis method, Lyapunov stability theory and some matrix inequality techniques, a number of less conservative mode-delay-dependent conditions are addressed to ensure estimation error system realize globally asymptotically mean-square stable with a specified H ∞ disturbance performance index. Moreover, the desired asynchronous estimator matrices can be precisely designed by solving achieved matrix inequalities. Finally, a numerical example with simulations is given to illustrate effectiveness and feasibility of estimation scheme. The special emphasis is that, for the first time, this paper explains that the number of unavailable elements in transition rate matrix has a specific effect on the H ∞ disturbance performance index, which expands on many existing relevant results.}, archive = {J_NEUCOM}, author = {Qiang Li and Kaisheng Zhang and Hanqing Wei and Fanrong Sun and Jinling Wang}, doi = {10.1016/j.neucom.2025.130292}, journal = {Neurocomputing}, month = {8}, pages = {130292}, shortjournal = {Neurocomputing}, title = {Non-fragile asynchronous h∞ estimation for piecewise-homogeneous markovian jumping neural networks with partly available transition rates: A dynamic event-triggered scheme}, volume = {640}, year = {2025}, }
• (2025). WARP-LCA: Efficient convolutional sparse coding with locally competitive algorithm. NEUCOM, 640, 130291. (www)

  @article{ , abstract = {The locally competitive algorithm (LCA) can solve sparse coding problems across a wide range of use cases. Recently, convolution-based LCA approaches have been shown to be highly effective for enhancing robustness for image recognition tasks in vision pipelines. To additionally maximize representational sparsity, LCA with hard-thresholding can be applied. While this combination often yields very good solutions satisfying an ℓ 0 sparsity criterion, it comes with significant drawbacks for practical application: (i) LCA is very inefficient, typically requiring hundreds of optimization cycles for convergence; (ii) the use of a hard-thresholding results in a non-convex loss function, which might lead to suboptimal minima. To address these issues, we propose the Locally Competitive Algorithm with State Warm-up via Predictive Priming (WARP-LCA), which leverages a predictor network to provide a suitable initial guess of the LCA state based on the current input. Our approach significantly improves both convergence speed and the quality of solutions, while maintaining and even enhancing the overall strengths of LCA. We demonstrate that WARP-LCA converges faster by orders of magnitude and reaches better minima compared to conventional LCA. Moreover, the learned representations are more sparse and exhibit superior properties in terms of reconstruction and denoising quality as well as robustness when applied in deep recognition pipelines. Furthermore, we apply WARP-LCA to image denoising tasks, showcasing its robustness and practical effectiveness. Our findings confirm that the naive use of LCA with hard-thresholding results in suboptimal minima, whereas initializing LCA with a predictive guess results in much better outcomes.}, archive = {J_NEUCOM}, author = {Geoffrey Kasenbacher and Felix Ehret and Gerrit Ecke and Sebastian Otte}, doi = {10.1016/j.neucom.2025.130291}, journal = {Neurocomputing}, month = {8}, pages = {130291}, shortjournal = {Neurocomputing}, title = {WARP-LCA: Efficient convolutional sparse coding with locally competitive algorithm}, volume = {640}, year = {2025}, }
• (2025). Sample-customized implicit semantic data augmentation for neural networks regularization. NEUCOM, 640, 130288. (www)

  @article{ , abstract = {Data augmentation techniques have been extensively proven to have the remarkable capability of improving the generalization of deep neural networks via appropriate training data transformations. However, current research mainly focus on automatically learning the augmentation policy in a simply designed search space (e.g., typical pixel space geometric transformations or their combinations) for the whole training set, while overlooking the individual sample variations. As a result, coarse dataset-level augmentation policies with minimal transform operations are learned and applied to the training data, potentially introducing noisy samples and adversely impacting network training. In this paper, we propose a novel sample-level data augmentation method, named “SCAug”, by exploiting the sample-customized implicit semantic transformations as well as the adversarial learning scheme to enable more accurate and diversified data augmentation. Specifically, SCAug introduces a policy network to implicitly explore the proper augmentation policy in deep feature space with the consideration of individual sample features, and the augmentation policy learning procedure acts as one side of the adversarial learning aims to enrich the data diversity semantically and make the machine learning task more challenging. While the adversary, i.e., the task network, is trained to well tolerate these customized augmented samples, thereby becoming more robust and generalizable. Compared with existing methods, SCAug achieves superior performance on typical benchmarks including ImageNet, CIFAR and SVHN.}, archive = {J_NEUCOM}, author = {Xu Kang and Jia Jia and Bing Zhang and Xinzhi Li}, doi = {10.1016/j.neucom.2025.130288}, journal = {Neurocomputing}, month = {8}, pages = {130288}, shortjournal = {Neurocomputing}, title = {Sample-customized implicit semantic data augmentation for neural networks regularization}, volume = {640}, year = {2025}, }
• (2025). Temporal latent variable structural causal model for causal discovery under external interferences. NEUCOM, 640, 130281. (www)

  @article{ , abstract = {Inferring causal relationships from observed data is an important task, yet it becomes challenging when the data is subject to various external interferences. Most of these interferences are the additional effects of external factors on observed variables. Since these external factors are often unknown, we introduce latent variables to represent these unobserved factors that affect the observed data. Specifically, to capture the causal strength and adjacency information, we propose a new temporal latent variable structural causal model, incorporating causal strength and adjacency coefficients that represent the causal relationships between variables. Considering that expert knowledge can provide information about unknown interferences in certain scenarios, we develop a method that facilitates the incorporation of prior knowledge into parameters learning based on Variational Inference, to guide the model estimation. Experimental results demonstrate the stability and accuracy of our proposed method.}, archive = {J_NEUCOM}, author = {Ruichu Cai and Xiaokai Huang and Wei Chen and Zijian Li and Zhifeng Hao}, doi = {10.1016/j.neucom.2025.130281}, journal = {Neurocomputing}, month = {8}, pages = {130281}, shortjournal = {Neurocomputing}, title = {Temporal latent variable structural causal model for causal discovery under external interferences}, volume = {640}, year = {2025}, }
• (2025). ReKon3D: Relation-knowledge aware multi-modal embedding and contrastive GAN for zero-shot 3D recognition. NEUCOM, 640, 130264. (www)

  @article{ , abstract = {Zero-shot learning (ZSL) has emerged as a promising approach to categorizing unseen objects without labeled instances by leveraging external knowledge sources. However, in the domain of 3D object recognition, these sources often contain irrelevant information that hinders accurate object representation. Moreover, generalized zero-shot learning (GZSL) faces persistent challenges with a bias towards seen classes. To address these issues, we propose two main contributions: First, we introduce a novel multimodal embedding framework that enhances the quality of semantic information used in 3D object recognition. This framework aligns visual data from 2D images of 3D objects with graph-based semantic information from knowledge graphs and jointly learns embeddings from different modalities via an attention-based fusion module. These improved alignments and refined knowledge enable us to generate more accurate and discriminative attribute descriptions for each 3D object class. Second, to overcome the bias towards seen classes in GZSL, we propose a semi-supervised contrastive module combined with a generative adversarial network (GAN). This module integrates instance-level contrastive embedding and supervision, enhancing feature-level representations of unseen 3D object classes. Our approach improves generalization in GZSL scenarios, addressing the long-standing challenge of bias towards seen classes. We conduct extensive experiments on four state-of-the-art 3D object recognition datasets and evaluate performance in both inductive ZSL and GZSL tasks. The experimental results demonstrate the effectiveness of our method, achieving a 3% increase in the harmonic mean metric on the ModelNet40 dataset compared to the strongest baseline, 3DGenZ, and significant improvements over other state-of-the-art models.}, archive = {J_NEUCOM}, author = {Md Fokhrul Islam and Sejuti Rahman and Shafin Rahman}, doi = {10.1016/j.neucom.2025.130264}, journal = {Neurocomputing}, month = {8}, pages = {130264}, shortjournal = {Neurocomputing}, title = {ReKon3D: Relation-knowledge aware multi-modal embedding and contrastive GAN for zero-shot 3D recognition}, volume = {640}, year = {2025}, }
• (2025). 4D gaussian splatting for high-fidelity dynamic reconstruction of single-view scenes. NEUCOM, 640, 130262. (www)

  @article{ , abstract = {With the introduction of differentiable rendering frameworks, many reconstruction techniques for dynamic scenes have been proposed. Existing methods have poor reconstruction quality for videos recorded in a single view, including problems such as blurred reconstruction of high-frequency details of objects and inaccurate geometry estimation. Inspired by 3D Gaussian Splatting, we propose a framework SV-Gaussians specifically designed for single-view dynamic scene reconstruction and editing. Our method uses a sparse point cloud of the scene as the initialization of the 3D Gaussian and designs a deformation module to model the change of Gaussian properties over time. We propose a deformation regularization to ensure the consistency of adjacent Gaussian motion. We define color constraints based on obstacle masks, help 3D Gaussians remove obstacles and restore the appearance using contextual information of occluded areas, achieving more realistic object removal than image inpainting methods. Extensive experiments show that our method achieves higher reconstruction quality for single-view scenes than SOTA reconstruction methods and guarantees real-time rendering. More demos and code are available at https://sv-gaussians.github.io .}, archive = {J_NEUCOM}, author = {Xiaodong Yang and Weixing Xie and Sen Peng and Yihang Fu and Wentao Fan and Baorong Yang and Xiao Dong}, doi = {10.1016/j.neucom.2025.130262}, journal = {Neurocomputing}, month = {8}, pages = {130262}, shortjournal = {Neurocomputing}, title = {4D gaussian splatting for high-fidelity dynamic reconstruction of single-view scenes}, volume = {640}, year = {2025}, }
• (2025). A wavelet-enhanced boundary aware network with dynamic fusion for polyp segmentation. NEUCOM, 640, 130259. (www)

  @article{ , abstract = {Accurate segmentation of colonic polyps can help with the timely prevention and treatment of colorectal cancer. However, achieving automatic polyp segmentation in clinical practice remains challenging, mainly because of the morphological similarity between polyps and the surrounding tissues, which makes it difficult for models to effectively capture global structures and edge details of polyps. Additionally, the large variation in polyp scales limits the model’s generalization ability. This paper proposes a Wavelet-Enhanced Boundary Aware Network (WBANet), which aims to efficiently extract multi-scale semantic information of polyps while maintaining low computational overhead. Specifically, we propose a Polarized Multi-Scale Edge Perception (PMEP) module, which extracts the gradient of polyp boundaries from the encoder feature maps and optimizes multi-receptive field perception with global and local semantic information, guiding the model to effectively explore multi-scale polyp contours. Furthermore, we introduce a Wavelet-Guided Feature Enhancement (WFE) module, which improves the clarity of the global polyp structure by combining spatial and frequency domain information and dynamically adjusting feature responses. Finally, we construct a Hierarchical Attention Fusion (HAF) module, which guides the model’s focus to key areas using brightness information and dynamically adjusts the information fusion based on the contributions of adjacent layers, reducing the loss of details during context interaction. Experimental results on five challenging publicly available datasets demonstrate that our proposed WBANet outperforms 16 state-of-the-art methods in terms of learning ability and generalization performance. The source code for the WBANet is available at https://github.com/surpass-15/WBANet .}, archive = {J_NEUCOM}, author = {Yongpeng Wang and Qi Tian and Jinghui Chu and Wei Lu}, doi = {10.1016/j.neucom.2025.130259}, journal = {Neurocomputing}, month = {8}, pages = {130259}, shortjournal = {Neurocomputing}, title = {A wavelet-enhanced boundary aware network with dynamic fusion for polyp segmentation}, volume = {640}, year = {2025}, }
• (2025). CPT-SNN: A spiking neural network that can combine the previous timestep. NEUCOM, 640, 130253. (www)

  @article{ , abstract = {Spiking Neural Networks (SNNs), as brain-inspired neural networks, offer significant advantages over Artificial Neural Networks (ANNs) in terms of energy efficiency and temporal information processing. However, existing research on SNNs largely focuses on improving training strategies or spiking neuron models, with relatively limited exploration of network structure optimization tailored to the unique characteristics of SNNs. Addressing the multi-time-step nature of SNNs, we propose a spiking neural network that can combine the previous timestep (CPT-SNN). This design incorporates outputs from the previous time step into the current layer by converting them into inhibitory currents, which modulate the membrane potential at the current step. Compared to vanilla SNNs, CPT-SNN more effectively extracts information within a limited number of time steps, achieving superior performance with almost no additional energy cost. Furthermore, we introduced the dynamic confidence strategy into directly trained SNNs for the first time. This method dynamically adjusts the number of time steps based on confidence scores calculated from the network’s outputs. When combined with CPT-SNN, it enhances network performance while significantly reducing the required number of time steps. Our approach demonstrated outstanding results on both frame-based and event-based image datasets. For example, it achieved 95.44% accuracy on CIFAR-10 with an average of only 2.72 time steps and 79.68% accuracy on the CIFAR10-DVS dataset with an average of 8.89 time steps. These results highlight the potential of our method in advancing energy-efficient, high-performance spiking neural networks.}, archive = {J_NEUCOM}, author = {Qisheng Xia and Yang Yu and Zheng Chang and Bin Hui and Haibo Luo}, doi = {10.1016/j.neucom.2025.130253}, journal = {Neurocomputing}, month = {8}, pages = {130253}, shortjournal = {Neurocomputing}, title = {CPT-SNN: A spiking neural network that can combine the previous timestep}, volume = {640}, year = {2025}, }
• (2025). Frequency-space enhanced and temporal adaptative RGBT object tracking. NEUCOM, 640, 130240. (www)

  @article{ , abstract = {Recently, superior tracker of ViT in RGBT tracking has been widely applied due to its global and dynamic modeling capabilities. However, these ViT-based trackers often extract information from the template and search images independently, which leads to difficulties in focusing on the discriminative target features and makes them vulnerable to background interference. In addition, they typically neglect to associate the relationship between the current and historical frames, which causes drifting. To address these issues, we propose a RGBT object tracking network based on frequency-space enhancement and temporal adaptation (FETA). We use pre-trained ConvMAE as the feature extraction backbone. Specifically, we design a frequency-space enhancement (FSE) module to establish associations between the template and search images in both the channel and spatial dimensions to extract more discriminative features. FSE enables the tracker to learn features that effectively differentiate the target from the background, thereby mitigating the impact of challenging conditions such as low illumination and background interference on tracking performance. Besides, to adaptatively introduce temporal information, we propose an online score prediction (OSP) module to link the current frame with historical frames. OSP calculates prediction score based on the correlation between the current and historical frames, guiding the FETA to dynamically update the online template. This facilitates effective adaptation to changes in target appearance and background variations. Extensive quantitative and qualitative experiments on three RGBT benchmark datasets RGBT234, LasHeR and VTUAV demonstrate that our method outperforms existing state-of-the-art trackers in terms of both accuracy and robustness.}, archive = {J_NEUCOM}, author = {Shiguo Chen and Linzhi Xu and Xiangyang Li and Chunna Tian}, doi = {10.1016/j.neucom.2025.130240}, journal = {Neurocomputing}, month = {8}, pages = {130240}, shortjournal = {Neurocomputing}, title = {Frequency-space enhanced and temporal adaptative RGBT object tracking}, volume = {640}, year = {2025}, }
• (2025). Multi-feature adaptive-fusion enhanced graph neural network for open-set node classification. NEUCOM, 640, 130238. (www)

  @article{ , abstract = {Semi-supervised open-set node classification is vital in graph learning, aiming to classify seen class nodes while identifying unseen class ones. Existing methods face limitations: they use fixed weight or simple concatenation for multi-view fusion, ignoring node-specific heterogeneity and inter-view relationships, and rely on global thresholds for novelty detection, which overlooks class-wise distribution variations. To address these issues, we introduce a Multi-feature Adaptive-fusion Enhanced graph neural Network (MAEN) for open-set node classification. MAEN comprises two key components: (1) A view-aware adaptive fusion mechanism that dynamically integrates multi-view features using a mixture-of-experts-guided weight generation strategy, effectively capturing node-specific characteristics and nuanced inter-view dependencies. (2) A distribution-aware rejection strategy that constructs class-adaptive decision boundaries by modeling the probability density distributions of seen classes, ensuring precise identification of novel-class instances. Experiments show that MAEN outperforms baseline methods across benchmark datasets, achieving significant improvements in both closed-set and open-set tasks.}, archive = {J_NEUCOM}, author = {Xinxin Liu and Jie Cao and Weiren Yu and Zongxi Li and Yuyao Wang and Huanhuan Gu and Darko B. Vuković}, doi = {10.1016/j.neucom.2025.130238}, journal = {Neurocomputing}, month = {8}, pages = {130238}, shortjournal = {Neurocomputing}, title = {Multi-feature adaptive-fusion enhanced graph neural network for open-set node classification}, volume = {640}, year = {2025}, }
• (2025). Multi-exposure image fusion via distilled 3D LUT grid with editable mode. NEUCOM, 640, 130237. (www)

  @article{ , abstract = {With the rising imaging resolution of handheld devices, existing multi-exposure image fusion algorithms struggle to generate a high dynamic range image with ultra-high resolution in real-time. Furthermore, contemporary applications increasingly demand customizable and adaptable algorithms that can be fine-tuned to specific application scenarios, necessitating both efficiency and flexibility in model design. To tackle these issues, we introduce 3D LUT technology, which can enhance images with ultra-high-definition (UHD) resolution in real time on resource-constrained devices. However, the fusion of information from multiple images with varying exposure levels introduces significant uncertainty, which challenges the generalization capacity of standard 3D LUT grid architectures. To address this fundamental issue and establish a more robust learning paradigm, we propose a teacher–student network architecture specifically designed to model this uncertainty within the 3D LUT grid framework. To address this issue and ensure a robust learning space for the model, we propose using a teacher–student network to model the uncertainty on the 3D LUT grid. Furthermore, we provide an editable mode for the multi-exposure image fusion algorithm by using the implicit representation function to match the requirements in different scenarios. Extensive experiments demonstrate that our proposed method is highly competitive in efficiency and accuracy. The code is released at https://github.com/zzr-idam/UHD-multi-exposure-image-fusion-algorithm .}, archive = {J_NEUCOM}, author = {Xin Su and Zhuoran Zheng and Jialing Yang}, doi = {10.1016/j.neucom.2025.130237}, journal = {Neurocomputing}, month = {8}, pages = {130237}, shortjournal = {Neurocomputing}, title = {Multi-exposure image fusion via distilled 3D LUT grid with editable mode}, volume = {640}, year = {2025}, }
• (2025). Enhancing multi-task performance through associative adversarial learning based on selective attacks. NEUCOM, 640, 130229. (www)

  @article{ , abstract = {Adversarial learning can enhance model robustness against adversarial attacks. However, traditional adversarial learning generates adversarial samples by adding perturbations to entire images, causing excessive differences between adversarial samples and clean samples. To address this issue, we introduce the Associative Adversarial Learning (AAL) method based on a selective attack that can add perturbations to the images according to spatial attention maps, which demonstrate smaller differences from clean samples compared to those generated by global attacks. Thereby, it prevents the degradation of the model performance in clean samples. To further improve the robustness of the model, we develop a backtracking algorithm to effectively learn associative attention, enabling us to decouple attention from adversarial perturbations and mitigate the impact of adversarial attacks on the attention mechanism. Associative attention helps identify the key regions of adversarial examples, thereby enhancing the robustness of the models. We also introduce a projection function for associative attention, which allows AAL to select different attacking regions across various tasks, thus enabling its applicability to a variety of scenarios. Our experimental results show that AAL can be seamlessly integrated into various models, significantly improving adversarial robustness against multiple attacks. For example, AAL can increase InfraRed Small Target Detection mAP on SIRST-v2 by 4.5% and the accuracy of few-shot recognition on miniImageNet by 1.63%.}, archive = {J_NEUCOM}, author = {Zhiming Wang and Yuanglong Yang and Bilang Zhang and Runqi Wang and Liping Jing and Baochang Zhang}, doi = {10.1016/j.neucom.2025.130229}, journal = {Neurocomputing}, month = {8}, pages = {130229}, shortjournal = {Neurocomputing}, title = {Enhancing multi-task performance through associative adversarial learning based on selective attacks}, volume = {640}, year = {2025}, }
• (2025). PFCPNet: A progressive feature correction and prompt network for robust real-world image denoising. NEUCOM, 640, 130221. (www)

  @article{ , abstract = {With the continuous development of deep learning technology, deep learning-based methods have become the mainstream in image denoising due to their superior performance. However, most denoising methods are non-blind, making them ineffective at handling unknown, complex, and diverse degradation in real-world cases. In this case, effective prompt information is needed to guide the denoising process with good generalization for various degradations. Therefore, we design a Progressive Feature Correction and Prompt network (PFCPNet) for robust real-world denoising. Unlike traditional end-to-end networks, we use the degradation feature as a priori and prompt information to prompt PFCPNet for image denoising. By using potential degradation features as the prompt for denoising, the image reconstruction process can avoid the dependence on the degradation estimation accuracy and thus increase the stability and generalization of the method. Specifically, we construct an effective prompt loss constraint to obtain a useful prompt feature. To further enhance the initial prompt feature, we propose a progressive prompt correction structure to aid the denoising sub-network. Subsequently, based on the aforementioned prompt feature information, we design feature domain guided denoising (FDGD) sub-network to tackle diverse real-world noise effectively. Moreover, we consider more comprehensive and better-matched real-world noise modeling to construct the dataset for training the proposed framework. Numerous experiments show that the PFCPNet, using the same weights, achieves excellent denoising results compared to the comparison methods on mainstream datasets, demonstrating its robustness and superiority.}, archive = {J_NEUCOM}, author = {Shouyi Wang and Yizhong Pan and Xiaohai He and Zhengyong Wang and Chao Ren}, doi = {10.1016/j.neucom.2025.130221}, journal = {Neurocomputing}, month = {8}, pages = {130221}, shortjournal = {Neurocomputing}, title = {PFCPNet: A progressive feature correction and prompt network for robust real-world image denoising}, volume = {640}, year = {2025}, }
• (2025). Flexible variational information bottleneck: Achieving diverse compression with a single training. NEUCOM, 640, 130198. (www)

  @article{ , abstract = {Information Bottleneck (IB) is a widely used framework that enables the extraction of information related to a target random variable from a source random variable. In the objective function, IB controls the trade-off between data compression and predictiveness through the Lagrange multiplier β . Traditionally, to find the trade-off to be learned, IB requires a search for β through multiple training cycles, which is computationally expensive. In this study, we introduce Flexible Variational Information Bottleneck (FVIB), an innovative framework for classification task that can obtain optimal models for all values of β with single, computationally efficient training. We theoretically demonstrate that across all values of reasonable β , FVIB can simultaneously maximize an approximation of the objective function for Variational Information Bottleneck (VIB), the conventional IB method. Then we empirically show that FVIB can learn the VIB objective as effectively as VIB. Furthermore, in terms of calibration performance, FVIB outperforms other IB and calibration methods by enabling continuous optimization of β . Our codes are available at https://github.com/sotakudo/fvib .}, archive = {J_NEUCOM}, author = {Sota Kudo and Naoaki Ono and Shigehiko Kanaya and Ming Huang}, doi = {10.1016/j.neucom.2025.130198}, journal = {Neurocomputing}, month = {8}, pages = {130198}, shortjournal = {Neurocomputing}, title = {Flexible variational information bottleneck: Achieving diverse compression with a single training}, volume = {640}, year = {2025}, }
• (2025). ReHyGen: Relational hypergraph enhanced generative aspect sentiment triplet extraction. NEUCOM, 639, 130312. (www)

  @article{ , abstract = {Aspect Sentiment Triplet Extraction (ASTE) has emerged as a pivotal task in sentiment analysis, focusing on extracting the aspect terms along with the corresponding opinion terms and the expressed sentiments. Recently, generative models have achieved significant success in ASTE task. However, existing generative approaches fail to further model the specific relations within the context for ASTE at the encoding phase, making it difficult to establish the nuanced connections between aspect and opinion terms. Additionally, these approaches rely on simple structured templates at the decoding phase to pair aspect terms with opinion terms, which fails to provide effective relation information for the decoding process. To address the aforementioned issues, we propose ReHyGen, a novel relational hypergraph enhanced framework designed to enhance the relational modeling capabilities of generative ASTE models during both the encoding and decoding phases. Specifically, ReHyGen comprises two core components: the Relational Hypergraph Enhanced Module (RHEM) and the Relational Prompt Module (RPM). RHEM leverages the hypergraph attention network and auxiliary relation classification to capture high-order word interactions and boundary-sensitive word pair relations. RPM incorporates relational information into the decoding phase by providing relation-aware prompts, guiding the generation of more accurate target sequences. Extensive experiments on benchmark datasets demonstrate that our proposed framework significantly improve the performance of generative ASTE models.}, archive = {J_NEUCOM}, author = {Zehong Lin and Weibo Chen and Yun Xue and Fenghuan Li}, doi = {10.1016/j.neucom.2025.130312}, journal = {Neurocomputing}, month = {7}, pages = {130312}, shortjournal = {Neurocomputing}, title = {ReHyGen: Relational hypergraph enhanced generative aspect sentiment triplet extraction}, volume = {639}, year = {2025}, }
• (2025). MT-CDGAT: A multi-label diagnosis model for untrained planetary gearbox compound faults based on multi-task cross dynamic graph attention networks. NEUCOM, 639, 130307. (www)

  @article{ , abstract = {Compound fault diagnosis of planetary gearboxes plays a vital role in ensuring the stable operation of mechanical equipment. However, in actual complex industrial environments, obtaining sufficient and labeled multiple components compound fault data is unpractical. Using single fault data from multiple components for compound fault diagnosis is currently a significant challenge. Moreover, compound faults are frequently regarded as a new type of fault in current fault diagnosis research, ignoring the connection between different fault labels. In order to improve the above problems, the multi-task cross dynamic graph attention network model (MT-CDGAT) is proposed in this paper, which can accurately diagnose unknown compound faults of gears and bearings based solely on the single fault data. First, the feature mode decomposition (FMD) method optimized by grey wolf optimization (GWO) is adopted to reduce noise interference and enhance fault features. Then, the multi-task cross-architecture combined with multi-head dynamic graph attention network is designed to update node information from various perspectives, extract, and separate unique node features of single faults. At the same time, the proposed method can also learn the common node features of two subtasks through a shared block. Finally, the output results of two subtasks are spliced to realize multi-label fault diagnosis. The test bench dataset under different operating conditions is utilized to demonstrate the effectiveness and superiority by comparing with multiple methods.}, archive = {J_NEUCOM}, author = {Lixiao Cao and Yixu Wang and Jimeng Li and Zheng Qian and Zong Meng and Miaomiao Liu}, doi = {10.1016/j.neucom.2025.130307}, journal = {Neurocomputing}, month = {7}, pages = {130307}, shortjournal = {Neurocomputing}, title = {MT-CDGAT: A multi-label diagnosis model for untrained planetary gearbox compound faults based on multi-task cross dynamic graph attention networks}, volume = {639}, year = {2025}, }
• (2025). Efficient adaptive label refinement for label noise learning. NEUCOM, 639, 130305. (www)

  @article{ , abstract = {Deep neural networks are highly susceptible to overfitting noisy labels, which leads to degraded performance. Existing methods address this issue by employing complex manually designed strategies, aiming to achieve optimal partitioning in each iteration to avoid fitting noisy labels while thoroughly learning clean samples. However, this often results in models that are overly sensitive to hyperparameters and challenging to train. To address this issue, we decouple the tasks of avoiding fitting incorrect labels and thoroughly learning clean samples, proposing a simple yet highly applicable method called Adaptive Label Refinement (ALR). First, inspired by label refurbishment techniques, we update the original hard labels to soft labels using the model’s predictions to reduce the risk of fitting incorrect labels. Then, by introducing entropy loss, the model adaptively ’harden’ the high-confidence soft labels based on the confidence differences, guiding the model to focus more on learning from clean samples. This approach is simple and efficient, requiring no auxiliary datasets and minimal hyperparameter tuning, making it more adaptable than existing methods that require careful adjustments for different datasets. We validate ALR’s effectiveness through experiments on benchmark datasets with artificial label noise (CIFAR-10/100) and real-world datasets with inherent noise (ANIMAL-10N, Clothing1M, WebVision). The results show that ALR outperforms state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Wenzhen Zhang and Debo Cheng and Guangquan Lu and Bo Zhou and Jiaye Li and Shichao Zhang}, doi = {10.1016/j.neucom.2025.130305}, journal = {Neurocomputing}, month = {7}, pages = {130305}, shortjournal = {Neurocomputing}, title = {Efficient adaptive label refinement for label noise learning}, volume = {639}, year = {2025}, }
• (2025). Adaptive spiking convolutional network for decoding motor imagery patterns. NEUCOM, 639, 130303. (www)

  @article{ , abstract = {Motor imagery (MI) decoding is crucial for the advancement of brain-computer interface (BCI) technologies. However, existing models often suffer from susceptibility to noise and lack biological interpretability. In this study, we introduce the Adaptive Firing Threshold Spiking Convolutional Network (AFTSC-Net), which enhances the biological relevance of spiking neural networks (SNNs) by integrating an adaptive firing rate mechanism with the spatial feature extraction capabilities of convolutional neural networks (CNNs). Additionally, we refined surrogate gradient functions through enhanced spiking neuron mechanisms, significantly reducing computational power consumption while improving the accuracy of MI pattern recognition from electroencephalography (EEG) signals. To validate the efficacy of AFTSC-Net, we conducted experiments with 36 elite athletes from soccer, basketball, and table tennis, performing a comprehensive analysis of neural activity across various motor imagery tasks. The model not only demonstrated superior performance on the athlete dataset but also achieved the state-of-the-art results on public benchmark datasets, surpassing existing methods in terms of accuracy and computational efficiency. These findings highlight the potential of biologically inspired neural networks to enhance MI decoding accuracy and robustness, setting a new standard for real-time BCI applications.}, archive = {J_NEUCOM}, author = {Zimeng Zhu and Fei Liu and Siqi Mu and Kuan Tao}, doi = {10.1016/j.neucom.2025.130303}, journal = {Neurocomputing}, month = {7}, pages = {130303}, shortjournal = {Neurocomputing}, title = {Adaptive spiking convolutional network for decoding motor imagery patterns}, volume = {639}, year = {2025}, }
• (2025). Feature-level attention network with group-aware interest modeling for sequential recommendation. NEUCOM, 639, 130302. (www)

  @article{ , abstract = {Sequential recommendation focuses on modeling user preferences based on their historical interaction sequences to predict future behaviors with greater precision. Incorporating feature-level information beyond item IDs has become a crucial approach to improving the performance of the recommendation system. However, existing methods overlook the hierarchical group relationships among users. This limitation prevents these methods from fully capturing user preferences, leading to an incomplete understanding of their true interests. Meanwhile, effectively leveraging multi-source information in recommendation systems remains a significant challenge. Existing methods typically rely on simple techniques such as pooling or concatenation to integrate information from different sources, which could degrade overall performance. To address these limitations, we propose a novel approach: F eature-level A ttention N etwork with G roup-aware I nterest M odeling for Sequential Recommendation (FANGIM). Specifically, we first employ two distinct encoders to generate user embeddings at different level. Next, we introduce a group clustering module, which identifies potential interest groups at multiple granularities and derives user group interest embeddings for both item and feature level interactions. Furthermore, we design a multi-source representation fusion module that effectively integrates information from diverse sources, reducing the semantic gap between different representation spaces. Additionally, we incorporate contrastive learning within this module to ensure consistency between the different levels of representations. Finally, extensive experiments demonstrate that FANGIM outperforms state-of-the-art baselines across four datasets.}, archive = {J_NEUCOM}, author = {Wei Jiang and Yongquan Fan and Yajun Du and Xianyong Li and Xiaomin Wang}, doi = {10.1016/j.neucom.2025.130302}, journal = {Neurocomputing}, month = {7}, pages = {130302}, shortjournal = {Neurocomputing}, title = {Feature-level attention network with group-aware interest modeling for sequential recommendation}, volume = {639}, year = {2025}, }
• (2025). Fixed-time privacy-preserving synchronization and energy consumption prediction for stochastic neural networks. NEUCOM, 639, 130300. (www)

  @article{ , abstract = {Sensitive information is often at risk of leakage during synchronization and communication in stochastic neural networks. To address this problem, this paper introduces a fixed-time privacy-preserving approach for ensuring leader-follower synchronization and predicting energy consumption in stochastic neural networks. First, an output mask function strategy is proposed to mask the true state of the controlled network and remain indiscernible to other followers throughout the information exchange. Unlike traditional privacy protection techniques, this strategy supports real-time monitoring of the system’s status. Second, by using the fixed-time stability criterion, the synchronization criterion for leader-follower neural networks are obtained, which helps to calculate the system’s working time. Third, the energy consumption of the system is predicted with privacy-preserving, which optimizes energy consumption strategy and improve efficiency of the system. Finally, the effectiveness of this method was demonstrated through simulation and comparative analysis.}, archive = {J_NEUCOM}, author = {Qinjie Jiang and Dongbing Tong and Fei Hu}, doi = {10.1016/j.neucom.2025.130300}, journal = {Neurocomputing}, month = {7}, pages = {130300}, shortjournal = {Neurocomputing}, title = {Fixed-time privacy-preserving synchronization and energy consumption prediction for stochastic neural networks}, volume = {639}, year = {2025}, }
• (2025). Invisible trigger image: A dynamic neural backdoor attack based on hidden feature. NEUCOM, 639, 130296. (www)

  @article{ , abstract = {Neural backdoor is one of the most significant supply chain threat faced by deep neural networks (DNNs). A common attack method for implanting backdoors in DNNs involves injecting poisoned training data that contains triggers. Triggers are commonly categorized into pixel-level and feature-level triggers. Feature-level triggers typically offer higher stealthiness and robustness compared to pixel-level triggers, but may sometimes be less effective. Additionally, the current generation of feature-level triggers requires unrealistic capabilities from the attacker, such as to control the entire training process or directly manipulate the victim model. To address above challenges, we propose a novel dynamic neural backdoor attack using hidden features. We design a trigger generator with a multi-feature constraint algorithm to overlay hidden features onto clean images in an ’invisible’ manner. The trigger generated by our method is dynamic, and the attacker does not need to control the entire training process or directly manipulate the victim model. Experiments show our method outperforms six existing attacks, achieving nearly 100% success with a low poisoning rate of 0.01 while maintaining high accuracy on clean samples. Additionally, it bypasses existing representative backdoor defenses, including three input-level and two model-level defenses.}, archive = {J_NEUCOM}, author = {Xiong Chen and Mohan Li and Yanbin Sun and Zhihong Tian}, doi = {10.1016/j.neucom.2025.130296}, journal = {Neurocomputing}, month = {7}, pages = {130296}, shortjournal = {Neurocomputing}, title = {Invisible trigger image: A dynamic neural backdoor attack based on hidden feature}, volume = {639}, year = {2025}, }
• (2025). Pick and mix reliable pseudo labels for scribble-supervised medical image segmentation. NEUCOM, 639, 130293. (www)

  @article{ , abstract = {Scribble-supervised segmentation methods have attracted significant attention in the field of medical imaging because of their potential to alleviate the data annotation burden. However, these methods often underperform due to a lack of sufficient supervision. Various methods have attempted to enrich the supervisory signals in different ways, including mixing pseudo labels from different samples (referred as Mixup-based method). However, these methods primarily focus on the quantity of enriched supervisory signals, disregarding their quality. This oversight presents a major drawback in that low-quality signals are often contaminated with the noise, thus can lead to undermine performance. Therefore, it is crucial to not only introduce diverse supervisory signals but also ensure their quality and reliability. Motivated by this understanding, we propose a new framework named Pick & Mix, which builds upon the Mixup-based method. In the first step, we leverage the consistency of intra-class features to assess the reliability of pseudo-labels. To enhance the quality of pseudo labels, we assign lower weights to those unreliable pseudo-labels to mitigate the noise effect in the training process. Furthermore, we utilize a threshold to pick reliable pseudo-labels based on their reliability score. In the second step, we mix the reliable pseudo-labels from various samples and generate corresponding mixed images to provide richer supervisory signals for model training. In this manner, we enhance the quality of supervisory signals by generating and picking reliable ones, as well as enrich the quantity of these signals through a process of mixing. Finally, we evaluated our framework on three publicly available datasets: ACDC, MSCMRseg, and BraTS2020. The experimental results demonstrate that our approach achieves state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Jiawei Su and Zhiming Luo and Dazhen Lin and Lihui Lin and Shaozi Li}, doi = {10.1016/j.neucom.2025.130293}, journal = {Neurocomputing}, month = {7}, pages = {130293}, shortjournal = {Neurocomputing}, title = {Pick and mix reliable pseudo labels for scribble-supervised medical image segmentation}, volume = {639}, year = {2025}, }
• (2025). Dynamic anchor-based deep multi-view subspace clustering network. NEUCOM, 639, 130290. (www)

  @article{ , abstract = {Deep multi-view subspace clustering methods have attracted increasing attention due to their excellent performance. However, existing methods based on self-representation property are limited in analyzing large-scale data because of the high time and space complexity. To address this problem, this paper proposes a dynamic anchor-based deep multi-view subspace clustering network (DADMVSC) by exploring the representation relationships between latent and anchor features of multi-view data. DADMVSC integrates view-specific latent feature extraction, dynamic anchor learning, and weighted anchor-sample representation relationship learning into a unified end-to-end network. Specifically, a shared dynamic anchor learning network cooperatively working with the deep auto-encoder is designed to extract the view-specific latent features and anchor features autonomously for multiple views. Based on the anchor and latent features, the fully connected layers are inserted between the encoder and dynamic anchor learning network to learn the anchor-sample representation matrices of multi-view data. We construct the anchor-sample representation tensor and employ a channel attention network to model their higher-order relations as discriminative weights for view-specific representation matrices. Subsequently, the weighted anchor-sample representation matrices are fused by concatenation operation to obtain the multi-view consensus representation. Finally, fast spectral clustering is performed on the consensus representation to achieve efficient clustering. The effectiveness of DADMVSC is demonstrated on large-scale multi-view datasets, and the efficiency is verified by algorithm complexity theory analysis and running time comparison experiments. DADMVSC provides a simple and scalable but effective network framework for deep multi-view subspace clustering analysis on large-scale data.}, archive = {J_NEUCOM}, author = {Jiao Wang and Fei Zhou and Zhenwen Ren}, doi = {10.1016/j.neucom.2025.130290}, journal = {Neurocomputing}, month = {7}, pages = {130290}, shortjournal = {Neurocomputing}, title = {Dynamic anchor-based deep multi-view subspace clustering network}, volume = {639}, year = {2025}, }
• (2025). SwinLip: An efficient visual speech encoder for lip reading using swin transformer. NEUCOM, 639, 130289. (www)

  @article{ , abstract = {This paper presents an efficient visual speech encoder for lip reading. While most recent lip reading studies have been based on the ResNet architecture and have achieved significant success, they are not sufficiently suitable for efficiently capturing lip reading features due to high computational complexity in modeling spatio-temporal information. Additionally, using a complex visual model not only increases the complexity of lip reading models but also induces delays in the overall network for multi-modal studies (e.g., audio–visual speech recognition, speech enhancement, and speech separation). To overcome the limitations of Convolutional Neural Network (CNN)-based models, we apply the hierarchical structure and window self-attention of the Swin Transformer to lip reading. We configure a new lightweight scale of the Swin Transformer suitable for processing lip reading data and present the SwinLip visual speech encoder, which efficiently reduces computational load by integrating modified Convolution-augmented Transformer (Conformer) temporal embeddings with conventional spatial embeddings in the hierarchical structure. Through extensive experiments, we have validated that our SwinLip successfully improves the performance and inference speed of the lip reading network when applied to various backbones for word and sentence recognition, reducing computational load. In particular, our SwinLip demonstrated robust performance in both English LRW and Mandarin LRW-1000 datasets and achieved state-of-the-art performance on the Mandarin LRW-1000 dataset with less computation compared to the existing state-of-the-art model.}, archive = {J_NEUCOM}, author = {Young-Hu Park and Rae-Hong Park and Hyung-Min Park}, doi = {10.1016/j.neucom.2025.130289}, journal = {Neurocomputing}, month = {7}, pages = {130289}, shortjournal = {Neurocomputing}, title = {SwinLip: An efficient visual speech encoder for lip reading using swin transformer}, volume = {639}, year = {2025}, }
• (2025). Improving generalization capability of object detectors via efficient global attention. NEUCOM, 639, 130287. (www)

  @article{ , abstract = {Real-world applications are highly variable, and object detectors must handle various environmental changes, such as lighting, weather conditions, and viewpoint differences. Therefore, these models need strong generalization capability. According to Gestalt theory, human-level generalization emerges from global perception abilities. In this paper, we aim to enhance the generalization capability of YOLO series detectors by introducing global self-attention mechanisms. Specifically, to address the quadratic complexity issue of the vanilla self-attention module, we propose a new efficient attention module. Additionally, to enhance the model’s feature extraction capability, we introduce a new local feature extraction module. Building on these two components, we construct a new backbone network that can be seamlessly integrated into the YOLO series detectors. Thanks to the global perception capability of the efficient global attention module and the strong learning ability of the local feature extraction module, our method achieved a +1.8 AP improvement on YOLOv8 and a +1.0 AP improvement on YOLOv10 in experiments on the COCO dataset, demonstrating the superior generalization capability. Ablation studies on the Pascal VOC dataset further validated the effectiveness and efficiency of each design component. Finally, experiments on the Diverse-Weather dataset confirmed the robust generalization capability of our method, which significantly outperforms state-of-the-art methods. The source code will be available at https://github.com/chenjie04/SGD-PyTorch .}, archive = {J_NEUCOM}, author = {Jie Chen and Meng Joo Er}, doi = {10.1016/j.neucom.2025.130287}, journal = {Neurocomputing}, month = {7}, pages = {130287}, shortjournal = {Neurocomputing}, title = {Improving generalization capability of object detectors via efficient global attention}, volume = {639}, year = {2025}, }
• (2025). Enhance the old representations’ adaptability dynamically for exemplar-free continual learning. NEUCOM, 639, 130286. (www)

  @article{ , abstract = {In continual learning, new data often falls outside the distribution of previous data. Since the old model is trained solely on past tasks and has not encountered the new data, its learned representations lack the adaptability needed to accommodate these new inputs. This mismatch creates a conflict between the learned representations and the concepts of new classes, resulting in a significant representation shift during model updates and exacerbating the issue of catastrophic forgetting. In this work, we presents a two-stage training framework to e nhance the learned r epresentations’ a daptability dynamically for e xemplar- f ree c ontinual l earning (EFCL), dubbed ERA-EFCL. The first stage involves synthesizing data for the old classes using DeepInversion and combining them with the new data to train an expanded module and a feature fusion network (FFN). The expanded module identifies critical features overlooked by the old model, while the FFN integrates the old representations with these newly discovered features. By incorporating additional information that distinguishes between the old and new classes, the FFN generates fused representations enriched with more transferable features for the old classes. This helps align the old representations with the new concepts, enhancing their adaptability and reducing the representation shift. Next, we introduce global task-wise knowledge distillation to ensure balanced knowledge transfer alongside other losses, thereby improving representation learning during the training of the new model. Furthermore, the new classifier is refined using a class-balanced training strategy. Extensive experiments demonstrate that ERA-EFCL achieves favorable results across four benchmark datasets. The code is available at: “ https://github.com/CSTiger77/ExemplarFreeCL ”.}, archive = {J_NEUCOM}, author = {Kunchi Li and Chaoyue Ding and Jun Wan and Shan Yu}, doi = {10.1016/j.neucom.2025.130286}, journal = {Neurocomputing}, month = {7}, pages = {130286}, shortjournal = {Neurocomputing}, title = {Enhance the old representations’ adaptability dynamically for exemplar-free continual learning}, volume = {639}, year = {2025}, }
• (2025). Radar working mode amplitude waveform augmentation based on generative adversarial networks. NEUCOM, 639, 130285. (www)

  @article{ , abstract = {In radar working mode recognition, one of the most widely utilized data formats is working mode data in the form of amplitude waveform sequences. However, the capacity to train big, sophisticated deep learning models is somewhat limited because this type of data involves military secrets and is costly to gather and annotate. This research suggests a generative adversarial network-based radar working mode amplitude waveform augmentation technique to address this issue. Firstly, the technique aims to amplify the radar working mode data set by artificially synthesizing limited amplitude waveform data using generative adversarial networks (GANs) and their variations: Wasserstein GANs (WGANs), Auto-encoders GANs (AE-GANs), and Variational Auto-encoders GANs (VAE-GANs). Secondly, this research suggests a type of Cross-domain Generative Adversarial Networks (CD-GANs) that use amplitude and waveform sequence data to generate each other by utilizing the possible mapping link between amplitude and waveform. To achieve data augmentation, the network can be trained on pre-existing datasets to produce class-specific synthetic data of any length, extending or substituting real amplitude waveform sequences of the same mode of working. Then, to assess the quality of the synthetic data produced by the aforementioned model, the qualitative visual assessment of the synthetic data was conducted using the visualization evaluation metrics of principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), while the quantitative evaluation metrics of wavelet coherence scores were used to compare the similarity between the synthetic and real data. Finally, by training the classifier utilizing a hybrid strategy of synthetic and real data using a standard Convolutional Neural Networks (CNNs) classification model, the effectiveness of synthetic data in the training phase of the radar working pattern recognition model is confirmed. In summary, the simulation experiment confirms the superiority of the CD-GANs model and shows that its generated data is more accurate than that of other GANs models. It can also achieve more realistic and trustworthy augmentation of radar amplitude waveform data.}, archive = {J_NEUCOM}, author = {Wenbo Li and Yang-Yang Dong and Chunxi Dong}, doi = {10.1016/j.neucom.2025.130285}, journal = {Neurocomputing}, month = {7}, pages = {130285}, shortjournal = {Neurocomputing}, title = {Radar working mode amplitude waveform augmentation based on generative adversarial networks}, volume = {639}, year = {2025}, }
• (2025). Multi-strategy ensemble binary RIME optimization for feature selection. NEUCOM, 639, 130284. (www)

  @article{ , abstract = {Feature selection is a critical process in machine learning and data mining that selects a subset of relevant features for model construction. This process helps reduce the dimensionality of the data, minimize computational cost, improve model performance, and enhance interpretability. An effective feature selection method can significantly impact the success of predictive modeling tasks. Recently, numerous algorithms have been developed to address the challenges associated with high-dimensional datasets. These algorithms range from traditional statistical methods to advanced metaheuristic methods. The RIME approach, enhanced through reinforcement learning, presents a viable answer to the optimization challenge. Through reinforcement learning that switches between the dispersed foraging and hard-RIME puncture mechanisms and comprehensive learning strategy, the multi-strategy ensembled RIME method, named QCLFRIME, reaches a more effective balance between local search and global exploration. This algorithm is validated on the IEEE CEC 2017 function test and compared with traditional and advanced algorithms. Subsequently, we develop a wrapper feature selection model based on the binary QCLFRIME model. Besides, the binary QCLFRIME-KNN classifier has excellent performance for the feature selection regarding fitness, error rate, the number of selected features and computational time on 14 publicly available high-dimensional datasets compared to the traditional and enhanced meta-heuristic algorithms. Relevant experimental findings highlight the proposed algorithm's outstanding efficacy, which surpasses most existing algorithms. Therefore, the proposed method is a valuable wrapper-mode feature selection tool.}, archive = {J_NEUCOM}, author = {Sudan Yu and Huiling Chen and Ali Asghar Heidari and Guoxi Liang and Yi Chen and Zhiqing Chen and Xiaoxia Jin}, doi = {10.1016/j.neucom.2025.130284}, journal = {Neurocomputing}, month = {7}, pages = {130284}, shortjournal = {Neurocomputing}, title = {Multi-strategy ensemble binary RIME optimization for feature selection}, volume = {639}, year = {2025}, }
• (2025). Prompt enhanced neural machine translation with POS tags. NEUCOM, 639, 130283. (www)

  @article{ , abstract = {In recent years, neural machine translation (NMT) has achieved significant breakthroughs due to the widespread deployment of large language models (LLMs) and corresponding prompt methods. Though powerful and efficient, the complexities and enormous computational costs for training scaled-up models have caused severe inconvenience. To this end, we propose a simple but efficient approach named Automated Prompt NMT (AP-NMT), which contains an automatic prompt construction module, and further utilizes generated prompts to incorporate syntactic information in training portable-sized Transformer-based translation model. In this way, our proposed method enables the NMT model to obtain comprehensive knowledge of the target language pattern without model structure expansion. Extensive experiments on low-resource IWSLT datasets demonstrate the effectiveness of AP-NMT in improving translation accuracy. We further provide a comprehensive analysis of our method’s influence factors and detailed benefits.}, archive = {J_NEUCOM}, author = {Zhiying Mu and Shengchuan Lin and Sensen Guo and Shanqing Yu and Dehong Gao}, doi = {10.1016/j.neucom.2025.130283}, journal = {Neurocomputing}, month = {7}, pages = {130283}, shortjournal = {Neurocomputing}, title = {Prompt enhanced neural machine translation with POS tags}, volume = {639}, year = {2025}, }
• (2025). Fast blind image deblurring via patch-wise maximum content-weighted prior. NEUCOM, 639, 130267. (www)

  @article{ , abstract = {Blind image deblurring aims to derive the kernel and corresponding clear version solely from blurred images. This paper introduces an innovative blind image deblurring method based on the patch-wise maximum content-weighted prior ( PMCW ). Our work originates from the intuitive observation that the maximum content-weighted value of non-overlapping patches will significantly decrease after blurring degradation, which we demonstrate both mathematically and empirically. Building upon this observation, we propose a novel blind deblurring model combining L 0 -regularized PMCW and L 0 -regularized gradient prior, and develop an efficient solution algorithm utilizing projected alternating minimization (PAM). Qualitative and quantitative evaluation results on multiple benchmark datasets indicate that our proposed model achieves optimal performance, surpassing state-of-the-art algorithms in solving efficiency and various quantitative metrics.}, archive = {J_NEUCOM}, author = {Zheng Guo and Wei Yan and Zirui Zhang and Zhixiang Wu and Zhenhua Xu and Chunyong Wang and Jiancheng Lai and Zhenhua Li}, doi = {10.1016/j.neucom.2025.130267}, journal = {Neurocomputing}, month = {7}, pages = {130267}, shortjournal = {Neurocomputing}, title = {Fast blind image deblurring via patch-wise maximum content-weighted prior}, volume = {639}, year = {2025}, }
• (2025). A fast solution method for the dynamic flexible pickup and delivery problem with task allocation fairness for multiple vehicles. NEUCOM, 639, 130266. (www)

  @article{ , abstract = {The Dynamic Flexible Pickup and Delivery Problem (DFPDP) originates from the actual needs of multi-warehouse management strategies and is one of the important challenges currently facing the field of logistics and distribution. In DFPDP, it is necessary to address dynamic order fluctuations, quickly plan heterogeneous fleet routes, ensure fairness in task allocation, and minimize total travel time under time window constraints. However, there is currently little research on this issue, and traditional heuristic algorithms make it difficult to quickly find a solution to this problem. First, we propose a Multimodal Constraint Dynamic Scheduling Mechanism (MCDSM) to select a vehicle with the lowest current time consumption to make task allocation between vehicles as fair as possible. Second, we propose a Parallel Encoder-Serial Decoder model integrating Variable-length Sequences (PESDVS), in which the variable-length sequences designed can effectively handle the generation of dynamic orders and the changes in the number of pickup and delivery locations, while the trained model can adapt itself to different order scenarios. In addition, the model improves the quality of order decisions through a parallel encoder and serial decoder structure to minimize the total traveling time of the fleet. Extensive experimental results demonstrate that our method has excellent performance and good generalization ability under different order sizes. At the same time, compared with heuristic algorithms, our method can quickly find a feasible solution to the problem and the task allocation between vehicles is relatively fair.}, archive = {J_NEUCOM}, author = {Zhihui Sun and Ran Tian and Jiarui Wu and Xin Lu and Jinshi Wang}, doi = {10.1016/j.neucom.2025.130266}, journal = {Neurocomputing}, month = {7}, pages = {130266}, shortjournal = {Neurocomputing}, title = {A fast solution method for the dynamic flexible pickup and delivery problem with task allocation fairness for multiple vehicles}, volume = {639}, year = {2025}, }
• (2025). Uncertainty and diversity-based active learning for UAV tracking. NEUCOM, 639, 130265. (www)

  @article{ , abstract = {Unmanned aerial vehicles (UAVs) are increasingly utilized in target tracking scenarios due to their compact size and agile movement. With the rapid advancement of artificial intelligence, an increasing number of deep learning algorithms, particularly Transformer-based trackers, are being employed in UAV tracking. However, these algorithms typically have substantial data requirements. This paper investigates the integration of active learning and UAV tracking to mitigate the dataset demands of these models. We propose an active learning framework tailored for UAV target tracking scenarios. Our method, called the Uncertainty and Diversity-based Active Learning for UAV Tracking (UDALT), aims to develop a high-performance tracking model by selecting the most informative samples based on video-level uncertainty and diversity. Specifically, for the uncertainty of the tracked object, we introduce a new entropy-based evaluation formula to assess unlabeled samples and identify more challenging ones. For diversity, we first represent object types by leveraging intermediate features from the model, then apply the K-means clustering algorithm to determine the cluster centers of known object types. By calculating the distances of unlabeled samples from these centers, we ensure a balanced distribution of object types when selecting new samples. This combination of uncertainty and diversity effectively reduces labeling costs while maintaining high tracking accuracy. To further enhance tracking performance in challenging UAV scenarios, we replace the traditional Intersection over Union (IoU) computation with Foacler-IoU. This adjustment allows the trained model to better align with difficult samples, thereby improving model robustness. Finally, we evaluate our algorithm on the UAV123, UAVDT, and DTB70 datasets. The results demonstrate that our UDALT method outperforms several existing active learning methods, validating the effectiveness of our proposed tracking method.}, archive = {J_NEUCOM}, author = {Yingqin Liang and Feng Huang and Zhaobing Qiu and Xiu Shu and Qiao Liu and Di Yuan}, doi = {10.1016/j.neucom.2025.130265}, journal = {Neurocomputing}, month = {7}, pages = {130265}, shortjournal = {Neurocomputing}, title = {Uncertainty and diversity-based active learning for UAV tracking}, volume = {639}, year = {2025}, }
• (2025). Skeleton-based action recognition through dual-granularity feature fusion with self-adapting graph convolution and multi-scale temporal convolution. NEUCOM, 639, 130261. (www)

  @article{ , abstract = {3-D skeleton sequences have been widely applied in human recognition tasks due to their ability of filtering out the interferences caused by multiple uncertainties such as background noise and illumination changing. Simultaneously, the Graph Convolutional Networks (GCNs) were developed to perform feature learning from the skeleton sequences, thereby providing a pivotal method for human action recognition. However, in related researches, the complexity and computation of such models need to be greatly increased to obtain a small accuracy improvement. To deploy such models on edge devices or mobile devices, it is necessary to reduce the complexity and computational overhead of GCNs without significantly affecting their performance. This study developed a novel GCN lightweight method for human action recognition. Firstly, we proposed to extract and fuse the meta-action features and local joint motion features from 3-D skeleton sequences, replacing the multi-stream architecture or multiple convolutional layers, for mining out the key motion features, which helps to significantly reduce the number of network parameters and overall computational load. Next, we designed the MTGCN network to perform action classification with a self-adaption graph convolution module for capturing deeper dependencies between joints during movement, and an attention mechanism based multi-scale temporal convolution module for handling the actions with different limb movement patterns over multiple time spans. Comparative experiments with 11 typical methods on two datasets, NTU RGB D+60 and NTU RGB D+120, showed the proposed method, while nearing the recognition accuracy of existing SOTA methods, required only about one-twentieth of the number of parameters and computational cost.}, archive = {J_NEUCOM}, author = {Hao Chen and YiHao Shen and YuanXiang Zhang and XiaoYing Pan and Qing Yang}, doi = {10.1016/j.neucom.2025.130261}, journal = {Neurocomputing}, month = {7}, pages = {130261}, shortjournal = {Neurocomputing}, title = {Skeleton-based action recognition through dual-granularity feature fusion with self-adapting graph convolution and multi-scale temporal convolution}, volume = {639}, year = {2025}, }
• (2025). Variational methods with application to medical image segmentation: A survey. NEUCOM, 639, 130260. (www)

  @article{ , abstract = {The segmentation model based on the variational method is gradually applied to medical image segmentation, but medical image segmentation still faces many challenges, such as the interference of intensity inhomogeneity, noise and unclear boundaries, etc. To solve the difficult challenge of image segmentation, various variational methods have appeared in the past 20 years. Based on the variational method, there are many methods, which are divided into the region-based model, the edge-based model, introducing other algorithms and adding the constraint term, etc. Each of these methods has its advantages and disadvantages, but there is still a certain distance from being used in clinical medicine. Besides the gradient descent, there are also the split Bregman method and the iterative convolution-thresholding method to minimize the energy functional. This paper aims at a comprehensive review and summary of the classical and improved models in image segmentation. We detail a general statistical formulation for the variational method and introduce the basic principle, representative works and improvement of several medical image segmentation methods based on variational methods. Subsequently, we show the results and comparisons of some variational methods. We also found that the idea of variational methods can be combined with more recent comparative deep learning to achieve better performance. Finally, the shortcomings and future research directions of medical image segmentation are discussed.}, archive = {J_NEUCOM}, author = {Xiu Shu and Zhihui Li and Xiaojun Chang and Di Yuan}, doi = {10.1016/j.neucom.2025.130260}, journal = {Neurocomputing}, month = {7}, pages = {130260}, shortjournal = {Neurocomputing}, title = {Variational methods with application to medical image segmentation: A survey}, volume = {639}, year = {2025}, }
• (2025). Exploiting prior tacit knowledge to enhance alignment and verification in zero-shot video grounding. NEUCOM, 639, 130258. (www)

  @article{ , abstract = {Video Grounding (VG) aims to identify the moment-of-interest in unedited videos according to given language queries. Zero-shot methods seek to achieve grounding results without labeled data, thereby demonstrating robust generalization capabilities and extensive applicability. However, existing zero shot approaches face limitations when matching queries with videos that contain rich information, primarily due to the intricate structure and semantics of the queries. Firstly, these methods tend to overlook the inherent temporal structure of events described by queries, which includes distinct phases: the initial, climax, and decay stages. By treating all phases of an event uniformly, critical video segments may be inadvertently omitted. Secondly, for queries exhibiting similar semantics, ambiguity can occur during the grounding process, compounded by the lack of reliable verification mechanisms to assess or correct these results. To tackle these challenges, we propose a two-stage method: Exploiting Prior Tacit Knowledge to Enhance Alignment and Verification. Specifically, in the first stage, we introduce a Temporal Structure-Informed Event Proposal Generation (TSPG) module,which capitalizes on the temporal structure of events to effectively filter video candidate segments that encapsulate critical phases of the events. In the second stage, we present a Temporal Consistency Verification and Recalibration (TCVR) module, designed to rigorously examine and refine the grounding results in accordance with the established prior semantic temporal order. Extensive ablation on two datasets demonstrates the superiority of our method.}, archive = {J_NEUCOM}, author = {Jing Wang and Xianbing Zhao and Xiaojie Wang and Fangxiang Feng}, doi = {10.1016/j.neucom.2025.130258}, journal = {Neurocomputing}, month = {7}, pages = {130258}, shortjournal = {Neurocomputing}, title = {Exploiting prior tacit knowledge to enhance alignment and verification in zero-shot video grounding}, volume = {639}, year = {2025}, }
• (2025). An introspection of graph structure learning: A graph skeleton extraction via minimum dominating set. NEUCOM, 639, 130257. (www)

  @article{ , abstract = {Graph structure learning (GSL) is a data-driven learning approach that has garnered widespread attention in recent years. Nevertheless, the insufficient understanding of latent graph properties poses various challenges for effective graph modeling. This raises the following question: What type of graph skeleton can preserve the most crucial latent properties that significantly impact the performance of graph neural networks (GNNs) in downstream tasks? To this end, we have conducted a comprehensive study on three key graph properties: homophily, degree distribution, and connected components, and determined how these factors influence semi-supervised node classification tasks. Specifically, the influence of homophily on GNN performance is rigorously assessed. Motivated by the analysis, a dual-sparsity graph extraction method, based on the minimum dominating set (MDS), is proposed, to intelligently select informative edges under a given edge sampling ratio. This method effectively captures the scale-free characteristics of the degree distribution, and prioritizes the preservation of node connectivity. Experimental results show that homophily is a key factor in achieving high GNN accuracy. Additionally, the degree distribution and connected components describe the connectivity patterns of the graph from both local and global topological perspectives, which are highly correlated with node classification performance under the GNN message-passing mechanism. This work reveals the necessity of considering the graph skeleton and provides a stepping stone for facilitating GSL using these latent graph properties.}, archive = {J_NEUCOM}, author = {Zifeng Ye and Aifu Han and Guolin Chen and Xiaoxia Huang}, doi = {10.1016/j.neucom.2025.130257}, journal = {Neurocomputing}, month = {7}, pages = {130257}, shortjournal = {Neurocomputing}, title = {An introspection of graph structure learning: A graph skeleton extraction via minimum dominating set}, volume = {639}, year = {2025}, }
• (2025). Assortativity attention based multi-dim graph neural architecture search under distribution shifts. NEUCOM, 639, 130256. (www)

  @article{ , abstract = {Graph neural architecture search (GraphNAS) has demonstrated significant advantages in automatically designing graph neural network architectures for graph classification tasks. However, a major limitation of GraphNAS is its vulnerability to distribution shifts, where the optimal architectures searched for training graphs may not generalize well to unseen testing graphs. Recent approaches have addressed this issue by tailoring neural architectures for each graph through differentiable search and weight sharing. Nevertheless, these methods are constrained by the number of candidate operations and neglect the importance of extracting node link patterns to distinguish different distributions. In this paper, we propose Assortativity-attention-based Multi-Dim Graph neural Architecture Search (AMDGAS), which tailors architectures end-to-end for each graph with unknown distribution from multiple search dimensions. Specifically, we extend the vertical depth of the architecture search space, defining multiple parallel search dimensions. Different dimensions share an operation-embedding mapping, enabling the model to conduct a multi-dimension search within single-dimension complexity. In addition, we model node link patterns and propose assortativity attention, which captures the regularities of graph distributions and guides architecture search across different dimensions while cutting off architecture subspaces that may have negative impacts under particular distributions. Experimental results on synthetic and real-world datasets with distribution shifts demonstrate that our proposed method surpasses state-of-the-art algorithms and achieves excellent out-of-distribution generalization. Our code is available at https://github.com/networkanddatasciencelab/AMDGAS .}, archive = {J_NEUCOM}, author = {Yinning Shao and Jingwei Wang and Qiyi Wang and Yukai Zhao and Yunlong Ma and Min Liu and Yufeng Chen}, doi = {10.1016/j.neucom.2025.130256}, journal = {Neurocomputing}, month = {7}, pages = {130256}, shortjournal = {Neurocomputing}, title = {Assortativity attention based multi-dim graph neural architecture search under distribution shifts}, volume = {639}, year = {2025}, }
• (2025). Image manipulation localization via semantic-guided feature enhancement and deep multi-scale edge supervision. NEUCOM, 639, 130255. (www)

  @article{ , abstract = {With the widespread application of image editing software, image manipulation localization has become a focal point and promising research. Existing neural networks for image manipulation primarily rely on RGB and noise features to accurately identify tampered areas within images. However, in practical image manipulation localization tasks, noise features extracted from RGB images alone are often insufficient to effectively address tampering issues. Furthermore, existing encoder–decoder models for image manipulated localization often overlook the direct interactions between different layers during the decoding process, which hinders the effective transfer of deep semantic information to shallow features, thereby impacting the ability to accurately identify manipulated areas. To address the challenges previously identified, this paper presents a dynamically adaptive noise extraction module and achieves inter-layer information exchange in the decoder by fusing output features from different layers to extract edge information. We adaptively map RGB images to an appropriate color space using linear transformations and then extract noise features, leveraging the differences in color blocks to effectively uncover features of tampering. In addition, we integrate features across multiple decoder layers, employ deep multi-scale edge supervision to impose constraints, and introduce a dynamic ringed residual module to further enhance feature representation. Extensive experiments demonstrate that our approach achieves competitive results on diverse large-scale image datasets, exhibiting superior precision and robustness compared with most state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Haichou Wang and Hang Cheng and Yun Chen and Yongliang Xu and Meiqing Wang}, doi = {10.1016/j.neucom.2025.130255}, journal = {Neurocomputing}, month = {7}, pages = {130255}, shortjournal = {Neurocomputing}, title = {Image manipulation localization via semantic-guided feature enhancement and deep multi-scale edge supervision}, volume = {639}, year = {2025}, }
• (2025). Dual filtration subdomain adaptation network for cross-subject EEG emotion recognition. NEUCOM, 639, 130254. (www)

  @article{ , abstract = {Emotion recognition based on electroencephalogram (EEG) data holds pivotal importance for advancing affective brain-computer interfaces. However, in cross-subject emotion recognition scenarios, negative transfer is likely to happen due to EEG’s individual differences and inherent temporal variability. To solve these issues, this study proposes a novel domain adaptation architecture, named dual filtration subdomain adaptation network (DFSAN), to mitigate negative transfer and align subdomain features at a fine-grained category level. Firstly, the transferability of each subject was assessed to identify those with high transferability to serve as source domains. Then, with the feature alignment through subdomain metric learning, the transferable features could be obtained by dual filtration network. Finally, dual classifiers were employed to mitigate misclassifications near the decision boundary and output the recognition results. Multi-source cross-subject emotion recognition experiments were executed with SEED, SEED-IV, DEAP and SEED-V datasets, achieving recognition accuracy of 88.68 %, 67.61 %, 65.33 % and 65.57 %, respectively. Compared with other state-of-the-art domain adaptation methods, our proposed method achieved better results in cross-subject emotion recognition tasks, demonstrating the effectiveness and feasibility of DFSAN in handling negative transfer under multi-source transfer emotion recognition.}, archive = {J_NEUCOM}, author = {Qingshan She and Yipeng Li and Yun Chen and Ming Meng and Su Liu and Yingchun Zhang}, doi = {10.1016/j.neucom.2025.130254}, journal = {Neurocomputing}, month = {7}, pages = {130254}, shortjournal = {Neurocomputing}, title = {Dual filtration subdomain adaptation network for cross-subject EEG emotion recognition}, volume = {639}, year = {2025}, }
• (2025). Curiosity-driven exploration based on hierarchical vision transformer for deep reinforcement learning with sparse rewards. NEUCOM, 639, 130252. (www)

  @article{ , abstract = {The issue of sparse reward represents a prevalent challenge within the reinforcement learning domain. Curiosity-driven exploration has emerged as the predominant strategy to encourage reinforcement learning agents to engage in external interactions. This suite of methodologies appraises novelty by calculating the prediction error between state observations as an intrinsic reward. Initially, the Transformer, originating from machine translation task, is adapted for deep reinforcement learning. Hierarchical Vision Transformers are employed to refine the rational curiosity model and to devise an efficacious curiosity loss. This paper proposes a curiosity model based on the Dilated Neighborhood Attention Transformer(DiNAT-RCM), which can learn useful state features and enhance data utilization efficiency, thus mitigating the sparse reward problem. Moreover, an extended advantage actor-critic method, AW-A2C, is based on the self-attention mechanism. This approach increases the precision in selecting action spaces by effectively filtering action information using self-attention weights. To evaluate the proposed method, Atari 2600 environments in Gym were selected as the benchmark and compared among the baseline, Random Network Distillation(RND) and Latent Bayesian Surprise(LBS). The results demonstrate that the proposed DiNAT-RCM and AW-A2C are superior to the baseline and comparison methods.}, archive = {J_NEUCOM}, author = {Wanting Jiang and Guanwei Liu and Quanyang Leng and Nan Guo}, doi = {10.1016/j.neucom.2025.130252}, journal = {Neurocomputing}, month = {7}, pages = {130252}, shortjournal = {Neurocomputing}, title = {Curiosity-driven exploration based on hierarchical vision transformer for deep reinforcement learning with sparse rewards}, volume = {639}, year = {2025}, }
• (2025). Physics-informed echo state networks for modeling controllable dynamical systems. NEUCOM, 639, 130251. (www)

  @article{ , abstract = {Echo State Networks (ESNs) are recurrent neural networks usually employed for modeling nonlinear dynamic systems with relatively ease of training. By incorporating physical laws into the training of ESNs, Physics-Informed ESNs (PI-ESNs) were proposed initially to model chaotic dynamic systems without external inputs. They require less data for training since Ordinary Differential Equations (ODEs) of the considered system help to regularize the ESN. In this work, the PI-ESN is extended with external inputs to model controllable nonlinear dynamic systems. Additionally, an existing self-adaptive balancing loss method is employed to balance the contributions of the residual regression term and the physics-informed loss term in the total loss function. The experiments with two nonlinear systems modeled by ODEs, the Van der Pol oscillator and the four-tank system, and with one differential–algebraic (DAE) system, an electric submersible pump, revealed that the proposed PI-ESN outperforms the conventional ESN, especially in scenarios with limited data availability, showing that PI-ESNs can regularize an ESN model with external inputs previously trained on just a few datapoints, reducing its overfitting and improving its generalization error (up to 92% relative reduction in the test error). Further experiments demonstrated that the proposed PI-ESN is robust to parametric uncertainties in the ODE equations and that model predictive control using PI-ESN outperforms the one using plain ESN, particularly when training data is scarce.}, archive = {J_NEUCOM}, author = {Eric Mochiutti and Eric Aislan Antonelo and Eduardo Camponogara}, doi = {10.1016/j.neucom.2025.130251}, journal = {Neurocomputing}, month = {7}, pages = {130251}, shortjournal = {Neurocomputing}, title = {Physics-informed echo state networks for modeling controllable dynamical systems}, volume = {639}, year = {2025}, }
• (2025). SSSI-l2p: An EEG extended source imaging algorithm based on the structured sparse regularization with l2p-norm. NEUCOM, 639, 130250. (www)

  @article{ , abstract = {Electroencephalographic (EEG) source imaging (ESI) aims to estimate brain activity locations and extents. ESI is crucial for studying brain functions and detecting epileptic foci. However, accurately reconstructing extended sources remains challenging due to high susceptibility of EEG signals to interference and the underdetermined nature of the ESI problem. In this study, we introduce a new ESI algorithm, Structured Sparse Source Imaging based on L 2 p -norm (SSSI- L 2 p ), to estimate potential brain activities. SSSI- L 2 p utilizes the mixed L 2 p -norm ( 0 < p < 1 ) to enforce spatial–temporal constraints within a structured sparsity regularization framework. By leveraging the alternating direction method of multipliers (ADMM) and iteratively reweighted least squares (IRLS) algorithm, the challenging optimization problem of SSSI- L 2 p can be effectively solved. We showcase the superior performance of SSSI- L 2 p over benchmark ESI methods through numerical simulations and human clinical data. Our results demonstrate that sources reconstructed by SSSI- L 2 p exhibit high spatial resolution and clear boundaries, highlighting its potential as a robust and effective ESI technique. Additionally, we have shared the source code of SSSI- L 2 p at https://github.com/Mashirops/SSSI-L2p.git .}, archive = {J_NEUCOM}, author = {Shu Peng and Hongyu Li and Yujie Deng and Hong Yu and Weibo Yi and Ke Liu}, doi = {10.1016/j.neucom.2025.130250}, journal = {Neurocomputing}, month = {7}, pages = {130250}, shortjournal = {Neurocomputing}, title = {SSSI-l2p: An EEG extended source imaging algorithm based on the structured sparse regularization with l2p-norm}, volume = {639}, year = {2025}, }
• (2025). Deep learning-based uncertainty quantification for spatio-temporal environmental remote sensing: A systematic literature review. NEUCOM, 639, 130242. (www)

  @article{ , abstract = {Over the past few decades, a variety of Deep Learning (DL) techniques have traditionally played a significant role in environmental Remote Sensing (RS) research. With the increasing volume of large-scale data from earth observation and the rapid advancements in DL, new and exciting opportunities have emerged to enhance environmental monitoring, enable the detection of critical trends, and improve the accuracy of time series prediction. Predicting time series data remains a complex yet crucial task, especially in diverse spatio-temporal applications such as disaster response, vegetation monitoring, and forecasting droughts and floods. Due to the growing prevalence, confidence in DL forecasting has become increasingly crucial. However, basic neural networks do not provide confidence assessments and may suffer from either overconfidence or underconfidence, leading to poor calibration. Uncertainty Quantification (UQ) seeks to assess confidence in DL predictions beyond accuracy. Numerous UQ methods have emerged for DL in recent years. To fill the gaps, numerous researchers have focused on comprehending and measuring uncertainty in DL. As a consequence, diverse sources and types of uncertainty have been recognized, and multiple techniques to measure and quantify uncertainty in DL have been suggested. This study offers a comprehensive overview of uncertainty evaluation in neural networks, highlights recent advancements in the field, emphasizes present challenges, and identifies potential research opportunities. In light of this, this review will focus on the application of DL techniques to enhance the environmental RS process while addressing the inherent uncertainty in neural network predictions.}, archive = {J_NEUCOM}, author = {Aya Ferchichi and Ahlem Ferchichi and Fatma Hendaoui and Mejda Chihaoui and Radhia Toujani}, doi = {10.1016/j.neucom.2025.130242}, journal = {Neurocomputing}, month = {7}, pages = {130242}, shortjournal = {Neurocomputing}, title = {Deep learning-based uncertainty quantification for spatio-temporal environmental remote sensing: A systematic literature review}, volume = {639}, year = {2025}, }
• (2025). Weakly supervised point cloud semantic segmentation using pseudo-label reliability and consistency regularization. NEUCOM, 639, 130241. (www)

  @article{ , abstract = {Weakly supervised point cloud semantic segmentation using sparse 3D labels effectively reduces the high cost associated with annotating dense 3D labels. This paper explores pseudo-labeling and consistency regularization techniques in weakly supervised point cloud semantic segmentation, and proposes a consistency regularization sparse convolutional network based on pseudo-label reliability. We identify issues with pseudo-labels selected based on confidence, as high confidence but incorrect labels are selected, while low confidence but potentially useful labels are discarded. To address these challenges, we combine confidence and uncertainty to measure the reliability of pseudo-labels and apply different consistency constraints based on this reliability. We also employ entropy regularization loss to more accurately distinguish pseudo-label reliability. Considering the significant impact of training strategies on model performance, we investigate the influence of augmented point cloud and distribution alignment loss, both associated with consistency regularization, on model training. Extensive experiments conducted on two most popular large-scale benchmark datasets, S3DIS and ScanNet, demonstrate the effectiveness of our proposed method. The experimental results show that our method achieves optimal performance in weakly supervised point cloud semantic segmentation.}, archive = {J_NEUCOM}, author = {Lingfeng Shen and Yanlong Cao and Xiaoyao Wei and Shuai Li}, doi = {10.1016/j.neucom.2025.130241}, journal = {Neurocomputing}, month = {7}, pages = {130241}, shortjournal = {Neurocomputing}, title = {Weakly supervised point cloud semantic segmentation using pseudo-label reliability and consistency regularization}, volume = {639}, year = {2025}, }
• (2025). Privacy-preserving average consensus for second-order discrete-time multi-agent systems. NEUCOM, 639, 130239. (www)

  @article{ , abstract = {This study addresses the privacy-preserving average consensus problem in second-order discrete multi-agent systems under strongly connected and balanced graphs. When both velocity and position states of each agent are measurable, a novel lightweight algorithm is proposed by introducing perturbation signals into the transmitted information. Specifically, the algorithm is divided into two stages. In the initial stage, each agent introduces perturbation signals into its initial position and velocity states during transmission to confound potential attackers. In the subsequent stage, the agents use a standard average consensus algorithm to update their states, ensuring accurate convergence to the average of the initial states. Additionally, further considering the scenario where the velocity state is unavailable for each agent, an improved edge-based perturbation algorithm is introduced. Both algorithms not only effectively prevent the internal honest-but-curious agents from accurately inferring the initial states of other agents, except in the specific case where the curious agent is the sole neighbor of the target agent, but also protect privacy from the external eavesdroppers. Lastly, several numerical examples are conducted to validate the effectiveness of the proposed theoretical approaches.}, archive = {J_NEUCOM}, author = {Jie Wang and Na Huang and Yun Chen and Qiang Lu}, doi = {10.1016/j.neucom.2025.130239}, journal = {Neurocomputing}, month = {7}, pages = {130239}, shortjournal = {Neurocomputing}, title = {Privacy-preserving average consensus for second-order discrete-time multi-agent systems}, volume = {639}, year = {2025}, }
• (2025). Rethinking attention mechanism for enhanced pedestrian attribute recognition. NEUCOM, 639, 130236. (www)

  @article{ , abstract = {Pedestrian Attribute Recognition (PAR) plays a crucial role in various computer vision applications, demanding precise and reliable identification of attributes from pedestrian images. Traditional PAR methods, though effective in leveraging attention mechanisms, often suffer from the lack of direct supervision on attention, leading to potential overfitting and misallocation. This paper introduces a novel and model-agnostic approach, Attention-Aware Regularization (AAR), which rethinks the attention mechanism by integrating causal reasoning to provide direct supervision of attention maps. AAR employs perturbation techniques and a unique optimization objective to assess and refine attention quality, encouraging the model to prioritize attribute-specific regions. Our method demonstrates significant improvement in PAR performance by mitigating the effects of incorrect attention and fostering a more effective attention mechanism. Experiments on standard datasets showcase the superiority of our approach over existing methods, setting a new benchmark for attention-driven PAR models.}, archive = {J_NEUCOM}, author = {Junyi Wu and Yan Huang and Min Gao and Yuzhen Niu and Yuzhong Chen and Qiang Wu}, doi = {10.1016/j.neucom.2025.130236}, journal = {Neurocomputing}, month = {7}, pages = {130236}, shortjournal = {Neurocomputing}, title = {Rethinking attention mechanism for enhanced pedestrian attribute recognition}, volume = {639}, year = {2025}, }
• (2025). Cue RAG: Dynamic multi-output cue memory under h framework for retrieval-augmented generation. NEUCOM, 639, 130235. (www)

  @article{ , abstract = {How to find high-quality knowledge related to the input is crucial for retrieval-enhanced generation of RAG. The most relevant knowledge matched directly from the vector library is often the shallow knowledge that the model has already mastered. This paper introduces the concept of cue memory to help the model pay more attention to the underlying knowledge that matches the input. And through rigorous theoretical derivation, it is proved that cue memory is conducive to improving the answer accuracy of the model. At the same time, this paper proposes a new retrieval-enhanced generation framework-H framework. This framework incorporates the concept of cue memory, uses a global dependency encoder to obtain relevant contextual knowledge and feeds it into the information filter. The filter only retains the underlying knowledge that has not yet been learned to obtain cue memory. The cue memory is then output as a high-quality memory through a sequential decoder. Finally, an infinite memory pool is built, and all enhanced memories will be fed back to the pool. The quality of the memory pool continues to improve with continuous feedback. This paper has achieved good results on the public dialogue dataset DailyDialog, the translation dataset JRC-Acquis and private data. In particular, in the dialogue task, our H model achieved the best B-1/2 result, and the large model with cue memory also achieved SOTA results in the noise test. The relevant project code of this paper is publicly published at: https://github.com/Futr1/H-frame-work .}, archive = {J_NEUCOM}, author = {Yuanshuang Fu and Dan Liu and Bonan Zhang and Zhuotong Jiang and Haibo Mei and Jiajin Guan}, doi = {10.1016/j.neucom.2025.130235}, journal = {Neurocomputing}, month = {7}, pages = {130235}, shortjournal = {Neurocomputing}, title = {Cue RAG: Dynamic multi-output cue memory under h framework for retrieval-augmented generation}, volume = {639}, year = {2025}, }
• (2025). Kernel broad learning cauchy conjugate gradient algorithm for online chaotic time series prediction. NEUCOM, 639, 130234. (www)

  @article{ , abstract = {Accurate prediction of nonlinear systems in non-Gaussian noise environments has long been a significant challenge in the fields of statistical data analysis and time series modeling. To address this issue, this paper proposes an improved Cauchy Conjugate Gradient algorithm based on a kernel broad learning feature extraction strategy (Kernel Broad Learning Cauchy Conjugate Gradient, KBLCCG). This algorithm integrates kernel mapping with broad learning systems, forming a dual feature extraction mechanism that effectively captures the complex nonlinear structures of chaotic time series while preserving their inherent dynamic chaotic characteristics. The KBLCCG algorithm utilizes its robust feature extraction capabilities through the dual extraction mechanism of kernel mapping and broad learning systems, effectively capturing the intricate nonlinear structures present in time series data. The kernel broad learning strategy mitigates the phenomenon of kernel matrix size expansion during the iterative process, thereby reducing the computational burden and enhancing the algorithm's robustness. The Cauchy Conjugate Gradient method is employed to optimize the reduced-dimensional feature data, efficiently addressing the nonlinear prediction problem of the target sequence. Empirical analysis using simulation data and actual financial data (including the Lorenz system, Shanghai Composite Index, and CSI 300 Index) validates the performance of this method. Experimental results indicate that KBLCCG significantly outperforms existing adaptive filtering algorithms in terms of prediction accuracy, particularly demonstrating stronger generalization capabilities when dealing with complex chaotic systems. Compared to traditional methods, the kernel broad learning strategy markedly enhances the feature capturing and modeling effectiveness of chaotic time series, further validating the method's efficacy and robustness in nonlinear time series prediction. The KBLCCG algorithm not only exhibits superior predictive capabilities in complex non-Gaussian noise environments but also provides an innovative solution for handling the nonlinear and chaotic characteristics of time series prediction.}, archive = {J_NEUCOM}, author = {Liyun Su and Xiaoyi Wang}, doi = {10.1016/j.neucom.2025.130234}, journal = {Neurocomputing}, month = {7}, pages = {130234}, shortjournal = {Neurocomputing}, title = {Kernel broad learning cauchy conjugate gradient algorithm for online chaotic time series prediction}, volume = {639}, year = {2025}, }
• (2025). One-step multi-view clustering via label transmission and fusion. NEUCOM, 639, 130233. (www)

  @article{ , abstract = {Anchor graph-based large-scale multi-view clustering algorithms have demonstrated remarkable performance and efficiency, attracting significant attention in recent years. Despite their successes, existing methods often lack interpretability in the clustering process and fail to leverage complementary information across different views fully. To address these limitations, we propose the One-Step Multi-View Clustering via Label Transmission and Fusion (OSMVC-LTF). This method leverages the probabilistic nature of the anchor graph to establish probabilistic relationships among the anchor graph, sample labels, and anchor point labels. It achieves label transmission from a probabilistic perspective, jointly learning sample labels and anchor point labels while maintaining good interpretability. Furthermore, to better integrate information from different views, we impose a Schatten p -norm constraint on both the sample label tensor and the anchor point label tensor, enabling deep exploration of complementary information across different views from both the sample point and anchor point perspectives. Extensive experiments validate the effectiveness and robustness of OSMVC-LTF.}, archive = {J_NEUCOM}, author = {Wenhui Zhao and Sen Wang and Guangfei Li and Quanxue Gao and Cheng Deng}, doi = {10.1016/j.neucom.2025.130233}, journal = {Neurocomputing}, month = {7}, pages = {130233}, shortjournal = {Neurocomputing}, title = {One-step multi-view clustering via label transmission and fusion}, volume = {639}, year = {2025}, }
• (2025). Fixed/predefined-time synchronization of stochastic gene regulatory networks. NEUCOM, 639, 130231. (www)

  @article{ , abstract = {This paper is dedicated to researching fixed-time synchronization (FD-TS) and predefined-time synchronization (PD-TS) problems of stochastic gene regulatory networks (SGRNs). First, two new Lemmas are established based on the It o ˆ differential equation, variable substitution and some inequality techniques, which can be analyzed the issues concerning fixed/predefined-time stability and PD-TS as well as FD-TS for any stochastic systems based on Lyapunov function method. Second, based on the new Lemmas, state feedback controller maneuvers are devised to actualize FD-TS and PD-TS of SGRNs. In addition, sufficient PD-TS/FD-TS criteria are obtained that take advantage of the Lyapunov function approach and inequality technics based on control schemes and new Lemmas. The FD-TS/PD-TS criteria and the upper bound estimate formulas about the settling time are different from those in the existing FD-TS and PD-TS theorems. Finally, numerical example is offered to substantiate the availability of the theoretical methods.}, archive = {J_NEUCOM}, author = {Juanping Yang and Yuhong Sheng and Hong-Li Li and Shenglong Chen and Wei Chen}, doi = {10.1016/j.neucom.2025.130231}, journal = {Neurocomputing}, month = {7}, pages = {130231}, shortjournal = {Neurocomputing}, title = {Fixed/predefined-time synchronization of stochastic gene regulatory networks}, volume = {639}, year = {2025}, }
• (2025). Physiological navigation amplifier for remote extracting PPG signals from face video clips. NEUCOM, 639, 130228. (www)

  @article{ , abstract = {Remote photoplethysmography (rPPG) signal measurement based on face video clips is one of the essential methods for non-contact monitoring of human heart health. However, the subtle variations in face color associated with the rPPG signal can be easily contaminated by external noise, leading to a limited cross-domain generalization of existing models. In the context of short face video clips, we have observed that the influence of factors such as background, illumination, and other variables results in significant variation in noise between samples, while those within the sample typically exhibit similar patterns. Therefore, those distinctive features can be utilized to model noise variation. This paper presents a physiological navigation amplifier with a self-adaptive differential feature representation structure for rPPG signal measurement in challenging video clips. First, a noise representation encoder is proposed for self-adaptive differential feature representation. Second, a physiological navigation amplifier is designed to extract the subtle rPPG signal from facial feature sets by disentangling rPPG and noise features in the spatiotemporal space. Finally, to solve the problem of measured signal degradation caused by drastic external disturbances that result in network performance decline, a learnable signal rectification matrix is employed to reconstruct the measured rPPG signal. The experimental results, obtained through publicly available intra-datasets and cross-dataset validation, demonstrate that our proposed method reduces the mean absolute error (MAE) in rPPG signal and heart rate measurements by more than 17 % (OBF), 13 % (COHFACE), and 24 % (UBFC) compared to state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Bin Li and Wei Zhang and Hong Fu}, doi = {10.1016/j.neucom.2025.130228}, journal = {Neurocomputing}, month = {7}, pages = {130228}, shortjournal = {Neurocomputing}, title = {Physiological navigation amplifier for remote extracting PPG signals from face video clips}, volume = {639}, year = {2025}, }
• (2025). Towards faster seeding for k-means++ via lower bound and triangle inequality. NEUCOM, 639, 130227. (www)

  @article{ , abstract = {The k -means clustering is a commonly used and powerful algorithm in data science. Its performance depends heavily on the initial cluster centers. It has been widely accepted that k -means++ seeding ( KMS ) is the most prevalent approach to initialize centers. However, for each data sample x , naive KMS must calculate Euclidean distances between x and all selected centers, leading to unsatisfactory efficiency when performing it over large datasets. Previous work concentrated on approximating the seeding process of k -means++ in order to attain speedup. Nevertheless, they cannot achieve the same results as original method because they are approximation approaches inherently. This paper proposes a method called TILB (shorthand for T riangle I nequality and L ower B ound), which accelerates KMS procedure by means of (i) a lower bound scheme; and (ii) exploiting the triangle inequality. To the best of our knowledge, this comprehensive set of pruning techniques has not yet been applied to KMS . TILB is an exact method capable of performing KMS in less time while ensuring identical results. Two lower bounds built on progressive partial distance and piecewise aggregate approximation are used in TILB. The theoretical analysis is provided to demonstrate their computational efficiency and lower bound property. Experimental results show that in comparison with the current related competitors, the TILB performs fewer Euclidean distance computations and attains higher performance, especially for large k problems.}, archive = {J_NEUCOM}, author = {Haowen Zhang and Jing Li}, doi = {10.1016/j.neucom.2025.130227}, journal = {Neurocomputing}, month = {7}, pages = {130227}, shortjournal = {Neurocomputing}, title = {Towards faster seeding for k-means++ via lower bound and triangle inequality}, volume = {639}, year = {2025}, }
• (2025). Efficient and accurate graph statistics with adaptive personalized local differential privacy. NEUCOM, 639, 130224. (www)

  @article{ , abstract = {Graph analysis has attracted increasing attention in the era of the Internet of Everything. However, the extensive graph data analysis seriously violates and discloses users’ privacy. Local differential privacy (LDP) has been adopted to provide privacy guarantees for graph analysis. Nonetheless, subgraph analysis with LDP still faces serious challenges in terms of non-personalized privacy, high communication costs, and low data utility. This paper presents novel subgraph analysis algorithms customized by adaptive personalized privacy protection while achieving low communication costs and superior data utility. Considering that different edges have distinct sensitivity levels, we propose an adaptive personalized edge local differential privacy paradigm that can be applied to both k -stars counting and triangle counting with improved data utility. We further design the triangle counting mechanism that only downloads a 2-hop subgraph instead of the whole graph when interacting with users, which significantly reduces communication costs. Moreover, we leverage a variance-based threshold edge-clipping strategy to reconstruct the noisy graph, leading to high data utility and practicability. Both theoretical analyses and experimental results on real social graph datasets demonstrate that the proposed mechanisms achieve high accuracy and low communication costs while providing adaptive personalized privacy protection.}, archive = {J_NEUCOM}, author = {Teng Wang and Jiangguo Liang and Shuai Wang and Ling Zhao and Tengfei Yang}, doi = {10.1016/j.neucom.2025.130224}, journal = {Neurocomputing}, month = {7}, pages = {130224}, shortjournal = {Neurocomputing}, title = {Efficient and accurate graph statistics with adaptive personalized local differential privacy}, volume = {639}, year = {2025}, }
• (2025). An improved predictive function control algorithm via wavelet neural network for urban rail train tracking control. NEUCOM, 639, 130223. (www)

  @article{ , abstract = {This study proposes a novel, effective improved predictive function control based on a wavelet neural network (IPFC-WNN) for urban rail train tracking control. Specifically, the step function and Morlet wavelet function were chosen as the base function together, and an adaptive nonlinear online adjustment function of the softening factor was proposed based on the fuzzy satisfaction of system performance and optimisation factor. The maximum and minimum softening factors for a simple straight line can also be set appropriately by a wavelet neural network according to the actual situation. To effectively improve the control performance of the predictive function control algorithm for urban rail train tracking, calculation of additional resistance with a multiparticle model was adopted, and parameters for the adaptive nonlinear online softening factor adjustment function were set using a wavelet neural network to improve the comprehensive performance quality for urban rail train tracking control. Considering the scenario of urban rail train tracking control from Bayi Road to Yongan Four Seasons, which is located in the second-phase project of Dalian Urban Rail Transit Line 13, as the hardware-in-the-loop test object, the proposed IPFC-WNN and three improved control algorithms were used for comparative verification. The test results showed that the proposed IPFC-WNN can significantly improve the performance of the control system, and quality indicators such as energy saving, precise parking, punctuality, and comfort of the system were significantly improved. Hence, the good tracking control for train operation using the proposed IPFC-WNN was verified.}, archive = {J_NEUCOM}, author = {Longda Wang and Gang Liu and Chuanfang Xu}, doi = {10.1016/j.neucom.2025.130223}, journal = {Neurocomputing}, month = {7}, pages = {130223}, shortjournal = {Neurocomputing}, title = {An improved predictive function control algorithm via wavelet neural network for urban rail train tracking control}, volume = {639}, year = {2025}, }
• (2025). Semi-supervised structured nonnegative matrix factorization for anchor graph embedding. NEUCOM, 639, 130222. (www)

  @article{ , abstract = {Semi-supervised nonnegative matrix factorization (NMF) has been widely used in various clustering tasks due to its reliable performance. The key is how to use effectively a small amount of label information to obtain a more discriminative low-dimensional representation of data. In order to improve the clustering performance of semi-supervised NMF more effectively, this paper proposes a new semi-supervised NMF method, namely semi-supervised structured NMF for anchor graph embedding (AESSNMF). Specifically, AESSNMF uses three kinds of supervision information simultaneously, namely, pointwise constraints, pairwise constraints, and negative label information. Also, in order to handle mixed-sign data, AESSNMF uses a convex NMF form and only imposes nonnegative constraints on the coefficient matrix. AESSNMF constructs an anchor graph to embed the matrix factorization process, rather than performing the matrix factorization directly on the original data. We use the alternating iterative algorithm to optimize the objective function of AESSNMF. We also discuss the relationship between several related NMF based algorithms and AESSNMF. A large number of experimental results show that AESSNMF is superior to other related algorithms.}, archive = {J_NEUCOM}, author = {Xiangli Li and Jianping Mei and Yuanjian Mo}, doi = {10.1016/j.neucom.2025.130222}, journal = {Neurocomputing}, month = {7}, pages = {130222}, shortjournal = {Neurocomputing}, title = {Semi-supervised structured nonnegative matrix factorization for anchor graph embedding}, volume = {639}, year = {2025}, }
• (2025). Distributed event-triggered state estimation for renewable microgrids subject to incomplete observations. NEUCOM, 639, 130220. (www)

  @article{ , abstract = {In this paper, we investigate the distributed state estimation problem for renewable microgrids (RMGs) with incomplete observations, where information transmission is governed by the event-triggered communication criterion. The missing measurements with the description of uncertain occurrence probabilities (UOPs) are considered and modeled via the integration of nominal probabilities and the associated bounds. In addition, an event-triggered mechanism involving some parameters is employed to improve reliability of communication by transmitting measurements under specific triggered conditions. The aim of this paper is to design a distributed event-triggered state estimation algorithm against missing measurements under UOPs that guarantees the existence of an upper bound on the estimation error covariance (EEC) with satisfactory algorithm performance. Afterwards, the gain matrix of the corresponding state estimator is properly designed by minimizing the trace of the upper bound on the EEC. Besides, the boundedness of the upper bound of EEC is further ensured by providing a sufficient condition. Subsequently, we discuss the monotonicity relationship with respect to the trace of upper bound and the nominal occurrence probability of missing measurements. Finally, a simulation experiment with comparisons is conducted on RMGs with two distributed generation units to demonstrate the effectiveness of newly designed state estimation algorithm.}, archive = {J_NEUCOM}, author = {Peixia Gao and Wen Chen and Chaoqing Jia and Jiawen Zhang and Hongxu Zhang and Jun Hu}, doi = {10.1016/j.neucom.2025.130220}, journal = {Neurocomputing}, month = {7}, pages = {130220}, shortjournal = {Neurocomputing}, title = {Distributed event-triggered state estimation for renewable microgrids subject to incomplete observations}, volume = {639}, year = {2025}, }
• (2025). Estimating expert prior knowledge from optimization trajectories. NEUCOM, 639, 130219. (www)

  @article{ , abstract = {A recurring task in research is iterative optimization of a process that can be evaluated only by conducting an experiment. Powerful algorithms for assisting this process exist, but they largely ignore the valuable knowledge of expert scientists. We consider a problem within this general scope, not aiming to automate the optimization but instead studying how to infer tacit expert knowledge. This complements the current literature focusing on how such information is used in the optimization process, paying little attention on how the information is obtained. We consider a new formulation where the expertise is inferred by passively observing a human solving an optimization problem, without requiring explicit elicitation techniques. Our solution leverages concepts from Bayesian optimization (BO) commonly used for automating the optimization, but now these tools are used as a theoretical model for the user behavior instead. We assume the expert solves the task approximately in the same manner as a BO algorithm would and solve what kind of prior knowledge about the target function is consistent with the sequence of choices they made. We introduce the problem and a concrete solution, and show that the recovered priors match the true priors in controlled simulated studies. We also empirically evaluate the robustness of the method against violations of the modeling assumptions and demonstrate it on real user data.}, archive = {J_NEUCOM}, author = {Ville Tanskanen and Petrus Mikkola and Aras Erarslan and Arto Klami}, doi = {10.1016/j.neucom.2025.130219}, journal = {Neurocomputing}, month = {7}, pages = {130219}, shortjournal = {Neurocomputing}, title = {Estimating expert prior knowledge from optimization trajectories}, volume = {639}, year = {2025}, }
• (2025). A semi-supervised multi-task assisted method for ultrasound medical image segmentation. NEUCOM, 639, 130217. (www)

  @article{ , abstract = {The accurate segmentation of the left ventricle in echocardiography is critical for assessing cardiac function, but challenges such as blurred boundaries, high morphological variability, and limited annotated data hinder the performance of traditional methods. Recent advances in semi-supervised learning show promise in leveraging unlabeled data, yet existing techniques often rely on network or data-level perturbation, which may not fully exploit spatial and positional information essential for precise segmentation. To address these challenges, we propose a novel multi-task assisted semi-supervised segmentation framework. Our method combines segmentation, landmark detection, and image reconstruction into a unified model with a shared encoder and dual decoders. A feature cross-fusion module based on a cross-attention mechanism integrates features across tasks to enhance spatial and positional awareness. We have also introduced a contrastive learning mechanism to refine the segmentation boundaries, especially in areas where the edges are blurred. In addition, we utilize multi-scale supervision to better adapt the model to targets of different scales. The framework employs an exponential moving average student–teacher model to effectively utilize unlabeled data for training. Experiments on CAMUS and EchoNet-Dynamic datasets demonstrate that the proposed method achieves state-of-the-art performance, delivering near fully-supervised results with only 10%–20% labeled data. This highlights its potential for high-quality segmentation in low-annotation scenarios, outperforming existing semi-supervised learning methods.}, archive = {J_NEUCOM}, author = {Honghe Li and Jinzhu Yang and Mingjun Qu and Yong Feng}, doi = {10.1016/j.neucom.2025.130217}, journal = {Neurocomputing}, month = {7}, pages = {130217}, shortjournal = {Neurocomputing}, title = {A semi-supervised multi-task assisted method for ultrasound medical image segmentation}, volume = {639}, year = {2025}, }
• (2025). DRRP-net: Dense-res- recurrent prototypical networks for carbon emission prediction using satellite image time series. NEUCOM, 639, 130216. (www)

  @article{ , abstract = {Monitoring and predicting carbon emissions is crucial for understanding the impacts of environmental changes and implementing effective climate policies. Traditional methods of carbon emission estimation often rely on ground-based measurements or limited aerial surveys, which can be costly, labour-intensive, and difficult to scale. Satellite imagery offers a viable alternative by providing high-resolution, multispectral, and multi-temporal data, making it an invaluable resource for tracking land-use changes, vegetation cover, and urban development. However, existing models for carbon emission prediction using satellite imagery face challenges in accurately capturing the complex spatial and temporal dynamics involved in emission processes. To overcome, this an efficient Dense-Res-Recurrent Prototypical Network (DRRP-Net) is developed for predicting CO2 emission using satellite image time series. At first, pre-processing is conducted using a Kalman filter. Then, satellite image segmentation is conducted, wherein the images are segmented into buildings, roads, land, vegetation as well as water employing the DRRP-Net. The DRRP-Net is introduced by integrating Dense-Res-Inception Network (DRINet) and RP-Net. Moreover, the CO2 emission estimation is performed and the amount of CO2 emission estimated is compared with the original CO2 emission using the climate change Kalman to determine the accuracy. Finally, Trend Analysis is evaluated by combining all accuracy from diverse time series using Long-Short-Term Memory (LSTM). The measures used for the evaluation of DRRP-Net gained an error of 0.101, 90.40 % of accuracy and 0.934 of the Jaccard coefficient.}, archive = {J_NEUCOM}, author = {Choudari Lakshmi and Srinivasa Rao Konda}, doi = {10.1016/j.neucom.2025.130216}, journal = {Neurocomputing}, month = {7}, pages = {130216}, shortjournal = {Neurocomputing}, title = {DRRP-net: Dense-res- recurrent prototypical networks for carbon emission prediction using satellite image time series}, volume = {639}, year = {2025}, }
• (2025). Incremental learning tracking control of complex dynamical trajectories for robotic manipulator. NEUCOM, 639, 130215. (www)

  @article{ , abstract = {The inherent complexity and unpredictability of complex trajectories pose significant challenges for tracking them using a robotic manipulator’s end effector. To address this challenge, a novel incremental learning control scheme is proposed and applied to the tracking of complex dynamic trajectories by the end effector (TCDTEE). Stochastic input torque signals are exerted on the end effector of robotic manipulator so that the modeling data are collected. Then, the T-S fuzzy model (also known as an inverse model) is constructed for TCDTEE via an ensemble fuzzy framework, making it possible to obtain a control signal that can accomplish most control tasks under nominal conditions. However, the controller with fixed structure suffers from the accumulation of tracking errors and may be incapable of addressing some unexpected dynamic changes. To overcome such two problems, an interval type-2 evolving fuzzy neural network (IT2EFNN) is designed to estimate the ideal torque control rates. The IT2EFNN initializes its structure from scratch and optimizes it through the elimination of redundant rules based on the quality assessment of the fuzzy rules, thereby enhancing tracking control accuracy and system responsiveness. Meanwhile, with the help of Lyapunov analysis approach, the stability of the IT2EFNN is guaranteed. Finally, the effectiveness of the proposed method has been validated through simulations and experiments.}, archive = {J_NEUCOM}, author = {Hainan Yang and Tao Zhao and Jia-Yu Zhao and Jianjian Zhao and Peng Qin}, doi = {10.1016/j.neucom.2025.130215}, journal = {Neurocomputing}, month = {7}, pages = {130215}, shortjournal = {Neurocomputing}, title = {Incremental learning tracking control of complex dynamical trajectories for robotic manipulator}, volume = {639}, year = {2025}, }
• (2025). Explaining graph neural networks with mixed-integer programming. NEUCOM, 639, 130214. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) provide state-of-the-art graph learning performance, but their lack of transparency hinders our ability to understand and trust them, ultimately limiting the areas where they can be applied. Many methods exist to explain individual predictions made by GNNs, but there are fewer ways to gain more general insight into the patterns they have been trained to identify. Most existing methods for model-level GNN explanations attempt to generate graphs that exemplify these patterns, but the discreteness of graphs and the nonlinearity of deep GNNs make finding such graphs difficult. In this paper, we formulate the search for an explanatory graph as a mixed-integer programming (MIP) problem, in which decision variables specify the explanation graph and the objective function represents the quality of the graph as an explanation for a GNN’s predictions of an entire class in the dataset. This approach, which we call MIPExplainer, allows us to directly optimize over the discrete input space and find globally optimal solutions with a minimal number of hyperparameters. MIPExplainer outperforms existing methods in finding accurate and stable explanations on both synthetic and real-world datasets. Code is available at https://github.com/blake-gaines/MIPExplainer .}, archive = {J_NEUCOM}, author = {Blake B. Gaines and Chunjiang Zhu and Jinbo Bi}, doi = {10.1016/j.neucom.2025.130214}, journal = {Neurocomputing}, month = {7}, pages = {130214}, shortjournal = {Neurocomputing}, title = {Explaining graph neural networks with mixed-integer programming}, volume = {639}, year = {2025}, }
• (2025). Exploring inter- and intra-modal relations in compositional zero-shot learning. NEUCOM, 639, 130213. (www)

  @article{ , abstract = {Compositional Zero-Shot Learning (CZSL) aims to recognize unknown compositions by leveraging learned concepts of states and objects. Prior methods have typically emphasized either inter-modal relation for multi-modal fusion, ignoring the entanglement within state–object pairs, or solely intra-modal relation for enhancing representations, neglecting the association between vision and language domains. To tackle these limitations, we propose a CZSL framework that simultaneously learns inter- and intra-modal relations to improve image-label alignment. Firstly, we explore inter-modal relation to enable image features to grasp the cross-modal information from states and objects. The image–text fusion method facilitates the modeling of text-aware image features and image-aware text features, improving the model’s compositional recognition capability. Secondly, due to the contextuality within state–object pairs, we further explore intra-modal relation to exploit semantic information from various representation subspaces, facilitating the comprehensive semantic expression of text features. Moreover, we propose a composition fusion module to establish semantic entanglement within state–object compositions. Extensive experiments demonstrate that our method significantly surpasses the state-of-the-art methods in both closed-world and open-world settings.}, archive = {J_NEUCOM}, author = {Xiao Zhang and Hui Chen and Haodong Jing and Yongqiang Ma and Nanning Zheng}, doi = {10.1016/j.neucom.2025.130213}, journal = {Neurocomputing}, month = {7}, pages = {130213}, shortjournal = {Neurocomputing}, title = {Exploring inter- and intra-modal relations in compositional zero-shot learning}, volume = {639}, year = {2025}, }
• (2025). DSAA: Cross-modal transferable double sparse adversarial attacks from images to videos. NEUCOM, 639, 130212. (www)

  @article{ , abstract = {Recent preliminary research has examined cross-modal attacks, enabling the transition of unlabeled attacks from images to videos. The current cross-modal attacks are mainly dense attacks, while sparse attacks have not been explored. To bridge this gap, this paper proposes a cross-modal transferable double sparse adversarial attack (DSAA) method. Specifically, DSAA incorporates both spatial sparsity and temporal sparsity. Firstly, we design a novel double-cross-shaped patch paradigm to achieve spatial sparsity. Leveraging this patch for perturbation addition enables a more effective global impact on video frames. Subsequently, we propose a uniform temporal sparsity strategy to achieve temporal sparsity. Even if only half of the frames are perturbed, the resulting average attack success rate remains comparable to that of perturbing all frames. Furthermore, we take into account the luminance, contrast, and structure of images and propose a structural similarity loss function (SSLF). By utilizing SSLF to measure the similarity of intermediate feature maps and incorporating it into perturbation optimization, DSAA achieves excellent attack performance. Extensive experiments demonstrate DSAA’s superiority over state-of-the-art approaches in terms of attack success rate. The proposed DSAA uses video adversarial examples generated by four ImageNet-pretrained image models on Kinetics-400 and UCF-101 to attack six video recognition models, achieving average attack success rates of 85.83% and 82.02%, respectively.}, archive = {J_NEUCOM}, author = {Xueqiang Han and Shihui Zhang and Houlin Wang and Qing Tian}, doi = {10.1016/j.neucom.2025.130212}, journal = {Neurocomputing}, month = {7}, pages = {130212}, shortjournal = {Neurocomputing}, title = {DSAA: Cross-modal transferable double sparse adversarial attacks from images to videos}, volume = {639}, year = {2025}, }
• (2025). Tensor networks for explainable machine learning in cybersecurity. NEUCOM, 639, 130211. (www)

  @article{ , abstract = {In this paper we show how tensor networks help in developing explainability of machine learning algorithms. Specifically, we develop an unsupervised clustering algorithm based on Matrix Product States (MPS) and apply it in the context of a real use-case of adversary-generated threat intelligence. Our investigation proves that MPS rival traditional deep learning models such as autoencoders and GANs in terms of performance, while providing much richer model interpretability. Our approach naturally facilitates the extraction of feature-wise probabilities, Von Neumann Entropy, and mutual information, offering a compelling narrative for classification of anomalies and fostering an unprecedented level of transparency and interpretability, something fundamental to understand the rationale behind artificial intelligence decisions.}, archive = {J_NEUCOM}, author = {Borja Aizpurua and Samuel Palmer and Román Orús}, doi = {10.1016/j.neucom.2025.130211}, journal = {Neurocomputing}, month = {7}, pages = {130211}, shortjournal = {Neurocomputing}, title = {Tensor networks for explainable machine learning in cybersecurity}, volume = {639}, year = {2025}, }
• (2025). A review of transformer-based human pose estimation: Delving into the relation modeling. NEUCOM, 639, 130210. (www)

  @article{ , abstract = {Recent advancements in transformer have profoundly impacted the field of computer vision, marking a pivotal shift in Human Pose Estimation (HPE) research from conventional deep learning models to transformer-based approaches. These approaches are known for their superior representation capability, which significantly enhances model performance. This survey explores the reason for this transformation, providing a detailed examination of the evolution and methodologies of transformer-based HPE techniques. Initially, we present a theoretical and qualitative analysis emphasizing the efficacy and advantages of transformer in HPE. Following this, we introduce an innovative and comprehensive taxonomy to categorize transformer-based HPE methods, emphasizing the role of transformer in their distinctive paradigms and pipelines. Furthermore, we present a unified perspective on HPE paradigms and pipelines through relation modeling, a cornerstone concept in transformer model. To empirically validate the superiority of transformer models over traditional deep learning methodologies, we conduct a comprehensive review of commonly used datasets in HPE and compare the performance of various approaches on these datasets. Lastly, we identify current challenges and propose future avenues for research in this rapidly evolving field.}, archive = {J_NEUCOM}, author = {Jiangning Wei and Bo Yu and TianJian Zou and Yida Zheng and Xinzhu Qiu and Minzhen Hu and Hao Yu and Dandan Xiao and Yang Yu and Jiu Liu}, doi = {10.1016/j.neucom.2025.130210}, journal = {Neurocomputing}, month = {7}, pages = {130210}, shortjournal = {Neurocomputing}, title = {A review of transformer-based human pose estimation: Delving into the relation modeling}, volume = {639}, year = {2025}, }
• (2025). Co-manifold learning for semi-supervised medical image segmentation. NEUCOM, 639, 130208. (www)

  @article{ , abstract = {In this study, we investigate jointly learning Hyperbolic and Euclidean space representations and match the consistency for semi-supervised medical image segmentation. We argue that for complex medical volumetric data, hyperbolic spaces are beneficial to model data inductive biases. We propose an approach incorporating the two geometries to co-train a variational encoder–decoder model with a Hyperbolic probabilistic latent space and a separate variational encoder–decoder model with a Euclidean probabilistic latent space with complementary representations, thereby bridging the gap of co-training across manifolds (Co-Manifold learning) in a principled manner. To capture complementary information and hierarchical relationships, we propose a Latent Space Loss aimed at maximizing disagreement between embeddings across manifolds. Additionally, we employ adversarial learning to enhance segmentation performance by guiding the network in hyperbolic latent space using confident regions identified by the network in Euclidean space. Conversely, the network in Euclidean space is informed by hyperbolic uncertainty, creating a dual uncertainty-aware framework that enables the two spaces to collaboratively learn confident regions from each other. Our proposed method achieves competitive results on two benchmarks for semi-supervised medical image segmentation on medical scans. The code is publicly available at: https://github.com/himashi92/Co-Manifold .}, archive = {J_NEUCOM}, author = {Himashi Peiris and Zhaolin Chen and Gary Egan and Mehrtash Harandi}, doi = {10.1016/j.neucom.2025.130208}, journal = {Neurocomputing}, month = {7}, pages = {130208}, shortjournal = {Neurocomputing}, title = {Co-manifold learning for semi-supervised medical image segmentation}, volume = {639}, year = {2025}, }
• (2025). Design of UAV target detection network based on deep feature fusion and optimization with small targets in complex contexts. NEUCOM, 639, 130207. (www)

  @article{ , abstract = {Target detection of miniature Unmanned Aerial Vehicles (UAVs) in urban feature backgrounds faces significant challenges due to the complexity of urban environments and the small size and limited features of UAVs, leading to frequent omissions and misdetections. To address these challenges, this paper introduces a lightweight and efficient detection model for miniature UAVs in complex urban settings. A novel bounding box loss function named cover-type Minimum Points Distance-IoU (c-MPDIoU) is proposed, optimizing small target regression accuracy while reducing computational complexity. Meanwhile, the More Spatial Information Feature Pyramid Network (MSI-FPN) is designed, which uses the Receptive Field Coordinate Attention Convolution (RFCAConv) layer to solve the parameter sharing problem, enhance the extraction of spatial feature information and reduce the number of model parameters, and also extends the original Path Aggregation Network with Feature Pyramid Network (PAN-FPN) structure upwards to increase the fusion frequency of high-level and low-level feature information. In addition, the Efficient Multi-scale Attention (EMA) mothed is used to embed some modules in the neck part of the original network to further improve the efficiency of feature fusion. Furthermore, the detection head is redesigned with three larger-sized detection heads for improved accuracy in small target detection. Experimental results on the self-built dataset City-UAV show that the proposed model improves the AP 50 by 2.6% to 93.4% and reduces the model size by 17.9% compared to yolov8s. In addition, we conduct comparative experiments using the publicly available CVPR2023 Anti-UAV and the DUT Anti-UAV datasets to demonstrate the feasibility of the proposed model.}, archive = {J_NEUCOM}, author = {Jianzheng Liu and Bin Wen and Jiayi Xiao and Minghui Sun}, doi = {10.1016/j.neucom.2025.130207}, journal = {Neurocomputing}, month = {7}, pages = {130207}, shortjournal = {Neurocomputing}, title = {Design of UAV target detection network based on deep feature fusion and optimization with small targets in complex contexts}, volume = {639}, year = {2025}, }
• (2025). A riemannian convolutional neural network for EEG-based motor imagery decoding. NEUCOM, 639, 130206. (www)

  @article{ , abstract = {Decoding motor imagery (MI) intentions from non-invasive EEG signals is a persistent challenge in brain-computer interface (BCI) research due to the low signal-to-noise ratio and high variability inherent in EEG data. To tackle these issues, we introduce a Riemannian Convolutional Neural Network (RMCNN), designed to enhance MI decoding performance. RMCNN begins with a spatial–temporal convolutional layer that captures the local spatial and temporal dependencies in EEG data, followed by a Riemannian block for extracting global geometric features, and concludes with a fully connected layer for MI classification. By incorporating Riemannian geometry methods into CNN, RMCNN enables powerful feature extraction without complex structures. Additionally, the proposed method leverages Euclidean Alignment to make the data distributions from different sessions more similar, thereby reducing inter-session variability. To further improve robustness, RMCNN is trained using a dual-loss strategy that combines cross-entropy loss with triplet loss. We evaluated RMCNN on three publicly available MI datasets (BCI Competition IV 2a, OpenBMI and High Gamma), where it significantly outperformed several state-of-the-art MI decoding methods in cross-session scenarios. These results underline the effectiveness of RMCNN in addressing the challenges of EEG variability and enhancing MI decoding accuracy. By bridging advancements in Riemannian geometry and deep learning, this work represents a step forward in the development of more reliable and robust BCI systems.}, archive = {J_NEUCOM}, author = {Changchun Li and Zhenghui Gu}, doi = {10.1016/j.neucom.2025.130206}, journal = {Neurocomputing}, month = {7}, pages = {130206}, shortjournal = {Neurocomputing}, title = {A riemannian convolutional neural network for EEG-based motor imagery decoding}, volume = {639}, year = {2025}, }
• (2025). Region-aware discriminative learning GAN for super-resolution reconstruction of infrared imagery. NEUCOM, 639, 130202. (www)

  @article{ , abstract = {Infrared image super-resolution reconstruction is crucial for improving the quality of infrared images. The cutting-edge efforts often suffer from problems such as unnatural texture and fake details, undermining image realism. This study introduces a novel unified framework, which combines a multi-order degradation simulation model and region-aware discriminative learning with generative adversarial network, to train the optimal model for infrared super-resolution reconstruction. The multi-order degradation model is established to enhance the effectiveness of visual reconstruction by simulating different degradation instances in authentic infrared scenarios. The generator adopts residual-in-residual dense blocks to enhance detail preservation capability. The discriminator uses an encoder–decoder architecture for semantic discrimination. This enables simultaneous global and local assessments of images as real or fake, providing a more accurate discrimination of the generated image’s authenticity. Additionally, to discriminate high-frequency artifacts and authentic details in infrared images with intricate textures, a region-aware discriminative learning strategy is introduced. Furthermore, a hybrid loss function is developed, integrating local loss, adversarial loss, pixel-wise and perceptual loss together for advanced adversarial training, making the reconstructed images more realistic and natural. The method’s efficacy is demonstrated on simulated and real-world infrared image datasets, with comparative analysis showing significant improvements. Our model outperforms the state-of-the-art alternative solutions on multiple benchmarks.}, archive = {J_NEUCOM}, author = {Feng Deng and Shuaichao Wang and Jingjing Yang}, doi = {10.1016/j.neucom.2025.130202}, journal = {Neurocomputing}, month = {7}, pages = {130202}, shortjournal = {Neurocomputing}, title = {Region-aware discriminative learning GAN for super-resolution reconstruction of infrared imagery}, volume = {639}, year = {2025}, }
• (2025). E-mamba: An efficient mamba point cloud analysis method with enhanced feature representation. NEUCOM, 639, 130201. (www)

  @article{ , abstract = {As a key technology for three-dimensional space analysis, point cloud analysis is widely used in many fields such as automated machinery, unmanned vehicles and virtual reality. Learning local and global features of point cloud is crucial for gaining a deep understanding of point cloud data. In point cloud local feature learning, sub-clouds with center coordinates subtracted are usually used as point patches, which are then input into mini-PointNet to enhance the point cloud feature representation. However, this method has a high dependence on the point cloud density, which affects the model performance. In this work, we introduce E-Mamba, a new model for efficient point cloud analysis. We use Scalable Embedding to rescale and patch embedding sub-clouds, which improves the model’s feature representation and generalization capabilities for point cloud. In addition, we also introduced Holosync Reordering Pooling to reorder tokens while preserving the original sequence, and used the hybrid pooling method to extract global features. In this way, the model fully utilizes the periodicity of Mamba while achieving good generalization and global feature extraction capabilities. We conduct extensive experiments on ModelNet40, ScanObjectNN, and ShapeNetPart datasets. The results show that E-Mamba can achieve superior performance while significantly reducing GPU memory usage and FLOPs, whether pre-trained or not.}, archive = {J_NEUCOM}, author = {Dengao Li and Zhichao Gao and Shufeng Hao and Ziyou Xun and Jiajian Song and Jie Cheng and Jumin Zhao}, doi = {10.1016/j.neucom.2025.130201}, journal = {Neurocomputing}, month = {7}, pages = {130201}, shortjournal = {Neurocomputing}, title = {E-mamba: An efficient mamba point cloud analysis method with enhanced feature representation}, volume = {639}, year = {2025}, }
• (2025). Modeling detail feature connections for infrared image enhancement. NEUCOM, 639, 130200. (www)

  @article{ , abstract = {While current image super-resolution (SR) methods have shown remarkable achievements across diverse tasks, they often fall short in enhancing infrared (IR) images due to IR-specific characteristics. Notably, the longer wavelengths of infrared radiation, compared to visible light, lead to IR images characterized by lower spatial resolution and less detail. This inherent limitation compromises IR cameras’ capacity to capture high-frequency infrared details. Furthermore, the typical forward propagation in neural networks tends to diminish high-frequency information, making the infrared SR quality more sensitive to the restoration of high-frequency information in infrared images, which existing methods fall short of. To bridge these issues, we propose a long-range sequential processing SR paradigm tailored for infrared images. Specifically, we introduce the Mamba Channel Recursion Module (MCRM) that leverages the selective structured state space model (SSM)’s capability for modeling long-range dependencies to establish a global cross-layer relation that adeptly recovers high-frequency details. Furthermore, we introduce the adaptive deformable self-attention block (AD-SA) and the adaptive deformable cross-attention block (AD-CA) to dynamically offset sampling points, enabling the reconstruction to concentrate more on modal-specific features, and effectively restore infrared high-frequency details. Extensive experiments across various datasets demonstrate that our method notably surpasses existing SR methods, attaining state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Ziang Meng and Kaiqi Han and Yuchun He and Yuhan He and Xingyuan Li and Yang Zou}, doi = {10.1016/j.neucom.2025.130200}, journal = {Neurocomputing}, month = {7}, pages = {130200}, shortjournal = {Neurocomputing}, title = {Modeling detail feature connections for infrared image enhancement}, volume = {639}, year = {2025}, }
• (2025). Distributed consensus optimal fault-tolerant control for nonlinear multi-agent systems with external disturbance. NEUCOM, 639, 130199. (www)

  @article{ , abstract = {Under switching topology, a distributed optimal fault-tolerant control (FTC) law is presented for affine nonlinear multi-agent systems (MASs) with actuator faults and external disturbances. In control design, the control law consists of an optimal control law and a fault compensator. Firstly, by using the critic neural network (CNN) and the experience replay (ER), the integral reinforcement learning (IRL) algorithm is utilized to approximate the coupled Hamilton–Jacobi–Isaac (HJI) equation when fault-free. Then, a fault compensator is designed for actuator faults. It is worth mentioning that the proposed optimal FTC law ensures that all the closed-loop signals remain bounded by the Lyapunov stability theory and the average dwell time method. Finally, the simulation result illustrates the effectiveness of the proposed optimal FTC approach.}, archive = {J_NEUCOM}, author = {Guowei Dong and Kewen Li and Yongming Li}, doi = {10.1016/j.neucom.2025.130199}, journal = {Neurocomputing}, month = {7}, pages = {130199}, shortjournal = {Neurocomputing}, title = {Distributed consensus optimal fault-tolerant control for nonlinear multi-agent systems with external disturbance}, volume = {639}, year = {2025}, }
• (2025). Recursively summarizing enables long-term dialogue memory in large language models. NEUCOM, 639, 130193. (www)

  @article{ , abstract = {Recently, large language models (LLMs), such as GPT-4, stand out remarkable conversational abilities, enabling them to engage in dynamic and contextually relevant dialogues across a wide range of topics. However, in a long-term conversation, these chatbots fail to recall appropriate information from the past, resulting in inconsistent responses. To address this, we propose to recursively generate summaries/ memory using large language models to enhance their long-term dialog ability. Specifically, our method first stimulates the LLM to memorize small dialogue contexts. After that, the LLM recursively produces new memory using previous old memory and subsequent contexts. Finally, the chatbot is prompted to generate a response based on the latest memory. The experiments on widely used LLMs show that our method generates more consistent responses in long-term conversations, and it can be significantly enhanced with just two/ three dialog illustrations. Also, we find that our strategy could nicely complement both large context windows ( e.g ., 8K and 16K) and retrieval-enhanced LLMs, bringing further long-term dialogue performance. Notably, our method is a potential solution to enable the LLM to model the extremely long dialog context. The code will be released later.}, archive = {J_NEUCOM}, author = {Qingyue Wang and Yanhe Fu and Yanan Cao and Shuai Wang and Zhiliang Tian and Liang Ding}, doi = {10.1016/j.neucom.2025.130193}, journal = {Neurocomputing}, month = {7}, pages = {130193}, shortjournal = {Neurocomputing}, title = {Recursively summarizing enables long-term dialogue memory in large language models}, volume = {639}, year = {2025}, }
• (2025). Pose-independent efficient gauge equivariant network for 3D mesh aneurysm segmentation. NEUCOM, 639, 130188. (www)

  @article{ , abstract = {Accurate detection and segmentation of intracranial aneurysms (IAs) in three-dimensional medical imaging are crucial for effective diagnosis and therapeutic planning, given the complex nature of cerebral vascular structures. These aneurysms present significant challenges due to their variable sizes, unpredictable locations, diverse poses, and morphological heterogeneity, making consistent and reliable segmentation a critical yet demanding task in medical imaging. To effectively tackle these challenges, this study leverages the principles of anisotropic gauge equivariant convolutions (GEC) to propose a comprehensive framework that encompasses three key innovations: a Pose-Independent feature Representation (PIR) module, an Intra-layer Recurrent Convolutional (IRC) module and a boundary enhanced Laplacian loss. We introduce a novel PIR module, that utilizes anisotropic GEC to handle the inherent geometric complexities that independently of their poses. Our network, featuring a IRC module, strategically reuses convolutional weights to extend the global receptive field without additional parameters, thereby maintaining efficient high-resolution feature learning across mesh surfaces. Additionally, we incorporate a novel Laplacian loss that enforces boundary sharpness, greatly enhancing the delineation of aneurysm boundaries. The proposed method achieves superior segmentation accuracy, clearly delineating aneurysm necks without extensive data augmentation. Experiments on two datasets (IntrA and IntrANeurIST) verify the validity of the proposed method by outperforming the state of the art method together with ablation study. The method can be easily extended in the similar aneurysm segmentation such as abdominal aortic aneurysm and ophthalmic aneurysms.}, archive = {J_NEUCOM}, author = {Xudong Ru and Xingce Wang and Zhenhong Liu and Peng Du and Haichuan Zhao and Zhongke Wu and Xiaodong Ju and Shaolong Liu and Yi-Cheng Zhu and Alejandro F. Frangi}, doi = {10.1016/j.neucom.2025.130188}, journal = {Neurocomputing}, month = {7}, pages = {130188}, shortjournal = {Neurocomputing}, title = {Pose-independent efficient gauge equivariant network for 3D mesh aneurysm segmentation}, volume = {639}, year = {2025}, }
• (2025). Physics-informed causal learning network for fault diagnosis of rotating machinery under unseen operating conditions. NEUCOM, 639, 130187. (www)

  @article{ , abstract = {Domain generalization methods have been extensively utilized to fault diagnosis tasks under unseen operating conditions. However, the current mainstream approaches primarily focus on achieving statistical consistency across source domains without exploring their genuine physical correlations. This leads to a degradation of generalization performance when significant discrepancies exist in the target domain. To tackle this problem, this paper proposes a physics-informed causal learning network (PICL). This model achieves domain generalization by extracting features that exhibit both genuine physical and statistical correlations with fault labels. Firstly, a causal mechanism based on the physical features of fault signals is constructed, capable of characterizing domain-invariant features hidden in the raw signals. Secondly, a feature generator based on ResNet18 embedded with continuous wavelet transform and a neural network-based adversarial mask are designed. They can ensure the extracted features are correlated with fault categories in terms of physical attributes and statistics, making the model interpretable and highly generalizable. Thirdly, a causal feature separation module is established to provide a criterion for ensuring that the features fed into the classifier are domain-invariant, further enhancing the model’s generalization capabilities.Finally, experimental and engineering applications demonstrate that PICL outperforms other methods in fault diagnosis under unseen operating conditions.}, archive = {J_NEUCOM}, author = {Yiyi Huang and Baoping Tang and Qichao Yang and Zhen Ming}, doi = {10.1016/j.neucom.2025.130187}, journal = {Neurocomputing}, month = {7}, pages = {130187}, shortjournal = {Neurocomputing}, title = {Physics-informed causal learning network for fault diagnosis of rotating machinery under unseen operating conditions}, volume = {639}, year = {2025}, }
• (2025). Fixed-time adaptive consistent control of higher-order nonlinear multi-agent systems with full state constraints and input saturation. NEUCOM, 639, 130180. (www)

  @article{ , abstract = {This paper investigates high-order nonlinear multi-agent systems with state constraints and input saturation. A novel control scheme incorporating Neural Networks and Barrier Lyapunov Functions is designed to achieve adaptive fixed-time consensus control. This innovative scheme effectively addresses the complexity explosion problem typical in traditional controller designs while ensuring that the closed-loop system remains within its constraints. During the design process, a first-order sliding mode differentiator was introduced, and compensations were made for filter errors to ensure stability and consistency within a fixed-time. Additionally, experiments using Matlab numerical simulations and the StarSim semi-physical simulation platform confirm that the proposed control scheme significantly surpasses traditional methods in efficiency and accuracy, validating its effectiveness and practicality for solving the consensus problem in high-order nonlinear multi-agent systems.}, archive = {J_NEUCOM}, author = {Guoqiang Zhu and Xuecheng Zhang and Xiuyu Zhang and Chun-Yi Su}, doi = {10.1016/j.neucom.2025.130180}, journal = {Neurocomputing}, month = {7}, pages = {130180}, shortjournal = {Neurocomputing}, title = {Fixed-time adaptive consistent control of higher-order nonlinear multi-agent systems with full state constraints and input saturation}, volume = {639}, year = {2025}, }
• (2025). End-to-end video object detection based on dynamic anchor box spatiotemporal decoder and hybrid matching. NEUCOM, 639, 130177. (www)

  @article{ , abstract = {Despite the significant progress in object detection using single-frame images, when dealing with continuously changing video scenes, relying solely on information from individual frames often fails to fully exploit the dynamic continuity and consistency across the temporal dimension. Designing a model structure that can capture local details while also understanding the global spatiotemporal context remains a challenging problem. To address this, this paper proposes an end-to-end video object detection algorithm based on dynamic anchor box spatiotemporal decoder and hybrid matching (DAHM-Net). Specifically, a dynamic anchor box spatiotemporal decoder is proposed, where each decoder layer iteratively updates the position priors that account for object scales, enabling the model to better learn object features and improve detection performance. Additionally, a hybrid matching training strategy is proposed, combining one-to-one and one-to-many matching to enhance model performance and accelerate convergence, while maintaining end-to-end detection. Experimental results show that the proposed method outperforms TransVOD by 0.7, 1.0, and 0.5 AP 50 on three public datasets, and also surpasses recent state-of-the-art approaches.}, archive = {J_NEUCOM}, author = {Zhe Fu and Yuan Shuo and Pengjun Cao and Jing Wei and Heng Wang and Gaoxiang Zhang}, doi = {10.1016/j.neucom.2025.130177}, journal = {Neurocomputing}, month = {7}, pages = {130177}, shortjournal = {Neurocomputing}, title = {End-to-end video object detection based on dynamic anchor box spatiotemporal decoder and hybrid matching}, volume = {639}, year = {2025}, }
• (2025). Feature selections integrating algebraic and information perspectives in weighted incomplete neighborhood rough sets. NEUCOM, 639, 130164. (www)

  @article{ , abstract = {In data-driven scientific research and practical applications, data incompleteness and uncertainty are widespread issues that have become critical bottlenecks, restricting the accuracy of data analysis and the reliability of decision-making. Addressing the limitations of existing incomplete rough set models, which predominantly focus on uncertainty measurement adjustment while overlooking feature weighting and neighborhood relation construction, this paper proposes feature selection methods based on a weighted incomplete neighborhood rough set framework, integrating algebraic and information perspectives. Firstly, a weighted tolerance neighborhood relation is introduced to better quantify uncertainty, enhancing adaptability in classification and feature selection tasks. Secondly, from the algebraic perspective, three weighted measures are developed: weighted approximation accuracy, weighted information granularity, and weighted approximation precision based on information granularity. These measures are combined with information-theoretic metrics such as mutual information, complementary mutual information, and self-information to form nine fusion measures. Finally, a unified feature selection algorithmic framework is designed to comprehensively evaluate feature importance. Experimental results demonstrate that the proposed methods significantly improve classification accuracy across 12 datasets. Notably, under a 10% incompleteness rate, the GASI-FS, GMI-FS, and AMI-FS algorithms achieve classification accuracies of 87.31%, 85.87%, and 86.79% on KNN, CART, and SVM classifiers, respectively, outperforming other methods. These findings provide a robust theoretical foundation and practical tools for analyzing incomplete data in complex scenarios.}, archive = {J_NEUCOM}, author = {Shan Zhang and Jiucheng Xu and Qing Bai}, doi = {10.1016/j.neucom.2025.130164}, journal = {Neurocomputing}, month = {7}, pages = {130164}, shortjournal = {Neurocomputing}, title = {Feature selections integrating algebraic and information perspectives in weighted incomplete neighborhood rough sets}, volume = {639}, year = {2025}, }
• (2025). Modality-aware contrast and fusion for multi-modal summarization. NEUCOM, 639, 130094. (www)

  @article{ , abstract = {Multimodal Summarization with Multi-modal Output (MSMO) is an emerging field focused on generating reliable and high-quality summaries by integrating various media types, such as text and video. Current methods primarily focus on integrating features from different modalities, but often overlook further enhancement and optimization of the fused features. This limitation can reduce the representational capacity of the fusion, ultimately diminishing overall performance. To address these challenges, a novel Modality-aware Contrast and Fusion (MCF) network has been proposed. This network leverages contrastive learning to preserve the integrity of modality-specific semantics while promoting the complementary integration of different media types. The Multi-Modal Attention (MMA) module captures temporal dependencies and learns discriminative semantics for individual media types through uni-modal semantic attention, while aligning and integrating semantics from multiple sources via cross-modal semantic attention. The Uni-Cross Contrastive Learning (UCC) module minimizes modality-aware contrastive losses to enhance the distinctiveness of semantic representations. The Modality-Aware Fusion (MAF) module dynamically adjusts the contributions of uni-modal and cross-modal outputs during the summarization process, optimizing the integration based on the strengths of each modality. Extensive validation on the Bliss, Daily Mail, and CNN datasets demonstrates the state-of-the-art performance of the MCF network and confirms the effectiveness of its components.}, archive = {J_NEUCOM}, author = {Lixin Dai and Tingting Han and Zhou Yu and Jun Yu and Min Tan and Yang Liu}, doi = {10.1016/j.neucom.2025.130094}, journal = {Neurocomputing}, month = {7}, pages = {130094}, shortjournal = {Neurocomputing}, title = {Modality-aware contrast and fusion for multi-modal summarization}, volume = {639}, year = {2025}, }
• (2025). Array geometry dependency in sound speed reconstruction using deep neural network. NEUCOM, 639, 130076. (www)

  @article{ , abstract = {Deep Neural Networks (DNN) are expected to provide better solutions for the accuracy and processing time of sound speed reconstruction in Ultrasound CT (USCT). In this study, we investigated the effect of array geometry on the image quality of reconstruction by DNN in comparison with the conventional method: Full Waveform Inversion (FWI). Pressure signals for a single frequency were computed by the finite difference method for three array geometries with different element pitch sizes with a ring array transducer. The DNN was trained end-to-end on the task of reconstructing the sound speed distribution from the observed signal. For test data, we utilized numerical breast models which containing a tumor model with different diameters and sound speeds. For training data, we utilized artificial sound speed distribution based on a natural images dataset that is different from the test data and is rich in diverse patterns for appropriate evaluation of the generalization performance. The results for the test data showed that when the element pitch size is around half of a wavelength, DNN can improve the image quality by further reducing the pitch, while FWI cannot. In addition, when the element pitch is larger than one wavelength, DNN can reduce artifacts compared to FWI. When the element pitch is smaller than half a wavelength, DNN has the advantage of better reconstruction of tumor shape and smaller diameter error than FWI for large diameter tumors with large sound speed difference. On the other hand, for small diameter tumors with small sound speed difference, FWI has the advantage of capturing the structure that cannot be captured by DNN. From the above, DNN may be able to improve image quality by designing a probe with a finer pitch than that of conventional probes, while FWI is not able to. In addition, using a probe with a coarser pitch than that of a conventional one may reduce the number of AD converters in the probe and significantly lower the cost, since DNN does not produce the serious artifacts that arise in FWI.}, archive = {J_NEUCOM}, author = {Yoshiki Watanabe and Takashi Azuma and Shu Takagi}, doi = {10.1016/j.neucom.2025.130076}, journal = {Neurocomputing}, month = {7}, pages = {130076}, shortjournal = {Neurocomputing}, title = {Array geometry dependency in sound speed reconstruction using deep neural network}, volume = {639}, year = {2025}, }
• (2025). Overcoming the price of anarchy by steering with recommendations. NEUCOM, 639, 129993. (www)

  @article{ , abstract = {Varied real world systems such as transportation networks, supply chains and energy grids present coordination problems where many agents must learn to share resources. It is well known that the independent and selfish interactions of agents in these systems may lead to inefficiencies, often referred to as the ‘Price of Anarchy’. Effective interventions that reduce the Price of Anarchy while preserving individual autonomy are of great interest. In this paper we explore recommender systems as one such intervention mechanism. We start with the Braess Paradox, a congestion game model of a routing problem related to traffic on roads, packets on the internet, and electricity on power grids. Following recent literature, we model the interactions of agents as a repeated game between Q -learners, a common type of reinforcement learning agents. This work introduces the Learning Dynamic Manipulation Problem, where an external recommender system can strategically trigger behavior by picking the states observed by Q -learners during learning. Our computational contribution demonstrates that appropriately chosen recommendations can robustly steer the system towards convergence to the social optimum, even for many players. Our theoretical and empirical results highlight that increases in the recommendation space can increase the steering potential of a recommender system, which should be considered in the design of recommender systems.}, archive = {J_NEUCOM}, author = {Cesare Carissimo and Marcin Korecki and Damian Dailisan}, doi = {10.1016/j.neucom.2025.129993}, journal = {Neurocomputing}, month = {7}, pages = {129993}, shortjournal = {Neurocomputing}, title = {Overcoming the price of anarchy by steering with recommendations}, volume = {639}, year = {2025}, }
• (2025). Remaining useful life prediction with uncertainty quantification for rotating machinery: A method based on explainable variational deep gaussian process. NEUCOM, 638, 130232. (www)

  @article{ , abstract = {Remaining Useful Life (RUL) prediction is one of the key technologies to ensure the safety and reliability of mechanical equipment. To address the challenges of low prediction accuracy and insufficient uncertainty quantification in rotating machinery monitoring, this paper proposes an innovative Variational Deep Gaussian Process (VDGP) method. The proposed method adopts an Adaptive Inter-layer Variational Inference (AIVI) strategy, integrating inter-layer dependency modeling, adaptive inducing point optimization, and hierarchical variational lower bound, enhancing information transfer and feature extraction in multi-layer structures while reducing redundant computation and improving training efficiency. Through Shapley Additive exPlanations (SHAP) analysis and correlation analysis between input features and hidden layer node outputs, the key factors in RUL prediction are deeply explored. The VDGP method is validated using the C-MAPSS dataset and a wind turbine planetary gearbox dataset, and it is thoroughly compared with classical methods and state-of-the-art methods from the past three years. Experimental results show that the VDGP method not only achieves superior prediction accuracy but also effectively quantifies prediction uncertainty. Additionally, through SHAP analysis and correlation analysis between input features and hidden layer node outputs, low-contribution features are identified and removed, significantly reducing the number of required sensors and deployment costs. This improves computational efficiency and real-time performance, providing strong technical support for equipment health evaluation, sensor deployment optimization in industrial applications, cost reduction, and enhanced real-time monitoring and decision-making efficiency.}, archive = {J_NEUCOM}, author = {Xiuli Liu and Shuo Cui and Wan Qiao and Jianyu Liu and Guoxin Wu}, doi = {10.1016/j.neucom.2025.130232}, journal = {Neurocomputing}, month = {7}, pages = {130232}, shortjournal = {Neurocomputing}, title = {Remaining useful life prediction with uncertainty quantification for rotating machinery: A method based on explainable variational deep gaussian process}, volume = {638}, year = {2025}, }
• (2025). Bias-variance decomposition knowledge distillation for medical image segmentation. NEUCOM, 638, 130230. (www)

  @article{ , abstract = {Knowledge distillation essentially maximizes the mutual information between teacher and student networks. Typically, a variational distribution is introduced to maximize the variational lower bound. However, the heteroscedastic noises derived from this distribution are often unstable, leading to unreliable data-uncertainty modeling. Our research identifies that bias-variance coupling in knowledge distillation causes this instability. We thus propose Bias-variance dEcomposition kNowledge dIstillatioN (BENIN) approach. Initially, we use bias-variance decomposition to decouple these components. Subsequently, we design a lightweight Feature Frequency Expectation Estimation Module (FF-EEM) to estimate the student's prediction expectation, which helps compute bias and variance. Variance learning measures data uncertainty in the teacher's prediction. A balance factor addresses the bias-variance dilemma. Lastly, the bias-variance decomposition distillation loss enables the student to learn valuable knowledge while reducing noise. Experiments on Synapse and Lits17 medical-image-segmentation datasets validate BENIN's effectiveness. FF-EEM also mitigates high-frequency noise from high mask rates, enhancing data-uncertainty estimation and visualization. Our code is available at https://github.com/duanzhongjian/BENIN .}, archive = {J_NEUCOM}, author = {Xiangchun Yu and Longxiang Teng and Zhongjian Duan and Dingwen Zhang and Wei Pang and Miaomiao Liang and Jian Zheng and Liujin Qiu and Qing Xu}, doi = {10.1016/j.neucom.2025.130230}, journal = {Neurocomputing}, month = {7}, pages = {130230}, shortjournal = {Neurocomputing}, title = {Bias-variance decomposition knowledge distillation for medical image segmentation}, volume = {638}, year = {2025}, }
• (2025). Instance style-aware transformer for domain generalizable person re-identification. NEUCOM, 638, 130226. (www)

  @article{ , abstract = {Domain generalization person re-identification (DG ReID) is designed to deploy supervised training models over different source domains to unseen target domains without fine-tuning. The source domains and the target domains have significantly different domain spacing. At the same time, the existing methods overfit the source domain distribution and cannot generalize well to the unseen domains. This paper proposes a DG ReID model based on Instance Style-aware Transformer (ISTDG), focusing on cross-domain hard sampling, style exact-mixing, and attention-aware ViT to enhance the generalization ability. To implement cross-domain hard sampling and improve the efficiency of mini-batch training, we design a Cross-domain Instance-aware Hard-mining (CIH) module to mine the hard-batch and make the model to learn the domain generalization features. To enhance data augmentation and mix styles across instances in the channel dimension, we introduce a Style Exact-Mixing method (SEMix) that aligns the feature distributions of both source and target instances, thereby generating new instances that are pertinent to intra-batch style. To enable ViT to utilize the self-attention mechanism and focus on local fine-grained discriminant features associated with persons, we design an Attention-aware-based Local Feature Alignment (ALFA) module and attach it to the predicted layer of ViT. Extensive experiments on both single-source and multi-source environments verify that the performance of the proposed ISTDG significantly outperforms State-of-the-art (SOTA) methods.}, archive = {J_NEUCOM}, author = {Yingchun Guo and Shi Di and Yi Liu and Shu Chen and Xueqi Lv}, doi = {10.1016/j.neucom.2025.130226}, journal = {Neurocomputing}, month = {7}, pages = {130226}, shortjournal = {Neurocomputing}, title = {Instance style-aware transformer for domain generalizable person re-identification}, volume = {638}, year = {2025}, }
• (2025). Corrigendum to “DSTF: A diversified spatio-temporal feature extraction model for traffic flow prediction” [Neurocomputing 621 (2025) 129280]. NEUCOM, 638, 130209. (www)

  @article{ , archive = {J_NEUCOM}, author = {Xing Wang and Xiaojun Wang and Faliang Huang and Fumin Zou and Lyuchao Liao and Ruihao Zeng}, doi = {10.1016/j.neucom.2025.130209}, journal = {Neurocomputing}, month = {7}, pages = {130209}, shortjournal = {Neurocomputing}, title = {Corrigendum to “DSTF: A diversified spatio-temporal feature extraction model for traffic flow prediction” [Neurocomputing 621 (2025) 129280]}, volume = {638}, year = {2025}, }
• (2025). Visual context-aware attribute-preserving face de-identification. NEUCOM, 638, 130205. (www)

  @article{ , abstract = {Recent advancements in face de-identification using deep learning have achieved significant progress in generating realistic de-identified face images. However, most existing de-identification methods anonymize identities by inpainting the entire face or specific regions, often altering essential facial attributes such as emotion, gender, age, or ethnicity. This alteration can compromise the original purpose of collected data. While attribute-preserving de-identification methods have been proposed to address this issue, they rely on labeled data for the attributes to be preserved and require model retraining whenever the target attributes are modified. To overcome these limitations, we propose a visual context-aware attribute-preserving face de-identification method. This method leverages the semantic relationship between images and textual descriptions to identify and preserve visually significant features that encapsulate the overall context of the facial image, while effectively anonymizing identity-related information. We extract comprehensive facial attribute features through an attribute encoder trained on paired image and textual descriptions. Then, we map the extracted features onto a semantic space using a semantic map to preserve facial attributes by aligning them with the structural information of the face during de-identification. This allows for robust face de-identification while maintaining critical attributes. To assess the preservation of facial attributes such as gender, age, and race, we conduct a comparative evaluation using classification networks, confirming the effectiveness of our method in maintaining these attributes. Additionally, our method achieves 4–30 % improvement in image quality and 4–9 % in attribute similarity on CelebAMask, and 6–8 % and 1–6 % improvements, respectively, on FFHQ.}, archive = {J_NEUCOM}, author = {Hyeonwoo Kim and Jonghwa Shim and Sungwoo Park and Eenjun Hwang}, doi = {10.1016/j.neucom.2025.130205}, journal = {Neurocomputing}, month = {7}, pages = {130205}, shortjournal = {Neurocomputing}, title = {Visual context-aware attribute-preserving face de-identification}, volume = {638}, year = {2025}, }
• (2025). Reward shaping in reinforcement learning for robotic hand manipulation. NEUCOM, 638, 130204. (www)

  @article{ , abstract = {Reinforcement Learning (RL) enables robots to accomplish complex tasks. In the context of robotic hand manipulation, We present the design of an 8 Degree of Freedom (DoF) hand hardware platform and utilize PyBullet to construct a simulation environment for RL training. Robotic hand manipulation involves considering factors such as object models, specific tasks, and environmental interactions, making it a highly challenging problem. The advantage of RL is its ability to learn without requiring a model, relying solely on the definition of the reward function. While training RL demands significant computational time and engineering adjustments. To enhance the training efficiency and stability in RL, Reward Shaping is adopted to modify the reward function during training, resulting in improved training performance. When transferring the training policy to a real-world environment, it is necessary to address the differences between simulation and reality. This paper employs domain randomization to vary environmental parameters and robotic hand parameters, thus training a stable policy that can execute tasks in the real-world. With just 2 h of simulation training, we have successfully achieved the task of valve rotation in real-world scenarios, validating the effectiveness of the proposed approach in robotic hand manipulation.}, archive = {J_NEUCOM}, author = {Zelin Deng and Yunlong Dong and Xing Liu}, doi = {10.1016/j.neucom.2025.130204}, journal = {Neurocomputing}, month = {7}, pages = {130204}, shortjournal = {Neurocomputing}, title = {Reward shaping in reinforcement learning for robotic hand manipulation}, volume = {638}, year = {2025}, }
• (2025). Trajectory tracking control for wheeled mobile robot subject to slipping and skidding. NEUCOM, 638, 130203. (www)

  @article{ , abstract = {The trajectory tracking control for wheeled mobile robots (WMRs) subject to longitudinal slipping and lateral skidding is studied in this paper. Both the slipping and skidding are considered explicitly at the kinematic level. A new Radial Basis Function Neural Network (RBFNN) based disturbance observer is proposed to estimate the slipping and skidding, and the slipping and skidding estimation errors can be made arbitrarily small. By integrating the disturbance observer, a kinematic controller is designed based on a new Lyapunov function. Then the dynamic controller is designed and the uniform ultimate boundedness of the overall closed-loop system is proved. Finally, simulations are given to verify the effectiveness and advantages of the method proposed in this paper.}, archive = {J_NEUCOM}, author = {Jianjun Bai and Chen Liu and Shengzhe Zhou and Yun Chen and Xujun Li}, doi = {10.1016/j.neucom.2025.130203}, journal = {Neurocomputing}, month = {7}, pages = {130203}, shortjournal = {Neurocomputing}, title = {Trajectory tracking control for wheeled mobile robot subject to slipping and skidding}, volume = {638}, year = {2025}, }
• (2025). Auto-tuning strategy for deep koopman robust model predictive control design based on advanced metaheuristics. NEUCOM, 638, 130197. (www)

  @article{ , abstract = {Nonlinear systems are challenging to control due to their complexity and unpredictability, with existing methods often struggling to balance accuracy and computational efficiency. This paper addresses these challenges by proposing an auto-tuning strategy that integrates Koopman operator theory with model predictive control (MPC), optimized using metaheuristic algorithms. We use deep neural networks to learn the Koopman embedding function and operator, transforming the nonlinear system into an equivalent linear system in the embedding space. To enhance rapid setpoint tracking, metaheuristic algorithms automatically tune MPC parameters within the Koopman framework. Additionally, acknowledging deviations between the Koopman model and the actual nonlinear system, we implement robust constraint-tightening methods to ensure stable convergence of the system’s state to the setpoint within a bounded error. Simulation results using a nonlinear iron removal process validate the effectiveness of the proposed approach. While this study focuses on offline optimization, the proposed framework can be extended to real-time applications in future work, addressing challenges like computational efficiency and model accuracy.}, archive = {J_NEUCOM}, author = {Zijia Meng and Yu Xiao and Xiaodong Xu and Biao Luo and Yuncheng Du and Weihua Gui}, doi = {10.1016/j.neucom.2025.130197}, journal = {Neurocomputing}, month = {7}, pages = {130197}, shortjournal = {Neurocomputing}, title = {Auto-tuning strategy for deep koopman robust model predictive control design based on advanced metaheuristics}, volume = {638}, year = {2025}, }
• (2025). Conditional attention guided normalizing flow for low-light image enhancement. NEUCOM, 638, 130196. (www)

  @article{ , abstract = {Low-light imaging is inherently challenging because images captured under insufficient illumination often suffer from poor quality. In this paper, we introduce a new invertible flow framework, which maps low-light images to latent data distribution. Then the parameter of the conditional distribution can be estimated by the proposed attention-flow network. Specifically, the network combines multi-scale attention mechanisms with flow-based generative models to enhance low-light images. Owing to the property of invertibility in the network, it effectively captures and reconstructs crucial visual details and contextual information, producing images that are perceptually realistic and free from artifacts. Experimental results on several public low-light datasets demonstrate the superior performance of the proposed method compared to existing approaches, highlighting its robustness across a range of low-light conditions. The source code is available at https://github.com/Rehgar3/AttnFlow .}, archive = {J_NEUCOM}, author = {Chenyu Wang and Lin Ma and Hanlin Qin and Shuowen Yang and Ruiyun Li and Xiaotao Shi}, doi = {10.1016/j.neucom.2025.130196}, journal = {Neurocomputing}, month = {7}, pages = {130196}, shortjournal = {Neurocomputing}, title = {Conditional attention guided normalizing flow for low-light image enhancement}, volume = {638}, year = {2025}, }
• (2025). Generalizing medical image representations via quaternion wavelet networks. NEUCOM, 638, 130195. (www)

  @article{ , abstract = {Neural network generalizability is becoming a broad research field due to the increasing availability of datasets from different sources and for various tasks. This issue is even wider when processing medical data, where a lack of methodological standards causes large variations being provided by different imaging centers or acquired with various devices and cofactors. To overcome these limitations, we introduce a novel, generalizable, data- and task-agnostic framework able to extract salient features from medical images. The proposed quaternion wavelet network (QUAVE) can be easily integrated with any pre-existing medical image analysis or synthesis task, and it can be involved with real, quaternion, or hypercomplex-valued models, generalizing their adoption to single-channel data. QUAVE first extracts different sub-bands through the quaternion wavelet transform, resulting in both low-frequency/approximation bands and high-frequency/fine-grained features. Then, it weighs the most representative set of sub-bands to be involved as input to any other neural model for image processing, replacing standard data samples. We conduct an extensive experimental evaluation comprising different datasets, diverse image analysis, and synthesis tasks including reconstruction, segmentation, and modality translation. We also evaluate QUAVE in combination with both real and quaternion-valued models. Results demonstrate the effectiveness and the generalizability of the proposed framework that improves network performance while being flexible to be adopted in manifold scenarios and robust to domain shifts. The full code is available at: https://github.com/ispamm/QWT .}, archive = {J_NEUCOM}, author = {Luigi Sigillo and Eleonora Grassucci and Aurelio Uncini and Danilo Comminiello}, doi = {10.1016/j.neucom.2025.130195}, journal = {Neurocomputing}, month = {7}, pages = {130195}, shortjournal = {Neurocomputing}, title = {Generalizing medical image representations via quaternion wavelet networks}, volume = {638}, year = {2025}, }
• (2025). MDVL-edit: Mask-assisted highly disentangled text-driven face image editing based on vision-language alignment. NEUCOM, 638, 130194. (www)

  @article{ , abstract = {Text-driven face image manipulation has always been a challenging task. The previous methods are highly dependent on annotated training data and have poor generalization and generality. More and more methods are now being studied with the help of the powerful image-text representation of the CLIP model and various decoupled latent spaces of StyleGAN. This paper finds and expounds the common problems exposed by a series of methods to find a specific editing direction in StyleSpace (channel-wise style parameters space), the most disentangled latent space of StyleGAN at present: human face is a complex muscle group area and can suffer from strong coupling of the five senses, such as the eye region, or even coupling with clothing. We analyzed the root cause of the problem and determined that the problem lies in the fact that the CLIP model suffers from alignment and hallucination problems. But tuning a model trained with large-scale data is tedious, so we proposed our simple and efficient solution based on the CLIP model, namely the mask-assisted MDVL-Edit method based on vision-language alignment. First, the training phase maps the CLIP image feature embedding differences to the editing direction of the StyleGAN latent space. Secondly, the inference stage uses the embedding differences of text features to match the embedding differences of image features through the intermediate embedding space Delta CLIP, predicts the editing direction, and then uses the R s matrix and the Auto Mask module to optimize the editing results. In this way, the decoupling operation is carried out at the latent code level and the image level respectively, which solves the problem of strong coupling on some fine-grained attributes in classical methods. Our method is a type of text-free training approach that utilizes only pure image data. Finally, the experiments prove that our method obtains the best results under the widely used and authoritative FFHQ and CelebA-HQ face datasets, achieving the highest efficiency as well as realism at the least cost of time.}, archive = {J_NEUCOM}, author = {Haowei Hong and Zhi Zheng and Wenjun Huang and Xiaomin Lin}, doi = {10.1016/j.neucom.2025.130194}, journal = {Neurocomputing}, month = {7}, pages = {130194}, shortjournal = {Neurocomputing}, title = {MDVL-edit: Mask-assisted highly disentangled text-driven face image editing based on vision-language alignment}, volume = {638}, year = {2025}, }
• (2025). Resilient synchronization of interconnected dynamical networks with DoS attacks: An adaptive event-triggered control approach. NEUCOM, 638, 130192. (www)

  @article{ , abstract = {This paper is concerned with resilience and efficiency of a class of interconnected dynamical networks (IDNs) under limited communication resources and potential cyber-attacks. We first propose a novel framework for synchronization control in IDNs that integrates distributed event-triggered communication with resilience against denial-of-service (DoS) attacks. We start with developing an adaptive event-triggered mechanism (AETM), dependent on the attack parameter, to dynamically adjust threshold parameters for each node based on network conditions and severity of DoS attacks. Then, we design DoS-resilient event-triggered distributed synchronization controllers for each node to achieve the desired synchronization objective, even under sporadic event-based communication and disruptions caused by DoS attacks. Third, we establish co-design criteria to jointly determine the gain matrices of both the AETMs and synchronization controllers. These criteria provide valuable insights into the interplay between communication quality (e.g., triggering frequency and DoS severity) and control quality (e.g., convergence rate). Finally, we validate the effectiveness of the proposed approach through two illustrative examples.}, archive = {J_NEUCOM}, author = {Yao Cui and Xiaohua Ge and Pei Cheng and Xin Liu}, doi = {10.1016/j.neucom.2025.130192}, journal = {Neurocomputing}, month = {7}, pages = {130192}, shortjournal = {Neurocomputing}, title = {Resilient synchronization of interconnected dynamical networks with DoS attacks: An adaptive event-triggered control approach}, volume = {638}, year = {2025}, }
• (2025). Goal-driven navigation via variational sparse q network and transfer learning. NEUCOM, 638, 130191. (www)

  @article{ , abstract = {Compared to traditional map-based, goal-driven navigation methods, deep reinforcement learning (DRL)-based goal-driven navigation for mobile robots offers the advantage of not relying on prior map information, it enables autonomous decision-making through continuous interaction with the environment. However, DRL-based goal-driven navigation faces significant challenges in terms of low generalization ability and learning inefficiency. In this paper, we propose a DRL approach, variational sparsity Q network (VSQN), which leverages variational inference and transfer learning to achieve efficient goal-driven navigation. The variational inference framework models weight uncertainty within the network, thereby enhancing the agent’s generalization capability. Furthermore, a hierarchical learning network framework is adopted, and transfer learning is employed to incorporate prior knowledge from a pre-trained model into new navigation tasks. This enables the agent to rapidly adapt to novel tasks without the need for fine-tuning after selecting an optimal sub-goal. This improves the agent’s initial performance in previously unseen navigation tasks. The experimental results indicate that the proposed method achieves a success rate (SR) of 76% and a success weighted by inverse path length (SPL) of 0.52 in previously unencountered environments and target locations within the grid environment, and an SR of 81% with an SPL of 0.30 in the AI2-THOR environment. These findings demonstrate that the method substantially enhances the agent’s generalization capability.}, archive = {J_NEUCOM}, author = {Jiacheng Yao and Li He and Hongwei Wang and Wendong Xiao and Yangjun Du and Zhaoqing Lu and Yuxin Liao}, doi = {10.1016/j.neucom.2025.130191}, journal = {Neurocomputing}, month = {7}, pages = {130191}, shortjournal = {Neurocomputing}, title = {Goal-driven navigation via variational sparse q network and transfer learning}, volume = {638}, year = {2025}, }
• (2025). CR-FND: Robustness-enhanced multimodal fake news detection against content variance. NEUCOM, 638, 130190. (www)

  @article{ , abstract = {The accelerated development and expansion of social networking services have posed a significant challenge for fake news detection (FND). However, the prevailing approaches—content-based and propagation-based methods are studied separately and the integration of these complementary perspectives remains underexplored. They overlook the correlations between the news semantic information and its propagation structure, which could be served as auxiliary detection information to enhance model’s robustness against content variance. In this article, we propose a multimodal C ontent- R obust FND (CR-FND) framework based on contrastive learning (CL) to align the overall semantic information with the structural features of news propagation. Specifically, CR-FND enhances the efficiency of CL between propagation graph and multimodal semantic feature by designing Multi-View data augmentation method and optimizing the architecture of graph transformer and sampling strategy, which mitigate the impact of detailed information in news content that has little correlation with the propagation structure. Additionally, we incorporate emotional and thematic features during semantic multimodal fusion and introduce user profiles in construction of the propagation structure graph, thereby strengthening their correlations. In this way, the inconsistency between the news semantic information and propagation structure serves as auxiliary detection information, enhancing the model’s generalization ability in real-world scenarios, where the detection model is required to identify news from different sources and across different domains. Extensive experiments on five real-world datasets demonstrate that the CR-FND outperforms other state-of-art models for FND and exhibits significant improvement in robustness.}, archive = {J_NEUCOM}, author = {Tiansen Niu and Suchuan Ma and Dapeng Yan and Qingshu Guan and Peiqi Hou and Jiarun Sun and Hui Cao}, doi = {10.1016/j.neucom.2025.130190}, journal = {Neurocomputing}, month = {7}, pages = {130190}, shortjournal = {Neurocomputing}, title = {CR-FND: Robustness-enhanced multimodal fake news detection against content variance}, volume = {638}, year = {2025}, }
• (2025). Distributed event-triggered sliding-mode control of second-order multi-UAV system with a dynamic leader. NEUCOM, 638, 130189. (www)

  @article{ , abstract = {This article delves into the issue of formation control for multi-UAV system considering dynamic inputs. It introduces a distributed event-triggered sliding mode control design for second-order multi-UAV systems, grounded on leader-following principles. Initially, contemplate the utilization of sliding mode control methods to tackle disturbances and uncertainties within the formation system, ensuring robustness against external interference. Additionally, event-triggered strategies are employed, reducing unnecessary control actions and conserving energy. This minimizes the network communication load and enhances system response time. Finally, an appropriate consensus algorithm is proposed, designed on the basis of event-triggered control, to ensure the consistency tracking effect of each UAV. The proposed controller’s stability is confirmed by Lyapunov’s stability theory, showcasing its capability to stabilize the system. Additionally, a triggering scheme is introduced to efficiently reduce state updates and eliminate Zeno behavior. Lastly, simulations of a formation problem involving six quadrotor UAVs are presented, confirming the practicality of the theoretical findings.}, archive = {J_NEUCOM}, author = {Fuyin Yao and Hongwei Ren}, doi = {10.1016/j.neucom.2025.130189}, journal = {Neurocomputing}, month = {7}, pages = {130189}, shortjournal = {Neurocomputing}, title = {Distributed event-triggered sliding-mode control of second-order multi-UAV system with a dynamic leader}, volume = {638}, year = {2025}, }
• (2025). Stress assessment with EEG and machine learning in affective VR environments. NEUCOM, 638, 130185. (www)

  @article{ , abstract = {Stress is a reaction that occurs when a person perceives, with or without awareness, an imbalance between requests and available resources. Relying on this definition, we have carried out an experiment in a Virtual Reality environment to elicit (light) stress in the user and analyze the emotional responses with electroencephalography (EEG). The virtual environment is divided in eight parts; in each of them a stressor has been put in action, meaning that in every part the participants perform a task, but a specific resource is missing (time, knowledge, control, salvation, no or too many alternatives, engagement, self-confidence). EEG is used to assess the emotional response with the aid of Valence/Arousal/Dominance/Stress indicators presented in previous literature. Nine indicators calculated for 87 participants, labeled according to self-assessment replies (post-experimental questionnaires), were classified with eXtreme Gradient Boosting, k-Nearest Neighbor, Support Vector Machine and Random Forest classifiers. The lowest results in terms of accuracy were obtained with k-Nearest Neighbor (around 70 %), whilst the highest ones were obtained with eXtreme Gradient Boosting and Random Forest (above 98 %), showing that EEG could be a valuable tool to assess the emotional response in stressful situations, with a particular focus on the Stress indicators.}, archive = {J_NEUCOM}, author = {Federica Marcolin and Elena Carlotta Olivetti and Ivonne Angelica Castiblanco Jimenez and Giorgia Passavanti and Sandro Moos and Enrico Vezzetti and Alessia Celeghin}, doi = {10.1016/j.neucom.2025.130185}, journal = {Neurocomputing}, month = {7}, pages = {130185}, shortjournal = {Neurocomputing}, title = {Stress assessment with EEG and machine learning in affective VR environments}, volume = {638}, year = {2025}, }
• (2025). Dynamic and prioritized task scheduling of heterogeneous multi-robot systems using deep reinforcement learning. NEUCOM, 638, 130184. (www)

  @article{ , abstract = {With the ongoing development of heterogeneous multi-robot systems technology, the challenges related to scheduling extended sequences of dynamic and prioritized tasks have become progressively more pronounced. To address this issue, this paper proposes an online task scheduling framework for efficient collaboration in heterogeneous multi-robot systems. First, a collaboration strategy for heterogeneous multi-robot systems is proposed to facilitate efficient parallel work, incorporating online updates to task sequences and priorities. Second, a deep reinforcement learning based dynamic task scheduling method employing actor–critic attention mechanisms is proposed for online optimal task sequence generation. In addition, task priority is integrated into the objective function, encouraging the rapid execution of high-priority tasks. Experimental results with randomly generated tasks of varying scales demonstrate the flexibility, efficiency and generalization of the method. Task sequences for the heterogeneous robots are generated online as the task set to be executed evolves, optimizing the global travel distances while ensuring the timely execution of high-priority tasks.}, archive = {J_NEUCOM}, author = {Jiabing Zhang and Qingxuan Jia and Shiyu Zhang and Gang Chen}, doi = {10.1016/j.neucom.2025.130184}, journal = {Neurocomputing}, month = {7}, pages = {130184}, shortjournal = {Neurocomputing}, title = {Dynamic and prioritized task scheduling of heterogeneous multi-robot systems using deep reinforcement learning}, volume = {638}, year = {2025}, }
• (2025). IShapeInst: A novel and effective instance segmentation for irregularly shaped objects. NEUCOM, 638, 130183. (www)

  @article{ , abstract = {Instance segmentation of irregularly shaped objects presents significant challenges. Existing methods struggle to generate accurate representations for such objects, leading to suboptimal segmentation performance. This study introduces iShapeInst, a novel region-based instance segmentation method designed to address these challenges. iShapeInst dynamically generates convolution kernels based on the position and shape to represent the irregular objects. Specifically, the method begins by using semantic segmentation to predict the regions of all objects and generates multiple convolution kernels corresponding to these regions. Then, metric learning is employed to cluster the kernels for the same instance while separating between different instances. Additionally, kernel fusion is utilized to select high-quality kernels for each instance, reducing resource consumption. Instance masks are subsequently created by convolving the mask feature map with these high-quality kernel weights. Finally, an optimal instance mask is derived using Matrix NMS. Experiments demonstrate that iShapeInst achieves an mAP (mean Average Precision) of 85.2% on the iShape dataset, outperforming existing methods by 13.5% mAP, with a processing speed of 31.5 FPS. Furthermore, our method achieves competitive performance on the Surgical Instruments Dataset and arbitrary-shaped scene text datasets, demonstrating its strong generalization capability. The code is available at https://github.com/youngnuaa/iShapeInst .}, archive = {J_NEUCOM}, author = {Yang Wang and Wanlin Zhou and Jiakai Zhou}, doi = {10.1016/j.neucom.2025.130183}, journal = {Neurocomputing}, month = {7}, pages = {130183}, shortjournal = {Neurocomputing}, title = {IShapeInst: A novel and effective instance segmentation for irregularly shaped objects}, volume = {638}, year = {2025}, }
• (2025). WM-SORT: Modeling multi-object tracking motion patterns in real world. NEUCOM, 638, 130182. (www)

  @article{ , abstract = {Multi-Object Tracking (MOT) is a pivotal yet challenging computer vision task, particularly in scenarios involving occlusion. While the Kalman filter serves as a predominant motion model for predicting and refining object locations, its reliance can lead to fast accumulation of state errors during occlusions, resulting in identity switch issues. Most methods primarily focus on modeling non-linear pixel-level motions by mining bounding box information, which is inherently limited. To address this issue, we propose World Motion SORT (WM-SORT), an innovative MOT framework which models the motion patterns of objects in the three-dimensional world coordinate system instead of only on the two-dimensional image plane. WM-SORT is designed to mitigate the identity switch problem caused by occlusion. Specifically, we first introduce Linear Camera Model to project the pixel coordinates to the world coordinate system, and Dimension Recovery Module is deployed to obtain the lost dimension due to the camera characteristics. To model the motion patterns in real world, we develop World Motion Model which updates the locations of objects in the world coordinate system when objects are occluded. It simulates the movement of objects in the real world rather than only using the information of bounding boxes. In addition, we propose a 2D-3D combined data association to match the tracked objects not only in the image plane but also in the three-dimensional world. Extensive experiments on MOT17 and MOT20 datasets demonstrate that our proposed WM-SORT achieves outstanding results, particularly in identity preservation.}, archive = {J_NEUCOM}, author = {Yizhuo Jiang and Weiyu Zhao and Yan Gao and Xinhang Niu and Jie Li and Xinbo Gao}, doi = {10.1016/j.neucom.2025.130182}, journal = {Neurocomputing}, month = {7}, pages = {130182}, shortjournal = {Neurocomputing}, title = {WM-SORT: Modeling multi-object tracking motion patterns in real world}, volume = {638}, year = {2025}, }
• (2025). SOH estimation of lithium-ion batteries subject to partly missing data: A Kolmogorov–Arnold–Linformer model. NEUCOM, 638, 130181. (www)

  @article{ , abstract = {Accurate estimation of the state of health (SOH) is crucial for improving the safety and reliability of lithium-ion batteries. However, sensor measurements inevitably suffer from incomplete data due to sensor failures caused by factors such as component aging in practical applications. To address this issue, a joint estimation Kolmogorov–Arnold–Linformer (KAL) network model is proposed. Specifically, the Kolmogorov–Arnold Network (KAN) module is employed to replace the Multi-layer Perceptrons (MLP) module in the Linformer model, which enhances the representation of nonlinear features and improves the overall accuracy of the model. A dual-training model approach for SOH estimation is designed, which integrates the health feature (HF)-to-capacity model to infer trends in capacity changes using historical data. Based on these inferred trends, the model is trained to achieve accurate SOH estimation in scenarios with partly missing data. Validation on the publicly available Toyota-MIT-Stanford dataset demonstrates that, compared with other common deep learning methods, the KAL network model exhibits superior accuracy and reliability in scenarios with varying rates of partly missing data.}, archive = {J_NEUCOM}, author = {Liyuan Shao and Yong Zhang and Xiujuan Zheng and Rui Yang and Wei Zhou}, doi = {10.1016/j.neucom.2025.130181}, journal = {Neurocomputing}, month = {7}, pages = {130181}, shortjournal = {Neurocomputing}, title = {SOH estimation of lithium-ion batteries subject to partly missing data: A Kolmogorov–Arnold–Linformer model}, volume = {638}, year = {2025}, }
• (2025). Efficient energy-based neural topic modeling with embeddings. NEUCOM, 638, 130179. (www)

  @article{ , abstract = {Recently, topic models based on deep neural networks have garnered significant attention due to their robust ability to capture latent topic structures. However, existing neural topic models face challenges such as approximate posterior inference errors or the computational intensity of iterative transport distance calculations. Additionally, a common limitation of these models is their dependence on pre-defined priors to constrain latent topic structures. This reliance can limit the expressive capacity of topic variables and reduce topic heterogeneity across diverse corpora. To address these issues, we propose an efficient embedding-derived topic model framework that captures the interrelations among documents, topics, and words through embedding agreements, enabling faster training. Furthermore, we introduce a learnable topic prior constraint for latent topic structures, formulated through an energy function, and design an effective joint optimization for both embeddings and the learnable topic prior. Extensive experiments on six widely-used datasets demonstrate that our approach outperforms existing state-of-the-art topic models.}, archive = {J_NEUCOM}, author = {Yanyan Liu and Puiieng Lei and Yijun Sheng and Ximing Chen and Zhiguo Gong}, doi = {10.1016/j.neucom.2025.130179}, journal = {Neurocomputing}, month = {7}, pages = {130179}, shortjournal = {Neurocomputing}, title = {Efficient energy-based neural topic modeling with embeddings}, volume = {638}, year = {2025}, }
• (2025). A robust high-resolution remote sensing image hiding network. NEUCOM, 638, 130178. (www)

  @article{ , abstract = {High-resolution remote sensing imagery is indispensable for applications ranging from environmental monitoring to national defense, where secure data transmission is paramount. In this study, we introduce a novel steganography network specifically engineered for embedding remote sensing images within natural images, addressing the longstanding challenges of achieving high payload capacity, imperceptibility, and robustness. Our approach leverages an encoder-decoder architecture with independent encryption and decryption modules to reduce computational overhead and accelerate convergence. Key innovations include an enhanced Residual Convolutional Block Attention Module (Res-CBAM) and a Discrete Wavelet Transform (DWT)-based frequency-domain transformation, which together optimize feature extraction and reduce the visibility of secret image contours. A dual-stage training process further refines the model, with the first stage ensuring high-quality encryption and decryption and the second stage mitigating rounding errors to enable accurate recovery under lossy conditions. Comprehensive evaluations on standard benchmark datasets (DIV2K, COCO, and ImageNet) and our custom remote sensing-natural image dataset demonstrate that our method consistently outperforms 12 state-of-the-art techniques in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), while achieving steganalysis detection accuracy near random guessing levels. These results underscore the model’s superior performance, enhanced security, and broad applicability, making it a promising solution for robust, high-resolution image hiding in real-world scenarios.}, archive = {J_NEUCOM}, author = {Shiyuan Wang and Fumin Wang and Mingqiang Guo and Ying Huang and Wei Cao}, doi = {10.1016/j.neucom.2025.130178}, journal = {Neurocomputing}, month = {7}, pages = {130178}, shortjournal = {Neurocomputing}, title = {A robust high-resolution remote sensing image hiding network}, volume = {638}, year = {2025}, }
• (2025). High-feasibility parafoil dynamic formation real-time adaptive trajectory planning based on ABC-MASAC in complex environments. NEUCOM, 638, 130176. (www)

  @article{ , abstract = {In large-scale parafoil formation airdrop missions, unpowered parafoils face challenges in precise speed control and are highly susceptible to wind disturbances. Traditional fixed formation trajectories require precise speed control and cannot adapt in real-time to environmental changes. Furthermore, dynamic models based on rigid-body assumptions fail to accurately reflect the aerodynamic characteristics of actual systems, resulting in poor trajectory feasibility. To address the aforementioned issues, this paper proposes a high-feasibility, real-time adaptive trajectory planning method for parafoil airdrop dynamic formations. First, a high-precision flexible eight-degree-of-freedom (8-DOF) dynamic model is developed using fluid–structure interaction (FSI) techniques, which can simulate the real aerodynamic deformation and dynamic response of the parafoil, ensuring the planned trajectories are more consistent the behavior of actual systems. Additionally, an Adaptive Behavior Cloning Multi-Agent Soft Actor–Critic (ABC-MASAC) algorithm is proposed, which accelerates training by adaptively adjusting the follower policy, thereby improving the training efficiency of the trajectory planning model. Detailed simulation results demonstrate that the proposed parafoil airdrop formation trajectory planning method achieves high trajectory feasibility and real-time planning capability, while significantly accelerating training speed.}, archive = {J_NEUCOM}, author = {Jiaming Yu and Hao Sun and Qinglin Sun and Hong Zhu and Mingwei Sun and Zhengxiang Jin and Zengqiang Chen}, doi = {10.1016/j.neucom.2025.130176}, journal = {Neurocomputing}, month = {7}, pages = {130176}, shortjournal = {Neurocomputing}, title = {High-feasibility parafoil dynamic formation real-time adaptive trajectory planning based on ABC-MASAC in complex environments}, volume = {638}, year = {2025}, }
• (2025). Missing data as augmentation in the earth observation domain: A multi-view learning approach. NEUCOM, 638, 130175. (www)

  @article{ , abstract = {Multi-view learning (MVL) leverages multiple sources or views of data to enhance machine learning model performance and robustness. This approach has been successfully used in the Earth Observation (EO) domain, where views have a heterogeneous nature and can be affected by missing data. Despite the negative effect that missing data has on model predictions, the ML literature has used it as an augmentation technique to improve model generalization, like masking the input data. Inspired by this, we introduce novel methods for EO applications tailored to MVL with missing views. Our methods integrate the combination of a set to simulate all combinations of missing views as different training samples. Instead of replacing missing data with a numerical value, we use dynamic merge functions, like average, and more complex ones like Transformer. This allows the MVL model to entirely ignore the missing views, enhancing its predictive robustness. We experiment on four EO datasets with temporal and static views, including state-of-the-art methods from the EO domain. The results indicate that our methods improve model robustness under conditions of moderate missingness, and improve the predictive performance when all views are present. The proposed methods offer a single adaptive solution to operate effectively with any combination of available views.}, archive = {J_NEUCOM}, author = {Francisco Mena and Diego Arenas and Andreas Dengel}, doi = {10.1016/j.neucom.2025.130175}, journal = {Neurocomputing}, month = {7}, pages = {130175}, shortjournal = {Neurocomputing}, title = {Missing data as augmentation in the earth observation domain: A multi-view learning approach}, volume = {638}, year = {2025}, }
• (2025). Combination of phrase matchings based cross-modal retrieval. NEUCOM, 638, 130174. (www)

  @article{ , abstract = {Fine-grained cross-modal retrieval is a prominent research focus in the fields of information retrieval and multi-modal learning. Existing methods lack local fragment matching for image–text semantic similarity. This results in poor performance when measuring local similarity between image regions and words. Furthermore, they fail to explore relationships between local fragments on a larger scale. As a result, models struggle with distinguishing complex scenarios where identical local fragments have differing relationships. Therefore, we propose Combination of Phrase Matchings (CPM) based cross-modal retrieval. Specifically, we introduce stacked GCNs at different layers to model non-neighboring phrases within various ranges. This is done to compute image–text semantic similarity at complex scenario level. Simultaneously, we combine local fragment matching, neighborhood phrase matching and multiple non-neighboring phrase matching (referred to collectively as phrase matching) to more comprehensively reflect image–text semantic similarity. We conducted extensive experiments on the MS-COCO and Flicker30K datasets, which demonstrated the superior performance of CPM. The experiments also highlighted the roles of local fragment matching, neighboring phrase matching, and non-neighboring phrase matching at different scales in enhancing image–text matching.}, archive = {J_NEUCOM}, author = {Li Zhang and Yahu Yang and Shuheng Ge and Guanghui Sun}, doi = {10.1016/j.neucom.2025.130174}, journal = {Neurocomputing}, month = {7}, pages = {130174}, shortjournal = {Neurocomputing}, title = {Combination of phrase matchings based cross-modal retrieval}, volume = {638}, year = {2025}, }
• (2025). Topic-aware most influential community search in social networks. NEUCOM, 638, 130173. (www)

  @article{ , abstract = {Influential community search (ICS) finds a set of densely connected and high-impact vertices from a social network. Although great effort has been devoted to ICS problems, most existing methods do not consider how relevant the influential community found is to specific topics. A few attempts at topic-aware ICS problems cannot capture the stochastic nature of community formation and influence propagation in social networks. To address these issues, we introduce a novel problem of topic-aware most influential community search ( TAMICS ) to discover a set of vertices such that for a given topic vector q , they induce a ( k , l , η ) -core in an uncertain directed interaction graph and have the highest influence scores under the independent cascade (IC) model. We propose an online algorithm to provide an approximate result for any TAMICS query with bounded errors. Furthermore, we design two index structures and an index-based heuristic algorithm for efficient TAMICS query processing. Finally, we experimentally evaluate the efficacy and efficiency of our proposed approaches on various real-world datasets. The results show that (1) the communities of TAMICS have higher relevance and social influence w.r.t. the query topics as well as structural cohesiveness than those of several state-of-the-art topic-aware and influential CS methods and (2) the index-based algorithm achieves speed-ups of up to three orders of magnitude over the online algorithm with an affordable overhead for index construction.}, archive = {J_NEUCOM}, author = {Long Teng and Yanhao Wang and Zhe Lin and Fei Yu}, doi = {10.1016/j.neucom.2025.130173}, journal = {Neurocomputing}, month = {7}, pages = {130173}, shortjournal = {Neurocomputing}, title = {Topic-aware most influential community search in social networks}, volume = {638}, year = {2025}, }
• (2025). Multivariate spatiotemporal clustering based on spatial toplitz inverse covariance with adaptive information aggregation. NEUCOM, 638, 130172. (www)

  @article{ , abstract = {Spatiotemporal clustering is an important technique for discovering potential and useful patterns in many areas including public safety, ecology, epidemiology, earth sciences, etc. However, many spatiotemporal clustering methods fail to effectively capture the complex correlations between sequences in spatiotemporal data, which leads to clustering results that are spatially scattered. Hence, we propose a multivariate spatiotemporal clustering method based on spatial Toplitz Inverse Covariance with Adaptive Information Aggregation (TIC-AIA). Specifically, in the temporal dimension, we extend Gaussian Markov Random Field to extract temporal features from multivariate spatiotemporal data. In the spatial dimension, a new adaptive K-nearest neighbor algorithm is proposed to construct subregions of different sizes for each location. Then, an adaptive information aggregation mechanism is employed to integrate temporal features and spatial correlations to obtain the spatiotemporal information of subregions, which are applied to the spatial Toeplitz inverse covariance structure for clustering multivariate spatiotemporal data. Additionally, a new variable importance metric is developed to identify unimportant variables to optimize the clustering accuracy. Experimental results show that TIC-AIA outperforms existing techniques in clustering performance across various synthetic and real-world multivariate spatiotemporal datasets.}, archive = {J_NEUCOM}, author = {Ling Wang and Gang Wang}, doi = {10.1016/j.neucom.2025.130172}, journal = {Neurocomputing}, month = {7}, pages = {130172}, shortjournal = {Neurocomputing}, title = {Multivariate spatiotemporal clustering based on spatial toplitz inverse covariance with adaptive information aggregation}, volume = {638}, year = {2025}, }
• (2025). Quantile pattern regression using twin extreme learning machines with smooth pinball loss. NEUCOM, 638, 130171. (www)

  @article{ , abstract = {To address the current limitation of single value prediction models in ensuring high prediction accuracy, this paper introduces a novel smooth pinball loss twin extreme learning machine quantile regression (STELMQR) model. STELMQR incorporates quantile parameters into two non-parallel hyperplanes, effectively capturing the asymmetric and heterogeneous distribution characteristics of data points across different quantile levels. By utilizing smooth pinball quantile loss, the model minimizes the impact of noise, enhancing its robustness. Furthermore, STELMQR leverage a dual coordinate descent algorithm, which accelerates the model training process and optimizes fewer variables compared to other quantile regression models. To validate the efficacy of the proposed models, numerical experiments were conducted on three types of artificial datasets and nine UCI benchmark datasets. The results demonstrate the validity and superiority of STELMQR in handling complex data distributions and improving prediction accuracy.}, archive = {J_NEUCOM}, author = {Haoyu Wang and Guolin Yu and Jun Ma}, doi = {10.1016/j.neucom.2025.130171}, journal = {Neurocomputing}, month = {7}, pages = {130171}, shortjournal = {Neurocomputing}, title = {Quantile pattern regression using twin extreme learning machines with smooth pinball loss}, volume = {638}, year = {2025}, }
• (2025). Multimodal understanding of human values in videos: A benchmark dataset and PLM-based method. NEUCOM, 638, 130170. (www)

  @article{ , abstract = {Multimodal content has become the mainstream communication medium in video sharing platforms such as TikTok and Twitter, containing rich values information. Understanding human values is of great significance to multimodal content analysis and can be applied to downstream tasks such as recommendation systems and value alignment. However, current studies on human values mainly focus on text and lack a multimodal perspective. In this work, we present a new multimodal human values video dataset called VVALUES, which contains 5,104 annotated videos along with their titles. The dataset is labeled with coarse-grained polarity tags of positive and neutral, and fine-grained tags including 13 classes of value vocabulary. Based on VVALUES, we further develop a pre-trained language model (PLM)-based multimodal method adopting a dual-transformer variant for value recognition, MMVR. Extensive experiments demonstrate that our method significantly improves the performance of understanding values in videos. To the best of our knowledge, we are the first to try to incorporate human values in video understanding, and VVALUES is the first multimodal video dataset for human values.}, archive = {J_NEUCOM}, author = {Yuchen Wang and Zhulin Tao and He Chang and Nanxin Huang and Libiao Jin and Xiaofang Luo}, doi = {10.1016/j.neucom.2025.130170}, journal = {Neurocomputing}, month = {7}, pages = {130170}, shortjournal = {Neurocomputing}, title = {Multimodal understanding of human values in videos: A benchmark dataset and PLM-based method}, volume = {638}, year = {2025}, }
• (2025). SegmATRon: Embodied adaptive semantic segmentation for indoor environment. NEUCOM, 638, 130169. (www)

  @article{ , abstract = {The state-of-the-art methods for computer vision are often trained with large amounts of data collected from static cameras. In contrast, an embodied intelligent agent can interact with a continuous environment to improve the perception quality. Previous methods for embodied computer vision have not considered the task of semantic segmentation. This paper first introduces an adaptive transformer model for embodied image semantic segmentation named SegmATRon. Its distinctive feature is the adaptation of model weights during inference on several images using a hybrid multicomponent loss function. We studied this model on datasets collected in the photorealistic Habitat and the synthetic AI2-THOR simulators. We showed that obtaining additional images using the agent’s actions in an indoor environment can improve the quality of semantic segmentation.}, archive = {J_NEUCOM}, author = {Tatiana Zemskova and Margarita Kichik and Dmitry Yudin and Aleksei Staroverov and Aleksandr Panov}, doi = {10.1016/j.neucom.2025.130169}, journal = {Neurocomputing}, month = {7}, pages = {130169}, shortjournal = {Neurocomputing}, title = {SegmATRon: Embodied adaptive semantic segmentation for indoor environment}, volume = {638}, year = {2025}, }
• (2025). MCI-GRU: Stock prediction model based on multi-head cross-attention and improved GRU. NEUCOM, 638, 130168. (www)

  @article{ , abstract = {As financial markets become increasingly complex and the era of big data unfolds, accurate stock prediction has become more critical. Although traditional time series models, such as GRU, have been widely applied to stock prediction, they still exhibit limitations in addressing the intricate nonlinear dynamics of markets, particularly in the flexible selection and effective utilization of key historical information. In recent years, emerging methods like Graph Neural Networks and Reinforcement Learning have shown significant potential in stock prediction. However, these methods often demand high data quality and quantity, and they tend to exhibit instability when dealing with data sparsity and noise. Moreover, the training and inference processes for these models are typically complex and computationally expensive, limiting their broad deployment in practical applications. Existing approaches also generally struggle to capture unobservable latent market states effectively, such as market sentiment and expectations, microstructural factors, and participant behavior patterns, leading to an inadequate understanding of market dynamics and subsequently impact prediction accuracy. To address these challenges, this paper proposes a stock prediction model, MCI-GRU, based on a multi-head cross-attention mechanism and an improved GRU. First, we enhance the GRU model by replacing the reset gate with an attention mechanism, thereby increasing the model’s flexibility in selecting and utilizing historical information. Second, we design a multi-head cross-attention mechanism for learning unobservable latent market state representations, which are further enriched through interactions with both temporal features and cross-sectional features. Finally, extensive experiments conducted on the CSI 300 and CSI 500 datasets from the Chinese stock market, as well as the NASDAQ 100 and S&P 500 datasets from the U.S. stock market, demonstrate that the proposed method outperforms the current state-of-the-art methods across multiple metrics. Furthermore, this approach has been successfully applied in the real-world operations of a fund management company, validating its effectiveness and practicality in actual financial environments.}, archive = {J_NEUCOM}, author = {Peng Zhu and Yuante Li and Yifan Hu and Sheng Xiang and Qinyuan Liu and Dawei Cheng and Yuqi Liang}, doi = {10.1016/j.neucom.2025.130168}, journal = {Neurocomputing}, month = {7}, pages = {130168}, shortjournal = {Neurocomputing}, title = {MCI-GRU: Stock prediction model based on multi-head cross-attention and improved GRU}, volume = {638}, year = {2025}, }
• (2025). Improved physics-informed neural network in mitigating gradient-related failures. NEUCOM, 638, 130167. (www)

  @article{ , abstract = {Physics-informed neural network (PINN) integrates fundamental physical principles with advanced data-driven techniques, leading to significant advancements in scientific computing. However, PINN encounters persistent challenges related to stiffness in gradient flow, which limits their predictive capabilities. This paper introduces an improved PINN (I-PINN) designed to mitigate gradient-related failures. The core of I-PINN combines the respective strengths of neural networks with an improved architecture and adaptive weights that include upper bounds. I-PINN achieves improved accuracy by at least one order of magnitude and accelerate convergence without introducing additional computational complexity compared to the baseline model. Numerical experiments across a variety of benchmarks demonstrate the enhanced accuracy and generalization of I-PINN. The supporting data and code are accessible at https://github.com/PanChengN/I-PINN.git , facilitating broader research engagement.}, archive = {J_NEUCOM}, author = {Pancheng Niu and Jun Guo and Yongming Chen and Yuqian Zhou and Minfu Feng and Yanchao Shi}, doi = {10.1016/j.neucom.2025.130167}, journal = {Neurocomputing}, month = {7}, pages = {130167}, shortjournal = {Neurocomputing}, title = {Improved physics-informed neural network in mitigating gradient-related failures}, volume = {638}, year = {2025}, }
• (2025). MelodyTransformer: Improving lyric-to-melody generation by considering melodic features. NEUCOM, 638, 130166. (www)

  @article{ , abstract = {Lyric-to-melody (L2M) generation is a fascinating research topic in artificial intelligence. Traditional L2M approaches that rely on language models from Natural Language Processing (NLP) rarely consider music as a unique structure, resulting in a lack of coherence and innovation. In this paper, MelodyTransformer is proposed to alleviate the above problems, which inherits and modifies the transformer architecture while considering the melodic features entirely regarding the L2M task. Specifically, (1) a dual omnidirectional self-attention module is introduced to the architecture of the Transformer, which helps the model understand the input data’s structural relationships more deeply and improves the quality of coherence and wholeness of the generated melody. (2)A new loss function and positional encoding design are proposed for the model’s training process, which helps the model account for the uncertainty in the musical sequence’s length, the temporal order, and the complexity of the note elements, and significantly improving the innovation of the melody generation. In addition, our research improved the L2M generation process and proposed new evaluation criteria adapted to the melody generation task. Both objective and subjective evaluations indicate that our method not only improves the coherence and innovation of generated melodies but also effectively considers the needs of different populations for musical melody generation. Our code is available at https://github.com/MelodyTransformer . Our music are available at https://github.com/Melodysong .}, archive = {J_NEUCOM}, author = {Ruhan He and Ruixue Liu and Tao Peng and Xinrong Hu}, doi = {10.1016/j.neucom.2025.130166}, journal = {Neurocomputing}, month = {7}, pages = {130166}, shortjournal = {Neurocomputing}, title = {MelodyTransformer: Improving lyric-to-melody generation by considering melodic features}, volume = {638}, year = {2025}, }
• (2025). Bayesian generalization error in linear neural networks with concept bottleneck structure and multitask formulation. NEUCOM, 638, 130165. (www)

  @article{ , abstract = {Concept bottleneck model (CBM) is a ubiquitous method that can interpret neural networks using concepts. In CBM, concepts are inserted between the output layer and the last intermediate layer as observable values. This helps in understanding the reason behind the outputs generated by the neural networks: the weights corresponding to the concepts from the last hidden layer to the output layer. However, it has not yet been possible to understand the behavior of the generalization error in CBM since a neural network is a singular statistical model in general. When the model is singular, a one to one map from the parameters to probability distributions cannot be created. This non-identifiability makes it difficult to analyze the generalization performance. In this study, we mathematically clarify the Bayesian generalization error and free energy of CBM when its architecture is three-layered linear neural networks. We also consider a multitask problem where the neural network outputs not only the original output but also the concepts. The results show that CBM drastically changes the behavior of the parameter region and the Bayesian generalization error in three-layered linear neural networks as compared with the standard version, whereas the multitask formulation does not.}, archive = {J_NEUCOM}, author = {Naoki Hayashi and Yoshihide Sawada}, doi = {10.1016/j.neucom.2025.130165}, journal = {Neurocomputing}, month = {7}, pages = {130165}, shortjournal = {Neurocomputing}, title = {Bayesian generalization error in linear neural networks with concept bottleneck structure and multitask formulation}, volume = {638}, year = {2025}, }
• (2025). Distributed finite-time convergent algorithms for time-varying optimization with coupled equality constraint over weight-unbalanced digraphs. NEUCOM, 638, 130163. (www)

  @article{ , abstract = {This paper studies time-varying optimization problems solved by distributed finite-time convergent algorithms under arbitrary strongly connected digraphs. The primary objective is to propose distributed methodologies for minimizing time-varying objective function subject to coupled equality constraint, focusing on the challenging case of weight-unbalanced digraphs. This network topology is among the most general and complex, with limited results available in the literature. To address this, we present innovative theoretical results on directed graphs, which are pivotal to the finite-time convergence analysis of the proposed algorithms. Compared to existing results, the algorithms proposed in this paper achieve finite-time optimization for distributed time-varying problems without introducing additional estimators that would increase the algorithmic complexity. Finally, two simulation examples, including application to the target encirclement problem in multi-robot system, are provided to substantiate the theoretical results.}, archive = {J_NEUCOM}, author = {Yuanyuan Yue and Qingshan Liu and Haitao Wang and Jingyi Huang}, doi = {10.1016/j.neucom.2025.130163}, journal = {Neurocomputing}, month = {7}, pages = {130163}, shortjournal = {Neurocomputing}, title = {Distributed finite-time convergent algorithms for time-varying optimization with coupled equality constraint over weight-unbalanced digraphs}, volume = {638}, year = {2025}, }
• (2025). Learning deep feature representations for multi-modal MR brain tumor segmentation. NEUCOM, 638, 130162. (www)

  @article{ , abstract = {Brain tumor segmentation is crucial for accurate diagnosis, treatment planning, and patient monitoring. Different MRI sequences can provide unique and complementary information about various aspects of brain tumors. However, effectively integrating diverse data sources to achieve accurate segmentation remains a significant challenge due to the inherent complexity and variability of the data. To address this challenge, this paper proposes a deep learning framework designed to fuse multi-modal MRI data and enhance brain tumor segmentation accuracy. Specifically, the framework introduces two innovative modules: the modality-wise feature fusion module (MFFM) and the spatial and channel-wise feature fusion module (SCFFM). The MFFM aims to learn modality-specific features and integrate information from diverse modalities, thereby ensuring richer and more discriminative feature representations. Meanwhile, the SCFFM is designed to capture contextual information and achieve multi-channel data incorporation by emphasizing informative regions and highlighting critical features. Together, these modules collaboratively enhance the model’s capacity for feature learning, leading to more precise tumor segmentation. Experimental validation on two public datasets demonstrates the effectiveness of the proposed approach, achieving an average Dice similarity coefficient of 83.2% with an average 95% Hausdorff distance of 4.3 mm on the BraTS 2018 dataset, and an average Dice similarity coefficient of 82.9% with an average 95% Hausdorff distance of 5.5 mm on the BraTS 2019 dataset. This framework not only presents an effective method for precise multi-modal brain tumor segmentation but also provides a promising solution for other challenges in multi-modal data fusion.}, archive = {J_NEUCOM}, author = {Tongxue Zhou and Zheng Wang and Xiaohui Liu and Weibo Liu and Shan Zhu}, doi = {10.1016/j.neucom.2025.130162}, journal = {Neurocomputing}, month = {7}, pages = {130162}, shortjournal = {Neurocomputing}, title = {Learning deep feature representations for multi-modal MR brain tumor segmentation}, volume = {638}, year = {2025}, }
• (2025). MMamba: Enhancing image deraining with morton curve-driven locality learning. NEUCOM, 638, 130161. (www)

  @article{ , abstract = {Rain streaks in images exhibit complex forms and spatial structures. Their removal not only requires restoring the clarity of the image but also aims to preserve as much of the original information as possible. Transformer-based deep learning models have made significant progress in image deraining, effectively capturing the diversity of rain streaks and removing them through deep feature learning. However, when handling long sequence data, the computational complexity rises sharply, severely limiting the efficiency of practical applications. In recent years, the Mamba module, with its unique design, has demonstrated great potential in handling long sequences and enabling fast inference in image deraining tasks. It reduces computational complexity while preserving the underlying features of the image, thus improving efficiency. However, compared to Transformer-based models, the Mamba model still faces certain limitations in preserving local information and dealing with complex rain streaks. To address these issues, this paper proposes an efficient MMamba model, which combines the Morton curve-based state space model (SSM) to further enhance the ability to retain local information. Additionally, a dynamic channel attention mechanism is introduced, which significantly improves image detail restoration by partitioning channels and assigning learnable weights to each subset. Finally, for images with rich and complex rain streaks, a Selective Rain Stripe Compensator is proposed, effectively identifying and removing these intricate rain streaks. Extensive experiments on five commonly used benchmark datasets demonstrate that our method outperforms state-of-the-art techniques in terms of performance.}, archive = {J_NEUCOM}, author = {Zongzhi Ouyang and Wenhui Li}, doi = {10.1016/j.neucom.2025.130161}, journal = {Neurocomputing}, month = {7}, pages = {130161}, shortjournal = {Neurocomputing}, title = {MMamba: Enhancing image deraining with morton curve-driven locality learning}, volume = {638}, year = {2025}, }
• (2025). Fixed-time synchronization of discontinuous complex-valued BAM neural networks with time delays. NEUCOM, 638, 130160. (www)

  @article{ , abstract = {The fixed-time synchronization (FTS) of discontinuous complex-valued bidirectional associative memory neural networks (CVBAMNNs) with time delays is investigated in this paper. First, CVBAMNNs are separated into real and imaginary parts, the equivalent real-valued subsystems are constructed for the analysis. Then, two kinds of discontinuous feedback controllers are designed, and two forms of Lyapunov function based on p -norm and 1-norm are constructed to study the FTS of CVBAMNNs. By using differential inclusions theory and some inequality techniques, some neoteric criteria are established to achieve FTS, and the upper-bound of the setting time is more explicitly estimated as well. Finally, numerical simulations are presented to verify the effectiveness and applicability of the developed theoretical results.}, archive = {J_NEUCOM}, author = {Qian Wang and Yinjie Qian and Yuanhua Qiao}, doi = {10.1016/j.neucom.2025.130160}, journal = {Neurocomputing}, month = {7}, pages = {130160}, shortjournal = {Neurocomputing}, title = {Fixed-time synchronization of discontinuous complex-valued BAM neural networks with time delays}, volume = {638}, year = {2025}, }
• (2025). Enhancing the accuracy of generative adversarial networks with Fokker–Planck equations. NEUCOM, 638, 130158. (www)

  @article{ , abstract = {GANs have the advantage of much less computation cost and high-quality generative capabilities but occasionally encounter mode collapse. Previous works address the problems more from the perspective of the game between the generator and the discriminator, ignoring the accuracy of generator distribution learning, which is the key target of the generator. Fokker–Planck Equation (FPE) is a powerful tool to overcome above shortcomings as it is Wasserstein gradient flow (WGF) of the given function in probability space. In this paper, we propose a new scheme using FPE in the training process of GAN. Specifically, learning real distribution of the generator in a GAN is replaced by solving a specific FPE since both describe the evolution of a probability distribution. GAN applied with FPE converges faster and mitigates mode collapse because the optimization direction is along the WGF of given loss. We pull it back to parameter space for convenience of solving and develop a error-bounded numerical method to solve FPE, ensuring generative distribution closer to the real enough even in high-dimensional manifold. Applying our approach achieve competitive performance compared to the vanilla across multiple popular datasets. Furthermore, we apply FPE-based method to style transfer and demonstrate the generalization of our approach on various image generation tasks.}, archive = {J_NEUCOM}, author = {Ben Wan and Tianyi Zheng and Zhaoyu Chen and Jia Wang}, doi = {10.1016/j.neucom.2025.130158}, journal = {Neurocomputing}, month = {7}, pages = {130158}, shortjournal = {Neurocomputing}, title = {Enhancing the accuracy of generative adversarial networks with Fokker–Planck equations}, volume = {638}, year = {2025}, }
• (2025). Adaptive dynamic programming based event-triggered multi-h∞ control. NEUCOM, 638, 130157. (www)

  @article{ , abstract = {This article presents an adaptive dynamic programming (ADP) based event-triggered multi- H ∞ control method for the completely unknown systems. First, a weighted average method is used to ensure balanced disturbance resistance for each input. Next, neural networks (NNs) are employed to identify the dynamics of completely unknown systems. Then, critic NNs are used to approximate value functions to achieve Nash equilibrium . Additionally, event-triggered control and disturbance strategies with a new threshold are developed. Finally, the method’s effectiveness is validated through two numerical experiments.}, archive = {J_NEUCOM}, author = {Shan Xue and Zhe Liu and Liqi Wang and Weidong Zhang and Ke Ren and Xinhui Yang}, doi = {10.1016/j.neucom.2025.130157}, journal = {Neurocomputing}, month = {7}, pages = {130157}, shortjournal = {Neurocomputing}, title = {Adaptive dynamic programming based event-triggered multi-h∞ control}, volume = {638}, year = {2025}, }
• (2025). Determination of barrier surface in target-attacker-defender game with capture radius for superior pursuer. NEUCOM, 638, 130156. (www)

  @article{ , abstract = {In a military operation, a drone tasked with striking a high-value asset might be intercepted by a surface-to-air missile, which can be modeled as a Target-Attacker-Defender (TAD) game. In existing TAD games with a slower attacker, capture is modeled as point capture with a zero radius. However, in real scenarios, both the attacker and defender may have non-zero capture radii, as in aircraft or missile engagements, shifting the capture curve from an Apollonius circle to an Apollonius oval. Furthermore, when the attacker has a capture radius, they can destroy the target outside the interception curve formed by the defender-attacker. This renders existing methods based on center-to-center distance inadequate for solving the game of kind problem in TAD games. To address this, this paper first defines the interception curve between the attacker and defender, considering their initial positions and capture radii. It then constructs the attacker’s farthest firepower coverage curve. Using this, Algorithm 1 is proposed to determine whether the current state belongs to the attacker’s or the target-defender’s winning area, based on the intersection of the attacker’s firepower coverage curve and the attacker-target capture curve. Finally, by analyzing the geometric relationship between the farthest firepower coverage curve and the attacker-target capture curve as an extremum problem, we derive the barrier surface that separates the attacker’s and the target-defender’s winning areas. The proposed method is validated through simulation.}, archive = {J_NEUCOM}, author = {Yibo Shi and Chaoli Wang and Dong Liang}, doi = {10.1016/j.neucom.2025.130156}, journal = {Neurocomputing}, month = {7}, pages = {130156}, shortjournal = {Neurocomputing}, title = {Determination of barrier surface in target-attacker-defender game with capture radius for superior pursuer}, volume = {638}, year = {2025}, }
• (2025). FourierLoss: Shape-aware loss function with fourier descriptors. NEUCOM, 638, 130155. (www)

  @article{ , abstract = {Encoder–decoder networks are commonly used for medical image segmentation tasks. When they are trained with a standard loss function, these networks are not explicitly enforced to preserve the shape integrity of an object in an image. However, this ability of the network is important to obtain more accurate results, especially when there is a low-contrast difference between the object and its surroundings. To respond this issue, this work introduces a new shape-aware loss function, which we name FourierLoss . This loss function relies on quantifying the shape dissimilarity between the ground truth and the predicted segmentation maps through the Fourier descriptors calculated on the objects of these maps, and penalizing this dissimilarity in network training. Different than the previous studies, FourierLoss offers an adaptive loss function with trainable hyperparameters that control the importance of the level of the shape details that the network is enforced to learn in the training process. This control is achieved by the proposed adaptive loss update mechanism, which end-to-end learns the hyperparameters simultaneously with the network weights by backpropagation. As a result of using this mechanism, the network can dynamically change its attention from learning the general outline of an object to learning the details of its contour points, or vice versa, in different training epochs. Working on two different datasets, our experiments revealed that the proposed adaptive shape-aware loss function led to statistically significantly better results for liver segmentation, compared to its counterparts.}, archive = {J_NEUCOM}, author = {Mehmet Bahadir Erden and Selahattin Cansiz and Onur Caki and Haya Khattak and Durmus Etiz and Melek Cosar Yakar and Kerem Duruer and Berke Barut and Cigdem Gunduz-Demir}, doi = {10.1016/j.neucom.2025.130155}, journal = {Neurocomputing}, month = {7}, pages = {130155}, shortjournal = {Neurocomputing}, title = {FourierLoss: Shape-aware loss function with fourier descriptors}, volume = {638}, year = {2025}, }
• (2025). ShapeGPT and PointPainter for fast zero shot text-to-3d point cloud generation. NEUCOM, 638, 130154. (www)

  @article{ , abstract = {Many recent works have attempted to generate 3D shapes from textual prompts. However, some of these methods are limited by high computational demands and slow optimization speeds, while others are constrained by the lack of sufficient detail, color, and texture in the generated results. This article proposes a method for rapidly generating high-fidelity 3D point clouds with color and texture based on textual prompts. The method consists of two stages: in the first stage, we generate voxels based on the textual prompts, and in the second stage, we refine the generated voxels into point clouds with enhanced geometric details and generate colors for them. During the voxel generation stage, we innovatively treat voxels as a form of “shape language” and propose ShapeGPT, a text-conditional voxel generative model based on GPT. It fully leverages the excellent generative capabilities of the language model, enabling the generation of high-fidelity and diverse voxels without the need for text-3D shape pairs as training data. Furthermore, we introduce PointPainter to upsample voxels into high-resolution dense point clouds and add colors and textures to the point clouds based on textual prompts. When generating color and texture, we use a newly proposed differentiable renderer to render the point clouds into images from multiple viewpoints and match the rendered images with the target text using the CLIP model. PointPainter not only generates intricate textures for point clouds but also achieves a fast optimization speed. Extensive experimental results demonstrate that our approach outperforms the state-of-the-art methods in terms of generative quality and consistency with the input text.}, archive = {J_NEUCOM}, author = {Zeyun Wan and Shichao Fan and Huan Zhao and Jiawei Liu and Ying Chen}, doi = {10.1016/j.neucom.2025.130154}, journal = {Neurocomputing}, month = {7}, pages = {130154}, shortjournal = {Neurocomputing}, title = {ShapeGPT and PointPainter for fast zero shot text-to-3d point cloud generation}, volume = {638}, year = {2025}, }
• (2025). Embryo-net: A blastocyst image segmentation network based on spatial modeling to resolve the challenge of textural similarity between the TE and ICM. NEUCOM, 638, 130153. (www)

  @article{ , abstract = {This study introduces Embryo-Net, an innovative embryo segmentation network designed to address the challenge of textural similarity between the inner cell mass (ICM) and trophectoderm (TE) in blastocyst images during in vitro fertilization (IVF). Existing segmentation methods often fall short in terms of accuracy and interpretability when dealing with these regions. Embryo-Net tackles these issues by employing a multipath, multiloss framework combined with spatial modeling, specifically optimized for blastocyst stage images. This approach integrates multipath feature fusion and multi-objective loss optimization, significantly improving segmentation accuracy and model interpretability. We validated Embryo-Net using a publicly available blastocyst segmentation dataset, achieving intersection over union (IoU) scores of 81.20% for TE and 86.51% for ICM segmentation, surpassing existing state-of-the-art algorithms. The spatial modeling module plays a crucial role in accurately segmenting the TE and ICM regions, effectively addressing the challenge of textural similarity. Ablation experiments further demonstrate the vital contributions of these strategies to the overall performance of the model.}, archive = {J_NEUCOM}, author = {Kuo Chen and Zhenghang Li and Guoning Huang and Jinhong Guo}, doi = {10.1016/j.neucom.2025.130153}, journal = {Neurocomputing}, month = {7}, pages = {130153}, shortjournal = {Neurocomputing}, title = {Embryo-net: A blastocyst image segmentation network based on spatial modeling to resolve the challenge of textural similarity between the TE and ICM}, volume = {638}, year = {2025}, }
• (2025). Video-text retrieval based on multi-grained hierarchical aggregation and semantic similarity optimization. NEUCOM, 638, 130152. (www)

  @article{ , abstract = {Current video-text retrieval approaches face two major challenges. Firstly, the semantic information imbalance between modalities, primarily caused by differences in granularity of inter-modal feature correlations, poses challenges for accurate alignment. Secondly, the presence of irrelevant content, which can be considered as noise, within videos and texts frequently leads to incorrect retrieval results. To address these issues, we propose the multi-grained hierarchical aggregation and semantic similarity optimization network for VTR, which is named as MSNet. MSNet employs a multi-grained hierarchical aggregation module (MHA) we designed to integrate global and local information. It introduces the expectation–maximization contrastive learning (EMCL) strategy and the interactive noise reduction (INR) technique to effectively reduce the impact of noise. Moreover, MHA extracts the optimized frame feature by a contextual frame encoder we designed to mitigate semantic imbalance. Additionally, the proposed semantic similarity optimization module (SSO) refines multi-scale feature interactions. Extensive experiments on benchmark datasets, including MSR-VTT, MSVD and DiDeMo, demonstrate the superior performance and effectiveness of MSNet in VTR.}, archive = {J_NEUCOM}, author = {Jie Guo and Shujie Lan and Bin Song and Mengying Wang}, doi = {10.1016/j.neucom.2025.130152}, journal = {Neurocomputing}, month = {7}, pages = {130152}, shortjournal = {Neurocomputing}, title = {Video-text retrieval based on multi-grained hierarchical aggregation and semantic similarity optimization}, volume = {638}, year = {2025}, }
• (2025). Noise-robust multimodal remote sensing image matching via geometric analysis of embedded pre-trained manifolds. NEUCOM, 638, 130150. (www)

  @article{ , abstract = {Multimodal remote sensing image matching has seen significant advancements through handcrafted and learning-based methods. However, challenges persist due to limited local structure representations in conventional methods and data dependency in learning-based approaches. Accordingly, this research introduces an innovative strategy to enhance the generalization of pre-trained networks by embedding their manifolds based on intrinsic dimension estimation, aiming to mitigate the memorization effect from adapting to training data during the learning process. Maximum index maps of embedded kernels were integrated with an adaptive binning spatial configuration to formulate six distinct descriptors derived from prominent networks, including HardNet8, HardNet, HyNet, DeDoDe (Detect, Don’t Describe- Describe, Don’t Detect), OSflowNet (Optical-SAR flowNet), and SAM (Segment Anything Model). The outcome addresses the challenge of scale and local geometric distortions and exhibits robustness to noise and illumination variations. On average, our modified version of HardNet8, as the superior descriptor, achieved recall values of 66.43 % for noise-robust kernel analysis (compared to 50.95 % for RIFT and 50.92 % for the original network) and 54.37 % for integrated geometric-radiometric analysis (compared to 36.88 % for the original network and 33.46 % for average of HOSS, OS-SIFT, RISS, OSS, WSSF, and scale-invariant RIFT). Notably, a rotation-invariant scheme was also considered, which overperformed the recently proposed rotation-invariant methods. The demo code of the proposed method can be downloaded from https://github.com/AmSedaghat/PTEM-Descriptor.}, archive = {J_NEUCOM}, author = {Negar Jovhari and Amin Sedaghat and Reza Shah-Hosseini and Nazila Mohammadi and Mahdi Hasanlou}, doi = {10.1016/j.neucom.2025.130150}, journal = {Neurocomputing}, month = {7}, pages = {130150}, shortjournal = {Neurocomputing}, title = {Noise-robust multimodal remote sensing image matching via geometric analysis of embedded pre-trained manifolds}, volume = {638}, year = {2025}, }
• (2025). Turn the tables: Proactive deception defense decision-making based on bayesian attack graphs and stackelberg games. NEUCOM, 638, 130139. (www)

  @article{ , abstract = {As cyberattacks become increasingly sophisticated, traditional defense mechanisms are outmatched against advanced persistent threats (APTs). Active deception defense mechanisms that rely on traditional honeypots have proven to be able to sense threats to a certain extent, but their reliance on statically deployed honeypots leads to the risk of being bypassed and failing to sense covert attackers. In this paper, we address the limitations of traditional deception defense and introduce a new lightweight technique, Honeypoint, which detects hidden unknown threats early by disguising itself as a legitimate service. Further, to address the lack of dynamic changes in deception decoys and the fact that deploying honeypoints without a purpose leads to a great waste of resources and costs. We propose an active deception defense model based on Bayesian attack graph and Stackelberg game (BagSg) for guiding the deployment of honeypots. The Bayesian attack graph helps identify attack paths and vulnerabilities in the network, while the Stackelberg game models the strategic interactions between attackers and defenders, and discusses the impact of deception defense strategies on both regular and advanced attackers in a long-term attack and defense game. Experimental results show that Honeypoint significantly outperforms traditional honeypots in detecting covert attack behavior. Moreover, the deployment efficiency of deception strategies can be optimized by using the BagSg model, revealing the importance of using deception defense strategies consistently in long-term cyber security defense.}, archive = {J_NEUCOM}, author = {Rui Wang and Changjiang Yang and Xiangdong Deng and Yinghai Zhou and Yuan Liu and Zhihong Tian}, doi = {10.1016/j.neucom.2025.130139}, journal = {Neurocomputing}, month = {7}, pages = {130139}, shortjournal = {Neurocomputing}, title = {Turn the tables: Proactive deception defense decision-making based on bayesian attack graphs and stackelberg games}, volume = {638}, year = {2025}, }
• (2025). An evolutionary metaheuristic for forming teams in the classroom with constraints. NEUCOM, 638, 130138. (www)

  @article{ , abstract = {Team formation is essential for developing teamwork-related skills in educational settings. The problem of team formation in the classroom consists of partitioning a classroom into non-overlapping teams of students, including every single student. Several algorithms have been proposed to automate the formation of teams, each employing different criteria for guiding the team formation process. Traditionally, metaheuristics have been a common approach due to the combinatorial complexity of the problem. This paper introduces a novel and general evolutionary algorithm for team formation in the classroom guided by mutation, the general concept of synergy between team members, and local search. Our algorithm allows for flexible team size constraints and the inclusion of compulsory and forbidden student combinations, which are not considered in existing methods but are important for capturing human relationships in the classroom. In addition, our algorithm is independent of the specific objective function employed to evaluate the teams formed. We present experiments comparing our proposal with other state-of-the-art algorithms, demonstrating robust performance across different objective functions employed in the team formation literature, superior scalability as the problem size increases, and remarkable performance in settings with or without the aforementioned constraints.}, archive = {J_NEUCOM}, author = {Gonzalo Candel and Victor Sánchez-Anguix and Juan M. Alberola and Vicente Julián and Vicent Botti}, doi = {10.1016/j.neucom.2025.130138}, journal = {Neurocomputing}, month = {7}, pages = {130138}, shortjournal = {Neurocomputing}, title = {An evolutionary metaheuristic for forming teams in the classroom with constraints}, volume = {638}, year = {2025}, }
• (2025). A dynamic anchor-based online semi-supervised learning approach for fault diagnosis under variable operating conditions. NEUCOM, 638, 130137. (www)

  @article{ , abstract = {The diagnosis of multi-mode faults under time-varying operating conditions poses a significant challenge in process control. Production processes frequently transition through multiple steady states, influenced by external environmental factors, changes in production strategies, and fluctuations in raw material properties. In such contexts, the impact of operating conditions on monitoring data can overshadow the effect of faults, thereby undermining the accuracy of fault diagnosis. To address this, effective management of transitional conditions requires updating diagnostic models with real-time monitoring data, as relying solely on historical data is often impractical. This paper introduces an online dynamic semi-supervised broad learning system for real-time fault diagnosis in variable operating conditions. The proposed system employs a dynamic center-based anchor set construction method, designed to adapt to the evolving distribution of monitoring data over time. Additionally, we derive a model update rule incorporating a dynamic manifold regularization term to guide the direction of online model evolution. Experimental results from gearbox fault diagnosis under various operating conditions validate the proposed approach, demonstrating its ability to reduce the adverse effects of transitional conditions on diagnostic performance.}, archive = {J_NEUCOM}, author = {Wei Li and Zeyi Liu and Pengyu Han and Xiao He and Limin Wang and Tao Zhang}, doi = {10.1016/j.neucom.2025.130137}, journal = {Neurocomputing}, month = {7}, pages = {130137}, shortjournal = {Neurocomputing}, title = {A dynamic anchor-based online semi-supervised learning approach for fault diagnosis under variable operating conditions}, volume = {638}, year = {2025}, }
• (2025). A modified automatic label generation method under low-light environments for event-based semantic segmentation. NEUCOM, 638, 130136. (www)

  @article{ , abstract = {Event cameras are novel bio-inspired sensors, which have the advantages of high temporal resolution, high dynamic range and low latency compared with traditional cameras. Employing event cameras for semantic segmentation is an area that deserves in-depth research to exploit the great potential of event cameras. Due to the asynchronous nature of data from event cameras, it is challenging to produce accurate labels directly from events for the training of event-based semantic segmentation networks, especially in low-light environments. In this paper, we propose a new automatic label generation method under low-light environment for event-based semantic segmentation. Firstly we propose a modified automatic label generation module to generate more accurate labels for training. Meanwhile, we propose an event-assisted low-light enhancement module to improve the accuracy of the labels of low-light data. Finally we provide an event-based semantic segmentation dataset with high-precision paired labels for both low-light and normal environments. Experiments have shown that our method accomplishes high-quality event-based semantic segmentation in both low-light and normal environments by acquiring high-precision labels. To the best of our knowledge, our work is the first to realize event-based semantic segmentation in low-light environments in the field of event-based semantic segmentation.}, archive = {J_NEUCOM}, author = {Zehao Wu and Yuanqing Xia and Rui Hu}, doi = {10.1016/j.neucom.2025.130136}, journal = {Neurocomputing}, month = {7}, pages = {130136}, shortjournal = {Neurocomputing}, title = {A modified automatic label generation method under low-light environments for event-based semantic segmentation}, volume = {638}, year = {2025}, }
• (2025). Can large language models independently complete tasks? a dynamic evaluation framework for multi-turn task planning and completion. NEUCOM, 638, 130135. (www)

  @article{ , abstract = {Large language models (LLMs) are increasingly relied upon for multi-turn dialogue to conduct complex tasks. However, existing benchmarks mainly evaluate LLMs as agents, overlooking their potential as independent systems to accomplish complex tasks. In addition, these benchmarks typically evaluate the planning and completion capabilities of the models individually, rather than simultaneously. To address these issues, we propose a new Dynamic Evaluation Framework for Multi-Turn task planning and completion (DEF-MT) to assess the ability of LLM to independently complete complex tasks in multi-turn scenarios. Our approach quantifies the model’s planning capability by guiding it to generate planning and responses sequentially. Simultaneously, we use a dynamic approach to generate data that simulates the complex intents of real users. Finally, experiments conducted on 9 mainstream models using the Multiwoz 2.2 dataset, indicate that the existing models’ sub-task planning capabilities hinder their ability to complete complex tasks, providing a meaningful reference for the future optimization direction of LLM.}, archive = {J_NEUCOM}, author = {Jun Gao and Junlin Cui and Huijia Wu and Liuyu Xiang and Han Zhao and Xiangang Li and Meng Fang and Yaodong Yang and Zhaofeng He}, doi = {10.1016/j.neucom.2025.130135}, journal = {Neurocomputing}, month = {7}, pages = {130135}, shortjournal = {Neurocomputing}, title = {Can large language models independently complete tasks? a dynamic evaluation framework for multi-turn task planning and completion}, volume = {638}, year = {2025}, }
• (2025). Pyramid fusion network with gated transformer for free-form image inpainting. NEUCOM, 638, 130133. (www)

  @article{ , abstract = {Free-form image inpainting with large-area masks presents significant challenges for attention-driven methods. Current approaches suffer from two major limitations: (1) inadequate hierarchical integration of high-level semantics and low-level spatial features, causing unrealistic structural and color transitions, and (2) computationally intensive attention mechanisms that increase parameters and inference time. To address these issues, this paper proposes PFGT-GAN, a Pyramid Fusion network with the Gated Transformer architecture. Our framework introduces two key innovations: First, a pyramidal encoder–decoder structure enabling progressive cross-layer fusion of multi-scale features. Second, a lightweight gated Transformer module that dynamically weights feature importance for global context modeling while reducing computational overhead. Extensive experiments on Places2 (0%–50% masks) and Paris StreetView datasets (all masks) demonstrate state-of-the-art performance. Remarkably, our model achieves these results with only 4.9M parameters – 0.9% of RePaint’s and 21.4% of LaMa’s parameter counts – while delivering the fastest inference speed at 9 ms/image. Additional evaluations on real-world applications, such as logo removal and facial editing, further confirm practical effectiveness.}, archive = {J_NEUCOM}, author = {Wenxuan Yan and Ning Xu and Xin Su and Aimin Jiang}, doi = {10.1016/j.neucom.2025.130133}, journal = {Neurocomputing}, month = {7}, pages = {130133}, shortjournal = {Neurocomputing}, title = {Pyramid fusion network with gated transformer for free-form image inpainting}, volume = {638}, year = {2025}, }
• (2025). Multi-resolution patch-based fourier graph spectral network for spatiotemporal time series forecasting. NEUCOM, 638, 130132. (www)

  @article{ , abstract = {Spatiotemporal time series is of great significance in many fields such as finance and meteorology. Accurate forecasting depends on the effective understanding of their spatiotemporal correlations. However, existing methods have two major problems: firstly, the analysis of multi-resolution spatiotemporal correlations is insufficient; secondly, they are difficult to analyze in terms of spatiotemporal correlations efficiently. Some graph-based methods have high computational costs, while efficient methods struggle to learn spatiotemporal relationships. To overcome these problems, we introduce the Multi-resolution Patch-based Fourier Graph Spectral Network (MPFGSN), which leverages a hyperpatch graph and a novel graph spectral convolution to spatiotemporal time series prediction, enhancing temporal dependency learning. We first capture the multi-scale features of the time series through multi-resolution patches, then represent the spatiotemporal relations in the time series with the hyperpatch graph structure. Then, the Fourier Graph Spectral Network (FGSN) is proposed. The FGSN utilizes Fourier transformation and point multiplication operations in the spectral domain to implement efficient graph spectral convolution. And it maintains the capabilities of traditional multi-layer graph convolution while also being faster. Finally, it integrates information from different resolutions through a residual fusion mechanism, capturing both long-term trends and short-term fluctuations, thus enhancing the model’s robustness. Extensive experiments on public datasets confirm MPFGSN’s superiority in accuracy and generalization over existing models, validating its practical applicability in sophisticated forecasting scenarios.}, archive = {J_NEUCOM}, author = {Jiankai Zheng and Liang Xie and Haijiao Xu}, doi = {10.1016/j.neucom.2025.130132}, journal = {Neurocomputing}, month = {7}, pages = {130132}, shortjournal = {Neurocomputing}, title = {Multi-resolution patch-based fourier graph spectral network for spatiotemporal time series forecasting}, volume = {638}, year = {2025}, }
• (2025). Efficient active learning using recursive estimation of error reduction. NEUCOM, 638, 130131. (www)

  @article{ , abstract = {Efficient labeling is an open challenge in the big data era, where the rate of data collection far exceeds human capacity for data labeling required for training machine learning models. Active learning (AL) attempts to find the most valuable instances from a pool of unlabeled data by exploiting a limited annotation budget. In this paper, we focus on the estimated error reduction (EER) approach for AL. Given an initial labeled set, EER aims to identify the candidate unlabeled data points that, if used in training with other labeled data, minimize a weighted average of error estimates. As such, EER is a promising technique for AL because it directly examines the effect of labeling the candidate data points on the classifier’s performance. However, the main shortcoming of EER is attributed to its excessive computational cost, as it requires retraining the classifier for each candidate data point and every possible label. In this regard, we propose an efficient EER algorithm using an accurate closed-form error estimator of Regularized Linear Discriminant Analysis (AL-RLDA) for the case of binary classification. At the same time, we introduce a novel objective function for EER, that relies less on the estimates of posterior probabilities. The efficiency of the algorithm is attributed to: (i) using a closed-form error estimator; and (ii) updating the estimator expression recursively for each candidate data point. Our empirical results using real data show that the proposed method can outperform baseline and state-of-the-art AL methods.}, archive = {J_NEUCOM}, author = {Azamat Mukhamediya and Reza Sameni and Amin Zollanvari}, doi = {10.1016/j.neucom.2025.130131}, journal = {Neurocomputing}, month = {7}, pages = {130131}, shortjournal = {Neurocomputing}, title = {Efficient active learning using recursive estimation of error reduction}, volume = {638}, year = {2025}, }
• (2025). PT-VAE: Variational autoencoder with prior concept transformation. NEUCOM, 638, 130129. (www)

  @article{ , abstract = {Learning and disentangling coherent latent representations of variational autoencoders (VAEs) have recently attracted widespread attention. However, the latent space of the VAE model is constrained by the prior distribution, which can hinder the latent variables from accurately capturing semantic information, thereby limiting its disentanglement and interpretability. This paper proposes PT-VAE, which constructs the latent space by a well-constructed latent space rather than a carefully designed prior distribution to guide the latent variables. Firstly, we transform the initial constraints of the latent space into understandable latent variable distributions, the so-called prior concept, which can be introduced into the latent space. Then, we design the Gumbel softmax reparameterization trick to enhance the integration of the prior concept and latent variables. Furthermore, the training process of PT-VAE is guided by deriving a variational lower bound, which facilitates the construction of the latent space concept based on the prior concept. Compared with 8 state-of-the-art VAE models, the PT-VAE improves the average clustering accuracy by over 11 % on the Fashion MNSIT, MNIST, COIL20, and COIL10 datasets. Moreover, the PT-VAE elucidates the process of information aggregation within the model and uncovers disentangled representations. PT-VAE provides a novel and flexible approach to construct an interpretable latent space by embedding prior concepts and disentangling the latent variables.}, archive = {J_NEUCOM}, author = {Zitu Liu and Yue Liu and Zhenyao Yu and Zhengwei Yang and Qingshan Fu and Yike Guo and Qun Liu and Guoyin Wang}, doi = {10.1016/j.neucom.2025.130129}, journal = {Neurocomputing}, month = {7}, pages = {130129}, shortjournal = {Neurocomputing}, title = {PT-VAE: Variational autoencoder with prior concept transformation}, volume = {638}, year = {2025}, }
• (2025). Decouple and rebalance: Towards effective joint learning for EEG-fNIRS integration. NEUCOM, 638, 130128. (www)

  @article{ , abstract = {Combining the complementary advantages of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) shows promising potential for enhancing the decoding performance of brain-computer interfaces (BCI). However, integrating these signals through commonly used joint learning frameworks often fails to achieve the expected results. The two primary reasons are: 1) Inadequate decoupling and effective utilization of diversity features; and 2) Imbalance in the learning of multimodal features. To overcome these limitations, this paper proposes a feature decoupling and modality rebalancing framework for EEG and fNIRS integration. Specifically, a dual-decoder architecture is introduced within the standard joint learning framework. The Query decoder is employed to decouple the intermodal coupling features, while the Key decoder focuses on decoupling modality-specific features. This approach mitigates signal interference and enhances the extraction of task-relevant information. Furthermore, we introduce a Gradient Rebalancing (GradReb) Strategy to regulate gradient flow during training, ensuring balanced learning across both modalities and mitigating the risk of one modality predominating the process. Experimental results demonstrate that our method effectively leverages the complementary information from both EEG and fNIRS, significantly improving the performance of the joint learning framework. The proposed joint framework provides a scalable solution for hybrid BCI systems and offers new insights for other multimodal applications.}, archive = {J_NEUCOM}, author = {Ming Meng and Danyang Hao and Yunyuan Gao and Luya Li and Maofeng Wang and Xugang Xi}, doi = {10.1016/j.neucom.2025.130128}, journal = {Neurocomputing}, month = {7}, pages = {130128}, shortjournal = {Neurocomputing}, title = {Decouple and rebalance: Towards effective joint learning for EEG-fNIRS integration}, volume = {638}, year = {2025}, }
• (2025). Improved cost function-based fault tolerant control for nonlinear systems with simultaneous faults. NEUCOM, 638, 130125. (www)

  @article{ , abstract = {This paper develops a fault-tolerant control (FTC) approach for nonlinear systems with simultaneous actuator and sensor faults by utilizing the adaptive dynamic programming technique. The developed approach integrates a descriptor observer with a neural network (NN) fault estimator, which enables simultaneous and precise estimations of system states and faults. By incorporating the estimated actuator fault information into the design of an improved cost function, the FTC problem can be transformed into a general optimal control problem. The Hamilton–Jacobi–Bellman equation for the nominal system (the fault-free system) is solved by constructing a critic NN. Thus, an approximate optimal control law is obtained to mitigate the impact of actuator faults. The faulty signal is replaced by the accurate estimation of the system states, thereby mitigating the influence of sensor faults. It is proven that the developed approach ensures the uniform ultimate boundedness of the closed-loop faulty system. Simulation studies are presented to illustrate the effectiveness of the developed approach.}, archive = {J_NEUCOM}, author = {Chujian Zeng and Bo Zhao and Derong Liu}, doi = {10.1016/j.neucom.2025.130125}, journal = {Neurocomputing}, month = {7}, pages = {130125}, shortjournal = {Neurocomputing}, title = {Improved cost function-based fault tolerant control for nonlinear systems with simultaneous faults}, volume = {638}, year = {2025}, }
• (2025). Debate divides: Argument relation-based contrastive opinion summarization via multi-task learning for online discussions. NEUCOM, 638, 130124. (www)

  @article{ , abstract = {Contrastive opinion summarization(COS) aims to generate summaries that can see both sides of an argument. Existing studies mostly focus on summarizing popular opinions expressed in comments, often ignoring divergent opinions. While some studies use sentiment-based frameworks to compare opinions on the same topic, relying solely on sentiment polarity proves inadequate for distinguishing nuanced differences in opinion. Additionally, many methods extract topics directly from comments without considering their relevance to the associated reading content. In this paper, we propose a novel Topic-level Contrastive Opinion Summarization model for online discussions, called ARCOS, which leverages argument relations between comments to identify divergent opinion pairs. Inspired by cognitive map theory, ARCOS first extracts topics from a keyword co-occurrence graph generated from the reading content. It then uses a multi-task learning network to predict argument relations between comments and align them with the extracted content topics. Based on argument relations, ARCOS selects representative contrastive comment pairs from comments on the same topic, and employs a large language model to produce a divergent opinion summary. Experimental results on a newly created benchmark dataset, RNCOS, show that ARCOS outperforms baselines in various sub-tasks and generates high-quality summaries for divergent opinions about specific content topics.}, archive = {J_NEUCOM}, author = {Huawei Shan and Dongyuan Lu}, doi = {10.1016/j.neucom.2025.130124}, journal = {Neurocomputing}, month = {7}, pages = {130124}, shortjournal = {Neurocomputing}, title = {Debate divides: Argument relation-based contrastive opinion summarization via multi-task learning for online discussions}, volume = {638}, year = {2025}, }
• (2025). Enhanced inter-camera person re-identification leveraging mixed-order relation-aware recurrent neural network. NEUCOM, 638, 130123. (www)

  @article{ , abstract = {The person Re-Identification (Re-ID) requires a significant quantity of the costly label information, whereas unsupervised ones are still unable to provide satisfactory identification performance. These results in the poor scalability due to the requirement of the laborious data collection and annotation process in real-world Re-id applications. Unsupervised Re-ID techniques not require identity label data, but have significantly worse and inadequate model performance. In this paper, Enhanced Inter-Camera Person Re-identification leveraging Mixed-Order Relation-Aware Recurrent Neural Network (EICPR-MORRNN-TTAO) is proposed. The input images are collected from Market-1501, MSMT17, and Duke MTMC-reID datasets. Afterward, the input image is supplied to pre-processing. In preprocessing, Unsharp Structure Guided Filtering (USGF) is employed to enhance image quality. The pre-processed image is supplied to classification phase for Re-identifying the Inter-Camera Person as Same and Different utilizing Mixed-Order Relation-Aware Recurrent Neural Network (MORRNN). Generally, MORRNN does not adopt any optimization methods to determine the ideal parameters to assure accurate person Re-identification. Hence, Triangulation Topology Aggregation Optimizer (TTAO) is proposed to enhance the weight parameters of MORRNN. The EICPR-MORRNN-TTAO method is implemented in Python. The metrics, like Mean Average Precision (MAP), Cumulative Matching Characteristic (CMC), recall, Rank-1, Rank-10, Rank-20, Entropy, error rate, and Receiver Operating Characteristic (ROC) is considered. The EICPR-MORRNN-TTAO method attains 23.10 %, 27.54 % and 25.72 %, higher mAP, 21.48 %, 17.73 %, 25.32 % higher CMC and 20.98 %, 26.66 % and 16.32 % lower Error rate, are compared with existing techniques, like Intra-camera supervised Re-ID (ICSP-RI-PR), Offline-online associated camera-aware proxies for unsupervised Re-ID(OSC-UPRI-EIC), and Unsupervised Re-ID with stochastic training strategy (UPRI-EIC-STS) respectively.}, archive = {J_NEUCOM}, author = {Vidhyalakshmi MK and Bhuvanesh Unhelkar and Pravin R. Kshirsagar and R. Thiagarajan}, doi = {10.1016/j.neucom.2025.130123}, journal = {Neurocomputing}, month = {7}, pages = {130123}, shortjournal = {Neurocomputing}, title = {Enhanced inter-camera person re-identification leveraging mixed-order relation-aware recurrent neural network}, volume = {638}, year = {2025}, }
• (2025). A multi-task learning framework with prompt-based hybrid alignment for vision-language model. NEUCOM, 638, 130121. (www)

  @article{ , abstract = {Large-scale pre-trained vision-language models have demonstrated remarkable performance across diverse domains, indicating substantial application potential. Empirical studies reveal specific cross-domain task correlations that remain underutilized in existing research, presenting a critical direction for future exploration. Previous work has established that joint multi-task training enhances individual task performance. Building upon established methodologies, we propose a novel multi-task framework with prompt-based alignment (MTFPA) that enables efficient co-tuning of pre-trained vision-language models using limited image data in multi-task scenarios. Leveraging CLIP as the backbone architecture for its superior downstream task performance, we introduce relevance-based task grouping and develop a multi-task sharing network (MTSN). This architecture enables individual tasks to acquire features from semantically related intra-group tasks while maintaining partial inter-group task correlations. The MTSN parameters undergo optimization through comprehensive joint training across all tasks. Extensive experiments on large-scale multi-task datasets demonstrate that MTFPA achieves superior performance compared to state-of-the-art approaches, confirming the exceptional potential and broad applicability of vision-language models in multi-task learning scenarios.}, archive = {J_NEUCOM}, author = {Jiabao Han and Wei Zhong and Cheng Ding and Cheng Wang}, doi = {10.1016/j.neucom.2025.130121}, journal = {Neurocomputing}, month = {7}, pages = {130121}, shortjournal = {Neurocomputing}, title = {A multi-task learning framework with prompt-based hybrid alignment for vision-language model}, volume = {638}, year = {2025}, }
• (2025). Multi-receptive field feature disentanglement with distance-aware gaussian brightness augmentation for single-source domain generalization in medical image segmentation. NEUCOM, 638, 130120. (www)

  @article{ , abstract = {Deep learning has made significant advancements in medical image segmentation. However, in practical applications, domain shift between the source and target domains can severely degrade model performance. Challenges such as data scarcity and privacy concerns in healthcare settings further complicate the situation. To address these issues, this paper proposes Multi-Receptive Field Feature Disentanglement (MRFFD), which integrates multi-scale feature extraction and channel-level feature disentanglement. It utilizes convolutional kernels of different sizes to capture both fine-grained details and global context from the image, while effectively separating style and structural features. This approach enhances the model’s robustness when facing domain-specific style variations. In addition, addressing the uneven brightness distribution in medical images, we propose a novel data augmentation method—Distance-Aware Gaussian Brightness Augmentation (DAGBA). This method dynamically adjusts the brightness by incorporating the distance from pixels to the foreground and image edges, thereby further enhancing the model’s adaptability. Experiments conducted on multiple medical image datasets, including prostate and fundus images, demonstrate that MRFFD significantly outperforms traditional single-source domain generalization methods as well as current state-of-the-art models. The proposed method exhibits superior cross-domain adaptation and robustness. This study provides a novel solution for domain generalization in medical image segmentation and highlights its potential for practical applications. Our code and models are made public at: https://github.com/Heyprion/MRFFD .}, archive = {J_NEUCOM}, author = {Ziyu Wang and Jia-xuan Jiang and Shubin Zeng and Ling Zhou and Yuee Li and Zhong Wang}, doi = {10.1016/j.neucom.2025.130120}, journal = {Neurocomputing}, month = {7}, pages = {130120}, shortjournal = {Neurocomputing}, title = {Multi-receptive field feature disentanglement with distance-aware gaussian brightness augmentation for single-source domain generalization in medical image segmentation}, volume = {638}, year = {2025}, }
• (2025). Efficient spatio-temporal modeling and text-enhanced prototype for few-shot action recognition. NEUCOM, 638, 130119. (www)

  @article{ , abstract = {Few-Shot Action Recognition (FSAR) aims to classify new action categories accurately with only a limited number of labeled samples. Current methods face challenges in capturing spatiotemporal dynamics and integrating multimodal information effectively. This work presents a new framework that improves FSAR performance by improving spatiotemporal modeling and integrating cross-modal semantics. To capture complex spatiotemporal relationships in videos, we introduce two complementary modules: Temporal Enhancement Adaptation (TEA), which enhances temporal modeling capability, and Spatio-Temporal Fusion Adaptation (STFA), which integrates spatial and temporal features for better representations. Additionally, we propose the Text-Enhanced Prototype Module (TEPM), which strengthens prototype representations by fusing textual and visual features at multiple levels, improving the discriminability and generalization of prototypes. The experiments show that our approach achieves competitive performance on various benchmark datasets, confirming its effectiveness in FSAR.}, archive = {J_NEUCOM}, author = {Qian Zhang and Shuo Yan and Mingwen Shao and Hong Liang}, doi = {10.1016/j.neucom.2025.130119}, journal = {Neurocomputing}, month = {7}, pages = {130119}, shortjournal = {Neurocomputing}, title = {Efficient spatio-temporal modeling and text-enhanced prototype for few-shot action recognition}, volume = {638}, year = {2025}, }
• (2025). SAVETrack: Scene-adaptive visual enhancement transformer for robust tracking. NEUCOM, 638, 130118. (www)

  @article{ , abstract = {Visual object tracking remains a fundamental yet challenging computer vision task, particularly in complex scenarios involving occlusions, deformations, and appearance variations. Current approaches exhibit three critical limitations: fixed feature extraction strategies, insufficient multi-scale context modeling, and weak temporal feature correspondence. To address these challenges, we propose SAVETrack, a novel Scene-Adaptive Visual Enhancement Transformer framework with three key technical innovations: (1) a Scene-Adaptive Feature Selection module that dynamically processes features through specialized branches based on scene context, (2) a Hierarchical Context-Guided Feature Enhancement module that captures multi-scale relationships via progressive refinement, and (3) a Dynamic Feature Propagation and Optimization Network that maintains reliable temporal correspondences through quality-guided feature alignment. Extensive experiments on seven challenging benchmarks demonstrate that SAVETrack achieves state-of-the-art performance, with AUC scores of 72.2% on LaSOT and 85.3% on TrackingNet, while maintaining a real-time inference speed of 32 FPS. The proposed framework advances visual tracking by providing a unified solution that effectively addresses core challenges through adaptive feature processing and enhancement mechanisms.}, archive = {J_NEUCOM}, author = {Yongjun Wang and Xiaohui Hao}, doi = {10.1016/j.neucom.2025.130118}, journal = {Neurocomputing}, month = {7}, pages = {130118}, shortjournal = {Neurocomputing}, title = {SAVETrack: Scene-adaptive visual enhancement transformer for robust tracking}, volume = {638}, year = {2025}, }
• (2025). Urban traffic flow prediction based on multi-spatio-temporal feature fusion. NEUCOM, 638, 130117. (www)

  @article{ , abstract = {Urban traffic flow exhibits non-linear characteristics and highly complex spatio-temporal correlations, posing a huge challenge to traffic flow prediction. To address this challenge, we constructed a novel Multi-Spatio-Temporal Feature Fusion (MSTFF) model. This model accurately depicts the spatio-temporal correlations of traffic flow from multiple spatio-temporal perspectives. It decomposes the non-linear traffic flow into stable trend and random fluctuation, and decomposes the topological graph of the road network into first-order adjacency, second-order in-degree and second-order out-degree relationship graphs which describe the connectivity, input and output relationship of the nodes in road network respectively, thus constructing multiple perspectives that can exhibit different spatio-temporal properties. For each perspective, we use the self-attention mechanism to learn the long-term spatio-temporal correlations in the traffic flow, and use convolutional neural networks to learn the short-term spatio-temporal correlations. By fusing the features of spatio-temporal dependencies from multiple perspectives, we explored the high-dimensional spatio-temporal features of traffic flow and predicted the future traffic flow. We also developed the traffic flow prediction algorithm based on MSTFF, verified its performance using four real-world datasets, and compared it with twelve baseline models in the past five years. Experimental results show that the performance of the MSTFF model far exceeds that of existing related models, with average improvement rates of 15.10%, 19.63%, and 13.59% respectively in terms of the three indicators of MAE, RMSE, and MAPE.}, archive = {J_NEUCOM}, author = {Hui Chen and Jian Huang and Yong Lu and Jijie Huang}, doi = {10.1016/j.neucom.2025.130117}, journal = {Neurocomputing}, month = {7}, pages = {130117}, shortjournal = {Neurocomputing}, title = {Urban traffic flow prediction based on multi-spatio-temporal feature fusion}, volume = {638}, year = {2025}, }
• (2025). A reinforcement learning approach for optimized MRI sampling with region-specific fidelity. NEUCOM, 638, 130116. (www)

  @article{ , abstract = {Accelerating Magnetic Resonance Imaging (MRI) acquisition while preserving diagnostic quality remains a significant challenge in medical imaging. This paper introduces a novel reinforcement learning approach for optimizing k-space sampling, addressing the critical need for maintaining high-fidelity reconstructions, particularly in clinically significant regions. Our method bridges k-space and image domains by employing a multi-layer Fast Fourier Transform (FFT) network architecture coupled with a comprehensive reward function. The reward function integrates global SSIM, region-specific metrics, and k-space MSE, achieving SSIM scores of 0.9764 (global) and PSNR of 43.427 in cardiac imaging, with a Dice score of 0.9606 for critical regions. This function uniquely balances global image fidelity with region-specific accuracy, ensuring optimal sampling in areas of high diagnostic value. We present a reinforcement learning framework that maintains consistency with MRI physics throughout the optimization process by utilizing Proximal Policy Optimization (PPO) for concurrent refinement of policy and value networks, while minimizing unnecessary domain transformations. We validate our approach on two diverse datasets: the ACDC cardiac MRI and the FastMRI knee dataset, demonstrating its effectiveness across different regions of clinical interest and showing consistent improvements across 4x and 6x acceleration factors. The proposed k-space sampling optimization technique not only enhances reconstruction quality but also serves as a versatile front-end tool, adaptable to various MRI reconstruction algorithms. The code is publicly available at https://github.com/Ruru-Xu/KSRO}, archive = {J_NEUCOM}, author = {Ruru Xu and Ilkay Oksuz}, doi = {10.1016/j.neucom.2025.130116}, journal = {Neurocomputing}, month = {7}, pages = {130116}, shortjournal = {Neurocomputing}, title = {A reinforcement learning approach for optimized MRI sampling with region-specific fidelity}, volume = {638}, year = {2025}, }
• (2025). Be your own doctor: Temperature scaling self-knowledge distillation for medical image classification. NEUCOM, 638, 130115. (www)

  @article{ , abstract = {Self-knowledge distillation (self-KD), which uses the student network as the teacher model, allows the model to learn knowledge by itself. It has been widely studied in various medical image tasks for constructing lightweight models to alleviate the limitations of computing resources. However, existing self-KD methods use a single temperature for distillation, ignoring the effect of temperature on different classes. In this paper, we investigate the effects of target class temperature and non-target class temperature on the performance of self-KD. Based on the above study, a temperature scaling self-knowledge distillation (TSS-KD) model is proposed, which can better balance the target class knowledge and non-target class knowledge. By adjusting the temperature scaling of different classes, the model can learn better representations by distilling the well-proportioned features. To make the network focus more on the local lesions of medical images, a regional gamma augmentation (RGA) method is proposed, which provides stronger perturbations to the same sample to generate more differentiated features. By self-regularizing the consistency of these features, the model can learn more local knowledge. To evaluate the effectiveness of the proposed method, extensive experiments are conducted on nine medical image classification tasks of eight public datasets. Experimental results show that the proposed method outperforms state-of-the-art self-KD models and has strong generality. The code is available at https://github.com/JeaneyLau/TSS-KD .}, archive = {J_NEUCOM}, author = {Wenjie Liu and Lei Zhang and Xianliang Zhang and Xinyang Zhou and Xin Wei}, doi = {10.1016/j.neucom.2025.130115}, journal = {Neurocomputing}, month = {7}, pages = {130115}, shortjournal = {Neurocomputing}, title = {Be your own doctor: Temperature scaling self-knowledge distillation for medical image classification}, volume = {638}, year = {2025}, }
• (2025). Robust scene aware multi-object tracking for surveillance videos. NEUCOM, 638, 130114. (www)

  @article{ , abstract = {Multiple object tracking (MOT) in dynamic environments poses significant challenges due to occlusions, varying object appearances, and complex motion patterns. These difficulties are further compounded in surveillance videos with low quality and distant targets, especially when abnormal movements and events occur within the scene. To address these challenges, this paper proposes two novel frameworks that integrates scene information as prior knowledge into the MOT process. Using scene information enhances object association and tracking accuracy. The first algorithm, called Scene-Aware Adaptive Tracking Algorithm (SAATA), improves MOT by introducing a target-centric search area and employing the novel state assessment techniques for comprehensive scene analysis. The state assessment aspect involves a thorough analysis of the targets’ current state to a more accurate and context-aware tracking system that can handle occlusion. One notable advantage of SAATA is its ability to significantly reduce identity switches and track fragmentation compared to the conventional algorithms. The second algorithm, called Abnormality-Aware Tracking Algorithm (AATA), leverages the scene's normality or abnormality to label detections and tracks, incorporating this insight as prior knowledge into the construction of the cost matrix for association. Consequently, if a target exhibits abnormal behavior, the tracker is alerted and adjusts its actions to prevent loss of the abnormal target. AATA demonstrates enhanced performance in terms of multiple object tracking accuracy (MOTA), compared to other trackers, showing its ability to effectively track abnormal targets. Furthermore, through experiments and evaluations on diverse datasets, the robustness and efficiency of the two algorithms are demonstrated.}, archive = {J_NEUCOM}, author = {Fatemeh Jalali and Morteza Khademi and Abbas Ebrahimi Moghadam and Hadi Sadoghi Yazdi}, doi = {10.1016/j.neucom.2025.130114}, journal = {Neurocomputing}, month = {7}, pages = {130114}, shortjournal = {Neurocomputing}, title = {Robust scene aware multi-object tracking for surveillance videos}, volume = {638}, year = {2025}, }
• (2025). FS-net collaborative flattening transformation for high performance 3D-to-2D vessel segmentation in OCTA images. NEUCOM, 638, 130113. (www)

  @article{ , abstract = {Optical coherence tomography angiography (OCTA) represents a novel and highly effective imaging modality that is widely employed in the examination and evaluation of a range of fundus vascular diseases. Accurate segmentation of the fundus vascular network is vital for diagnosing and preventing ophthalmic diseases. However, air noise and artifact noise are far more abundant than vascular signals and are concentrated in specific layers within the 3D OCTA image. Current 3D-to-2D segmentation models require additional costs to eliminate noise. Furthermore, these models operate as a one-way process, resulting in underlying feature maps from the 3D pathway being skipped. Herein, we propose a novel model that integrates a 3D OCTA image preprocessing framework (“flattening transformation”) with a 3D-to-2D segmentation network (FS-Net) to maximize the utilization of 3D OCTA data while normalizing image structure. Specifically, the “flattening transformation” is used to extract OCTA layers containing the blood flow signals, which reduces data complexity. We proposed a feature extraction module that compresses the skipped feature maps from the 3D pathway into 2D and concatenates them with the feature maps from the 2D pathway. It is shown that the proposed model significantly outperforms the current state-of-the-art methods, with a dice coefficient of 93.03%, and also exceeds the performance of existing models in terms of generalization ability. The code is available at https://github.com/xingjw666/FS-Net.git.}, archive = {J_NEUCOM}, author = {Jiwei Xing and Songyi Jiang and Jing Xu and Dongbei Guo and Jinde Zhang and Peishan Li and Heng Zhang and Linyan Xue and Valery V. Tuchin and Xiufei Gu and Ronghui Li and Qingliang Zhao}, doi = {10.1016/j.neucom.2025.130113}, journal = {Neurocomputing}, month = {7}, pages = {130113}, shortjournal = {Neurocomputing}, title = {FS-net collaborative flattening transformation for high performance 3D-to-2D vessel segmentation in OCTA images}, volume = {638}, year = {2025}, }
• (2025). Information bottleneck with input sampling for attribution. NEUCOM, 638, 130112. (www)

  @article{ , abstract = {In order to facilitate the adoption of deep learning in areas where decisions are of critical importance, understanding the model’s internal workings is paramount. Nevertheless, since most models are considered black boxes, this task is usually not trivial, especially when the user does not have access to the network’s intermediate outputs. In this paper, we propose IBISA, a model-agnostic attribution method that reaches state-of-the-art performance by optimizing sampling masks using the Information Bottleneck Principle. Our method improves on the previously known RISE and IBA techniques by placing the bottleneck right after the image input without complex formulations to estimate the mutual information. The method also requires only twenty forward passes and ten backward passes through the network, which is significantly faster than RISE, which needs at least 4000 forward passes. We evaluated IBISA using a VGG-16 and a ResNET-50 model, showing that our method produces explanations comparable or superior to IBA, RISE, and Grad-CAM but much more efficiently.}, archive = {J_NEUCOM}, author = {Bruno Coelho and Jaime S. Cardoso}, doi = {10.1016/j.neucom.2025.130112}, journal = {Neurocomputing}, month = {7}, pages = {130112}, shortjournal = {Neurocomputing}, title = {Information bottleneck with input sampling for attribution}, volume = {638}, year = {2025}, }
• (2025). TransCMFD: An adaptive transformer for copy-move forgery detection. NEUCOM, 638, 130110. (www)

  @article{ , abstract = {Copy-move forgery is one of the most common image tampering methods. It is a frequently employed method for manipulating evidence or deceiving the public by hiding some objects in an image or replicating significant objects. Therefore, it is crucial to focus on copy-move forgery detection. In this paper, we propose TransCMFD as a new transformer-based model for detecting copy-move forgery. We propose an adaptive transformer encoder and combine the traditional convolution encoder–decoder to capture different global and local features of the forgery image, respectively. This can enhance the model’s comprehension of the impact of forgery across different tampered image regions. To allow the model to concentrate more on the tampered regions that resemble the original regions, we introduce a similarity detection module. Moreover, to enhance the localization accuracy of the tampered regions, we design an adaptive loss function combination strategy that incorporates the DICE coefficient loss and binary cross-entropy loss. We perform comprehensive experiments on both synthetic and four publicly available datasets. The results show that our model has better performance in copy-move forgery detection compared to baseline methods, and it has remarkable robustness to some common image attacks such as noise addition attacks, image blurring attacks, and color reduction attacks.}, archive = {J_NEUCOM}, author = {Enji Liang and Kuiyuan Zhang and Zhongyun Hua and Yuanman Li and Xiaohua Jia}, doi = {10.1016/j.neucom.2025.130110}, journal = {Neurocomputing}, month = {7}, pages = {130110}, shortjournal = {Neurocomputing}, title = {TransCMFD: An adaptive transformer for copy-move forgery detection}, volume = {638}, year = {2025}, }
• (2025). LLM-guided decision-making toolkit for multi-agent reinforcement learning. NEUCOM, 638, 130105. (www)

  @article{ , abstract = {Complex cooperative game tasks pose significant challenges to decision making in multi-agent systems. Despite advancements in multi-agent reinforcement learning (MARL), these systems often struggle in the initial learning phase, known as the cold-start problem, where devising effective strategies from scratch is difficult. To address this, we leverage the advanced reasoning capabilities of Large Language Models (LLMs) by introducing prompt learning in complex cooperative games. In our approach, tensor-formatted environmental information is converted into natural language and fed into LLMs during early exploration stages in MARL algorithms. The strategic suggestions generated by LLMs, combined with trajectories from random exploration, form an experience pool for initial training. Our experiments demonstrate that this approach significantly enhances cooperation among agents in complex games and mitigates the cold-start issue. Notably, we have developed a toolkit compatible with the StarCraft Multi-Agent Challenge (SMAC) environment, providing seamless integration between SMAC APIs, LLMs, and MARL algorithms. This includes automated prompt generation for any map and alignment of natural language strategies with tensor-formatted data. In addition, considering the importance of effective communication for enhancing collaboration efficiency, our toolkit includes a robust communication module, which supports the training, testing, and visualization of communication algorithms. We provide the source code on GitHub at https://github.com/lizhemin18/pymarl_LLM .}, archive = {J_NEUCOM}, author = {Zhemin Li and Ruobing Zhang and Zhengming Wang and Zheng Xie and Yiping Song}, doi = {10.1016/j.neucom.2025.130105}, journal = {Neurocomputing}, month = {7}, pages = {130105}, shortjournal = {Neurocomputing}, title = {LLM-guided decision-making toolkit for multi-agent reinforcement learning}, volume = {638}, year = {2025}, }
• (2025). FRIH: Fully reversible image hiding by invertible network. NEUCOM, 638, 130104. (www)

  @article{ , abstract = {Image hiding is the study of techniques for covert storage and transmission, which embeds a secret image into a container image and generates stego image to make it similar in appearance to a normal image. However, existing image hiding methods have a serious problem that the hiding and revealing process cannot be fully invertible, which results in the revealing network not being able to recover the secret image losslessly, which makes it impossible to simultaneously achieve high fidelity and secure transmission of the secret image in an insecure network environment. To solve this problem, this paper proposes a completely reversible image hiding architecture based on reversible neural network, which only inputs secret images into the hiding network without information loss, aiming to realize reversible hiding of secret images and realize reversible on both data and network. Based on this ingenious architecture, the method can withstand deep learning based image steganalysis. In addition, we also propose a field pixel filling method and an analog noise layer to enhance the loaded image after interference during transmission to improve its robustness. Experiments demonstrate that the FRIH proposed in this paper significantly outperforms other state-of-the-art image hiding methods in hiding a single image, and also significantly outperforms other state-of-the-art methods in robustness and security.}, archive = {J_NEUCOM}, author = {Lin Huo and Lang Huang and Zheng Gan and Xinrong He}, doi = {10.1016/j.neucom.2025.130104}, journal = {Neurocomputing}, month = {7}, pages = {130104}, shortjournal = {Neurocomputing}, title = {FRIH: Fully reversible image hiding by invertible network}, volume = {638}, year = {2025}, }
• (2025). Generative prompt controlled diffusion for weakly supervised semantic segmentation. NEUCOM, 638, 130103. (www)

  @article{ , abstract = {Weakly supervised semantic segmentation (WSSS), aiming to train segmentation models solely using image-level labels, has received significant attention. Existing approaches mainly concentrate on creating high-quality pseudo labels by utilizing existing images and their corresponding image-level labels. However, a major challenge arises when the available dataset is limited, as the quality of pseudo labels degrades significantly. In this paper, we tackle this challenge from a different perspective by introducing a novel approach called Generative Prompt Controlled Diffusion (GPCD) for data augmentation. This approach enhances the current labeled datasets by augmenting them with a variety of images, achieved through controlled diffusion guided by Generative Pre-trained Transformer (GPT) prompts. In this process, the existing images and image-level labels provide the necessary control information, while GPT enriches the prompts to generate diverse backgrounds. Moreover, we make an original contribution by integrating data source information as tokens into the Vision Transformer (ViT) framework, which improves the ability of downstream WSSS models to recognize the origins of augmented images. Our proposed GPCD approach clearly surpasses existing state-of-the-art methods, with its advantages being more pronounced when the available data is scarce, thereby demonstrating the effectiveness of our method. Our source code will be released.}, archive = {J_NEUCOM}, author = {Wangyu Wu and Tianhong Dai and Zhenhong Chen and Xiaowei Huang and Fei Ma and Jimin Xiao}, doi = {10.1016/j.neucom.2025.130103}, journal = {Neurocomputing}, month = {7}, pages = {130103}, shortjournal = {Neurocomputing}, title = {Generative prompt controlled diffusion for weakly supervised semantic segmentation}, volume = {638}, year = {2025}, }
• (2025). FCAnet: A novel feature fusion approach to EEG emotion recognition based on cross-attention networks. NEUCOM, 638, 130102. (www)

  @article{ , abstract = {Recognition of human emotions from complex, diverse, and variable electroencephalogram (EEG) signals based on feature fusion strategies is gaining significant attention in affective computing. However, existing strategies employ a unified learning paradigm to integrate multiple features, and they lead to the neglect of critical local patterns and intricate relationships among features. Furthermore, most of models are generally designed to be complex and difficult to leverage and interpret in practical applications. To address these issues, FCAnet, a novel generic cross-attention-based feature fusion approach is proposed. FCAnet views the embedding space of multiple features as interactive patterns and utilizes cross-attention mechanisms to analyze both spatial correlations and discrepancy information simultaneously. Specifically, a dual-branch feature extraction module (DBFE) is first designed to effectively capture the 3D differential entropy (DE) and 3D power spectral density (PSD) feature maps of EEG. Secondly, a cross-attention feature fusion network (CAFFN) integrates a designed discrepancy information injection block (DIIB) with a common information injection block (CIIB) unit, facilitating significant interaction between different features. Finally, a time augmentation block (TAB) is employed to recover information lost from high-level representations, reusing discriminative temporal feature maps. Experimental results on the datasets DEAP, SEED, SEED-IV, and MPED demonstrate that the proposed FCANet outperforms state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Mingjie Li and Heming Huang and Kedi Huang}, doi = {10.1016/j.neucom.2025.130102}, journal = {Neurocomputing}, month = {7}, pages = {130102}, shortjournal = {Neurocomputing}, title = {FCAnet: A novel feature fusion approach to EEG emotion recognition based on cross-attention networks}, volume = {638}, year = {2025}, }
• (2025). Improving chinese character representation with formation tree. NEUCOM, 638, 130098. (www)

  @article{ , abstract = {Learning effective representations for Chinese characters presents unique challenges, primarily due to the vast number of characters and their continuous growth, necessitating models that can handle an expanding category space. Additionally, the inherent sparsity of character usage complicates the generalization of learned representations. Prior research has explored radical-based sequences to overcome these issues, achieving progress in recognizing unseen characters. However, these approaches fail to fully exploit the inherent tree structure of such sequences. To address these limitations and leverage established data properties, we propose Formation Tree-CLIP (FT-CLIP). FT-CLIP utilizes formation trees to represent characters and incorporates a dedicated tree encoder, significantly improving performance in both seen and unseen character recognition tasks. We further introduce masking for both character images and tree nodes, enabling efficient and effective training. This approach accelerates training significantly (by a factor of two or more) while enhancing accuracy. Extensive experiments show that processing characters through formation trees aligns better with their inherent properties than direct sequential methods, significantly enhancing the generality and usability of the representations.}, archive = {J_NEUCOM}, author = {Yang Hong and Xiaojun Qiao and Yinfei Li and Rui Li and Junsong Zhang}, doi = {10.1016/j.neucom.2025.130098}, journal = {Neurocomputing}, month = {7}, pages = {130098}, shortjournal = {Neurocomputing}, title = {Improving chinese character representation with formation tree}, volume = {638}, year = {2025}, }
• (2025). Adaptive knowledge transfer for data-free low-bit quantization via tiered collaborative learning. NEUCOM, 638, 130097. (www)

  @article{ , abstract = {Efficient deployment of deep learning models on resource-constrained edge devices often requires reducing model complexity and memory usage, with low-bit quantization (LBQ) emerging as a promising solution. However, data-free quantization approaches frequently struggle with the knowledge gap between full-precision (FP) and low-bit-width models, which can degrade performance. This paper introduces “Adaptive Knowledge Transfer for Data-Free Low-Bit Quantization via Tiered Collaborative Learning” (abbreviated as TieredQ), a novel framework that leverages high-bit-width (HBQ) models as intermediate stages to facilitate progressive knowledge transfer and effectively bridge this gap. Our approach utilizes a Tiered Collaborative Learning (TCL) strategy, where HBQ models serve as both intermediate “student” models and auxiliary “teacher” models, thereby enhancing knowledge transfer to LBQ models. This hierarchical method minimizes information loss and improves model accuracy, even without labeled data. A key component of our framework is the Adaptive Loss Controller (ALC), which dynamically adjusts loss weights to balance contributions across different bit-width models, optimizing for the specific needs of each quantization level. The framework operates in two stages: first, we generate high-quality synthetic data by optimizing noise samples to promote internal diversity and external distinction in convolutional neural networks (CNNs) and to reflect sample difficulty in vision transformers (ViTs). In the second stage, we implement tiered collaborative learning with the ALC, utilizing Kullback–Leibler (KL) divergence to stabilize training, accelerate convergence, and enhance overall performance. Extensive evaluations on both CNN and ViT architectures across large-scale image classification tasks confirm the efficacy of our method, particularly for challenging LBQ models. Notably, on the ImageNet1K dataset, our framework achieves significant accuracy improvements of 7.15% on MobileNetV1 and 10.6% on DeiT-T, demonstrating its effectiveness and superior performance in data-free quantization.}, archive = {J_NEUCOM}, author = {Zelin Yang and Tong Lin and Chenyang Li and Bo Qian and Xinyu Yang and Xing Wei}, doi = {10.1016/j.neucom.2025.130097}, journal = {Neurocomputing}, month = {7}, pages = {130097}, shortjournal = {Neurocomputing}, title = {Adaptive knowledge transfer for data-free low-bit quantization via tiered collaborative learning}, volume = {638}, year = {2025}, }
• (2025). Optimal bipartite consensus for multi-agent systems using twin Q-learning deterministic policy gradient algorithm with adaptive learning rate. NEUCOM, 638, 130096. (www)

  @article{ , abstract = {We investigate the optimal bipartite consensus control (OBCC) problem for multi-agent systems (MASs) over a signed network. Due to the improper cooperation–competition strength (CCS) among agents, the system may be unstable or even non-convergent. Recognizing the close relationship between CCS and the training of the critic network, we propose a twin Q-learning deterministic policy gradient algorithm with an adaptive learning rate (ALR-TQDPG). First, an adaptive learning rate formula is established based on the CCS and historical temporal difference (TD) error variations. The weights of two factors are dynamically adjusted using the weight equation as training progresses, then dynamically adjusting the update magnitude (i.e., learning rate) of critic network weights. Second, to solve the underestimation problem of Q-value, a twin Q-learning algorithm is adopted to improve system performance. The addition of experience replay and target network methods enhances algorithm stability. Lyapunov stability theory and functional analysis are utilized to ensure the ALR-TQDPG algorithm’s convergence. Finally, numerical simulations confirm that the suggested approach is effective.}, archive = {J_NEUCOM}, author = {Lianghao Ji and Jiali Song and Cuijuan Zhang and Shasha Yang and Jun Li}, doi = {10.1016/j.neucom.2025.130096}, journal = {Neurocomputing}, month = {7}, pages = {130096}, shortjournal = {Neurocomputing}, title = {Optimal bipartite consensus for multi-agent systems using twin Q-learning deterministic policy gradient algorithm with adaptive learning rate}, volume = {638}, year = {2025}, }
• (2025). Optimal decorrelated score subsampling for cox regression with massive survival data. NEUCOM, 638, 130090. (www)

  @article{ , abstract = {This paper investigates optimal subsampling strategies for the preconceived low-dimensional parameters of main interest in the presence of the nuisance parameters for Cox regression with massive survival data. A general subsampling decorrelated score function based on the log-partial likelihood is constructed to reduce the influence of the less accurate nuisance parameter estimation with a possibly slow convergence rate. The consistency and asymptotic normality of the resultant subsample estimators are established. We derive unified optimal subsampling probabilities based on A- and L-optimality criteria. A two-step algorithm is further proposed to implement practically, and the asymptotic properties of the resultant estimators are also given. The satisfactory performance of our proposed subsample estimators is demonstrated by simulation results and an airline dataset.}, archive = {J_NEUCOM}, author = {Yujing Shao and Zhaohan Hou and Lei Wang and Heng Lian}, doi = {10.1016/j.neucom.2025.130090}, journal = {Neurocomputing}, month = {7}, pages = {130090}, shortjournal = {Neurocomputing}, title = {Optimal decorrelated score subsampling for cox regression with massive survival data}, volume = {638}, year = {2025}, }
• (2025). A hierarchical neural network structure and collaborative learning for incomplete data classification. NEUCOM, 638, 130081. (www)

  @article{ , abstract = {The existence of missing values increases the difficulty of classification in real world applications. The classification of incomplete data often requires the imputation of missing values. In the process of constructing the classifier, to make full use of the existing attribute values of incomplete samples, we introduce a tracking-removed autoencoder as the imputation model, which imputes missing values by constructing the mutual fitting relationship between attributes on one network structure. On this basis, we propose an imputation-classification hierarchical neural network structure, which realizes the classification of incomplete data by cascading fully connected layers to the tracking-removed autoencoder. For the training of the hierarchical neural network, we comprehensively consider regression error and classification error to design a cost function. Meanwhile, we propose a collaborative learning method that regards missing values as variables and dynamically optimizes them together with network parameters. This training scheme enables the hierarchical neural network to simultaneously match the existing attribute values and class structure of an incomplete dataset, thereby effectively improving the performance of the classifier. The experimental results after testing on several incomplete datasets demonstrate the effectiveness of the proposed imputation-classification hierarchical structure and collaborative learning method.}, archive = {J_NEUCOM}, author = {Zhi Sun and Hang Lu and Liyong Zhang and Xiaochen Lai and Wei Lu}, doi = {10.1016/j.neucom.2025.130081}, journal = {Neurocomputing}, month = {7}, pages = {130081}, shortjournal = {Neurocomputing}, title = {A hierarchical neural network structure and collaborative learning for incomplete data classification}, volume = {638}, year = {2025}, }
• (2025). Heteroscedastic ensemble deep random vector functional link neural network with multiple output layers for high frequency volatility forecasting and risk assessment. NEUCOM, 638, 130078. (www)

  @article{ , abstract = {Accurate volatility forecasting is crucial for the efficient management of financial systems. However, the dynamic nature and significant variability in financial time series data pose substantial challenges to achieving these forecasts. The paper introduces a novel Heteroscedastic ensemble deep random vector functional link (HedRVFL) network with multiple output layers for high frequency volatility forecasting and risk assessment. The hidden layers of the model are hierarchical and stacked for deep representation learning to extract complex patterns within the data. The neuron pruning strategy is utilized to eliminate noisy information from random features, thereby improving the network’s performance. The forecast is generated by combining the outputs of each layer through an ensemble method. A comparative analysis was conducted against several existing forecasting methods, utilizing error metrics and statistical tests on sixteen high frequency cryptocurrency time-series datasets, demonstrating that the proposed model outperforms others in terms of forecasting accuracy and risk assessment.}, archive = {J_NEUCOM}, author = {Aryan Bhambu and Ponnuthurai Nagaratnam Suganthan and Selvaraju Natarajan}, doi = {10.1016/j.neucom.2025.130078}, journal = {Neurocomputing}, month = {7}, pages = {130078}, shortjournal = {Neurocomputing}, title = {Heteroscedastic ensemble deep random vector functional link neural network with multiple output layers for high frequency volatility forecasting and risk assessment}, volume = {638}, year = {2025}, }
• (2025). AttenStyler: Text-image style transfer based on attention mechanism. NEUCOM, 638, 130075. (www)

  @article{ , abstract = {Currently, traditional image style transfer methods require the provision of a style image during training. If users do not have a corresponding style image or want to transfer an abstract or imagined style, these methods are not effective. With the introduction of the CLIP model, some text-based style transfer methods have been proposed, but these methods rely on pretrained generative models and do not effectively utilize feature information from different channels. To address these issues, we propose an attention-based text style transfer model called AttenStyler. We use a ”bigger-first, same-later” strategy to construct a biased channel attention module, which encourages the model to better focus on different channel features during training, thereby enhancing the model’s expressive capabilities. Additionally, we find that traditional bilinear interpolation upsampling struggles to learn content and style information for different images. Therefore, we use transpose convolutions instead and employ convolutions between skip connections to enhance the model’s information fusion capabilities. The paper includes qualitative analysis, quantitative analysis, and user studies to validate the effectiveness of the model. Through extensive experiments, it is demonstrated that this model can generate higher-quality style images and outperforms several representative methods mentioned in the paper.}, archive = {J_NEUCOM}, author = {Yongsheng Dong and Shichao Fan and Lintao Zheng and Mingchuan Zhang and Qingtao Wu}, doi = {10.1016/j.neucom.2025.130075}, journal = {Neurocomputing}, month = {7}, pages = {130075}, shortjournal = {Neurocomputing}, title = {AttenStyler: Text-image style transfer based on attention mechanism}, volume = {638}, year = {2025}, }
• (2025). FastTalker: An unified framework for generating speech and conversational gestures from text. NEUCOM, 638, 130074. (www)

  @article{ , abstract = {Generating 3D human gestures and speech from a text script is critical for creating realistic talking avatars. One typical approach is to use separate pipelines for text-to-speech (TTS) and speech-to-gesture (STG), but this often leads to mismatched speech and gestures and slow processing. In this paper, we introduce FastTalker, an efficient framework that generates high-quality speech and 3D gestures together with fast inference speeds. Our main insight is to reuse intermediate features from speech synthesis for gesture generation, as these features capture rhythmic details better than those re-extracted from generated speech. Specifically, 1) we present an end-to-end system that produces speech waveforms and full-body gestures at the same time, directly using intermediate speech features—like pitch, onset, energy, and duration—for gesture decoding; 2) we redesign the causal network architecture to avoid relying on future inputs, making it suitable for practical use; 3) we apply Reinforcement Learning-based Neural Architecture Search (NAS) to refine our network, improving both its performance and speed. Experiments on the BEAT2 dataset show that FastTalker sets new standards in both speech synthesis and gesture generation, achieving a processing speed of 0.17 s per second on an NVIDIA 3090.}, archive = {J_NEUCOM}, author = {Jian Zhang and Zixin Guo and Minggui He and Osamu Yoshie}, doi = {10.1016/j.neucom.2025.130074}, journal = {Neurocomputing}, month = {7}, pages = {130074}, shortjournal = {Neurocomputing}, title = {FastTalker: An unified framework for generating speech and conversational gestures from text}, volume = {638}, year = {2025}, }
• (2025). Stage-wise multi-focus fusion for numerous misaligned industrial images. NEUCOM, 638, 130062. (www)

  @article{ , abstract = {Multi-focus image fusion is a technique that combines information from multiple images to generate a single composite image that retains all the essential details and features of the original images. Traditional methods can achieve rapid fusion but struggle with image misalignment. Due to the strong expressive power of deep neural networks, deep learning methods can handle misaligned situations but face challenges when fusing many images. In this work, a stage-wise multi-focus fusion method is proposed. Firstly, in the coarse fusion stage, the number of images for subsequent fusion is reduced by fusing similar images. Here, several consecutive images are sequentially fused by using an indicator to distinguish between clear and blurred areas and then merging the clear regions. Subsequently, in the fine fusion stage, pairwise merging is adopted to obtain the final result. For each pair of sub-images, a blur filter, difference operator, and guided filter are utilized to create a decision map, followed by pixel-wise weighted averaging to fuse the source images. The proposed method is validated using two multi-focus image datasets with numerous misaligned industrial images, namely the Lampwick dataset and the Impurity dataset. Moreover, its generalization ability is demonstrated by conducting experiments on the Lytro and MFFW datasets. Our findings show that, despite the industrial images often possessing the characteristics of being numerous and misaligned, our method not only achieves fast fusion but also yields the best results.}, archive = {J_NEUCOM}, author = {Wenjie Zhu and Xurong Chi and Jingrun Chen}, doi = {10.1016/j.neucom.2025.130062}, journal = {Neurocomputing}, month = {7}, pages = {130062}, shortjournal = {Neurocomputing}, title = {Stage-wise multi-focus fusion for numerous misaligned industrial images}, volume = {638}, year = {2025}, }
• (2025). NeuroSimWorm: A multisensory framework for modeling and simulating neural circuits of caenorhabditis elegans. NEUCOM, 638, 130055. (www)

  @article{ , abstract = {Biological behaviors emerge from the dynamic interplay among the inner neurodynamic system, embodied mechanical structure, and external environmental inputs. Nonetheless, existing approaches simply consider the static brain model that cannot fully exploit the potential of continuous interaction and feedback from the body and the environment. To address these problems, we introduce NeuroSimWorm , a multisensory closed-loop neural circuit simulation approach of the widely studied organism, Caenorhabditis elegans ( C. elegans ). The full closed-loop simulation platform integrates four key subcomponents including Environment, Neural Computing, Biomechanical Model, and Visualization modules. We initially define multiple sensory environments for chemical, mechanical, and thermal stimuli. Subsequently, we construct four types of neural circuits including Locomotion, Chemosensation, Thermosensation, and Mechanosensation. Fitness functions based on multisensory optimization are proposed, enabling the virtual nematode to achieve complex intelligent behaviors such as autonomous locomotion, foraging, thermosensory regulation, and tactile avoidance. Thus, NeuroSimWorm presents a feasible way to understand the mechanism of biological intelligence by modeling the connectome and simulating behaviors in the physical surroundings.}, archive = {J_NEUCOM}, author = {Jiaxin Wang and Mengxiao Zhang and Kejun Wang and Lijun Kang and Liang Feng and Gang Pan and Huajin Tang}, doi = {10.1016/j.neucom.2025.130055}, journal = {Neurocomputing}, month = {7}, pages = {130055}, shortjournal = {Neurocomputing}, title = {NeuroSimWorm: A multisensory framework for modeling and simulating neural circuits of caenorhabditis elegans}, volume = {638}, year = {2025}, }
• (2025). Learning pose regression as reliable pixel-level matching for self-supervised depth estimation. NEUCOM, 638, 130048. (www)

  @article{ , abstract = {Existing self-supervised methods for depth-pose joint learning mainly focus on the design of sophisticated depth estimation networks, while pose estimation is mostly conducted using a vanilla black-box CNN-based PoseNet that directly regresses the relative pose from two frames. Since direct pose regression is a non-trivial task requiring abstract understanding of the scenes, self-supervised PoseNet can hardly generalize to challenging scenarios with complex camera motions, leading to inferior performance of both pose and depth estimation. To address this issue, we propose a pose estimation method that explicitly recovers relative poses from 3D–2D correspondences, in conjunction with depth estimation via a differentiable weighted Perspective-n-Point (PnP) module. Global pose estimation from input images then essentially amounts to local image matching, which mainly relies on low-level cues and is therefore more concrete with stronger generalization ability. In addition, a weighting module is learned to adaptively filter out outliers based on their appearance feature, since not all matches contribute equally in solving relative poses, and only reliable ones are needed for robust pose recovery. We implement the above pipeline using a fully differentiable network, which allows self-supervised joint depth-pose learning in an end-to-end manner. Extensive experiments show that our method achieves state-of-the-art performance on KITTI monocular depth estimation and visual odometry benchmarks, with a 2%–5% improvement in depth accuracy (Abs Rel) using the identical model architecture and 40% improvement in pose accuracy (ATE) compared with PoseNet. Furthermore, our approach demonstrates superior generalization ability on indoor scenarios, outperforming existing self-supervised counterparts on the NYUv2 dataset by 1.5% in depth accuracy (Abs Rel).}, archive = {J_NEUCOM}, author = {Junwen He and Yifan Wang and Lijun Wang and Huchuan Lu}, doi = {10.1016/j.neucom.2025.130048}, journal = {Neurocomputing}, month = {7}, pages = {130048}, shortjournal = {Neurocomputing}, title = {Learning pose regression as reliable pixel-level matching for self-supervised depth estimation}, volume = {638}, year = {2025}, }
• (2025). QuaDCNN: Quantized compression of deep CNN based on tensor-train decomposition with automatic rank determination. NEUCOM, 638, 130047. (www)

  @article{ , abstract = {Convolutional Neural Networks (CNNs) are widely used across diverse domains. However, they suffer from computational and storage costs due to their extensive parameters and complex architectures, limiting their deployment on resource-constrained devices. Therefore, it is crucial to compress CNNs to substantially decrease the model size while maintaining performance. In this study, we introduce QuaDCNN, a novel quantized low-rank compression algorithm for CNNs based on tensor-train (TT) decomposition of a permuted kernel weight tensor, incorporating an automated rank determination process. The rank of each layer is adaptively tuned in accordance with prespecified accuracy and compression rate. Subsequent to initial TT decomposition, the network undergoes additional compression via 8-bit integer quantization. We evaluated QuaDCNN on multiple CNN architectures across three datasets, achieving superior parameter reduction, model size minimization, and FLOPs decrease, with only a slight drop in classification accuracy compared to existing CNN compression methods.}, archive = {J_NEUCOM}, author = {Xiandong Sun and Enze Zhu and Guangzhi Qu and Feng Zhang}, doi = {10.1016/j.neucom.2025.130047}, journal = {Neurocomputing}, month = {7}, pages = {130047}, shortjournal = {Neurocomputing}, title = {QuaDCNN: Quantized compression of deep CNN based on tensor-train decomposition with automatic rank determination}, volume = {638}, year = {2025}, }
• (2025). Pursuit-evasion game of under-actuated ASVs based on deep reinforcement learning and model predictive path integral control. NEUCOM, 638, 130045. (www)

  @article{ , abstract = {This paper investigates a pursuit-evasion game scenario involving multiple under-actuated autonomous surface vehicles (ASVs) pursuing a single ASV in a closed and bounded environment in the presence of moving and stationary obstacles. The pursuer and evader ASVs suffer from entirely unknown models. An entirely data-driven learning and control scheme is designed to achieve pursuit-evasion game behavior by leveraging deep reinforcement learning (DRL) and model predictive path integral (MPPI) control. Specifically, a gated recurrent unit (GRU) neural network is first trained to learn the state transition function for each ASV by taking advantage of the logged input and output data. Then, by considering inter-vehicle collision avoidance as well as collision avoidance with moving and stationary obstacles, the MPPI control laws are designed to achieve the hunting of the pursuers against the evader based on the learned state transition functions. It is noted that after learning by exploiting random and new data, pursuit-evasion game is capable of being achieved in the absence of model information. Simulation results are conducted to validate the efficacy of the presented GRU neural network-based DRL method for pursuit-evasion game of under-actuated ASVs.}, archive = {J_NEUCOM}, author = {Chao Pan and Anqing Wang and Zhouhua Peng and Bing Han and Guanghao Lyu and Weidong Zhang}, doi = {10.1016/j.neucom.2025.130045}, journal = {Neurocomputing}, month = {7}, pages = {130045}, shortjournal = {Neurocomputing}, title = {Pursuit-evasion game of under-actuated ASVs based on deep reinforcement learning and model predictive path integral control}, volume = {638}, year = {2025}, }
• (2025). NLBAC: A neural ODE-based algorithm for state-wise stable and safe reinforcement learning. NEUCOM, 638, 130041. (www)

  @article{ , abstract = {Ensuring safety and stability is critical when using reinforcement learning (RL) to control safety-critical systems. However, model-free RL algorithms usually suffer from low sample efficiency, and employing widely-used methods like dual ascent to solve constrained RL problems may be challenging due to their sensitivity to hyperparameters. To address these difficulties, in this work, we first propose an augmented Lagrangian-based method to maintain safety and stability through state-wise control Lyapunov function (CLF) and pre-defined control barrier function (CBFs) constraints in non-constrained Markov decision process (non-CMDP) settings. To handle tasks without pre-defined CBFs, we extend this method by training a barrier certificate jointly with the control policy, supported by theoretical guarantees to ensure monotonically improved control performance. Moreover, we investigate the issue of infeasibility arising from the presence of multiple state-wise constraints. A practical algorithm, Neural ordinary differential equations-based Lyapunov-Barrier Actor-Critic (NLBAC), is further designed by integrating the proposed method with the Soft Actor-Critic (SAC) and leveraging neural ordinary differential equations (NODEs) for system modeling. Comparisons with baselines and ablation experiments demonstrate that our algorithm achieves superior performance in terms of safety and driving the system towards the desired state with higher sample efficiency.}, archive = {J_NEUCOM}, author = {Liqun Zhao and Keyan Miao and Hongpeng Cao and Konstantinos Gatsis and Antonis Papachristodoulou}, doi = {10.1016/j.neucom.2025.130041}, journal = {Neurocomputing}, month = {7}, pages = {130041}, shortjournal = {Neurocomputing}, title = {NLBAC: A neural ODE-based algorithm for state-wise stable and safe reinforcement learning}, volume = {638}, year = {2025}, }
• (2025). A meta-reinforcement learning method for adaptive payload transportation with variations. NEUCOM, 638, 130032. (www)

  @article{ , abstract = {The safe transport of cable-suspended payloads by a group of Unmanned Aerial Vehicles (UAVs) depends on their capacity to effectively respond to fluctuations in the dynamics caused by external variations, such as wind gusts. For group transportation with obstacles, internal variations, such as changes in formation, can also alter the space occupancy of the system related to collision detection. However, traditional adaptive learning methods are challenging to adapt to these two variations. In this paper, we present a learning-based method for collision-free dual-UAV–payload transportation in the presence of varied wind force and formation change. It consists of an adaptive trajectory tracking controller based on meta-model-based reinforcement learning with online adaptation and a novel correction policy, and a path planner that can sample collision-free goal states of the system for the controller based on the meta-collision predictor. The simulation results demonstrate that the proposed trajectory tracking controller outperforms state-of-the-art model-free, model-based, and variational inference methods in terms of payload tracking error reduction and robustness when dealing with the variations mentioned above. Specifically, the proposed controller reduces the average payload tracking error to less than 0.1 metres in most tasks without obstacles. Furthermore, by following the adapted paths generated by the path planner, the trajectory tracking controller can effectively track the payload while ensuring collision-free safety of the dual-UAV–payload system during navigation among obstacles. The success rate of the proposed method is more than 80% in all scenarios with obstacles. Our project website can be seen at https://sites.google.com/view/meta-payload-fly/ and the source code is available at https://github.com/wawachen/Meta-load-fly .}, archive = {J_NEUCOM}, author = {Jingyu Chen and Ruidong Ma and Meng Xu and Fethi Candan and Lyudmila Mihaylova and John Oyekan}, doi = {10.1016/j.neucom.2025.130032}, journal = {Neurocomputing}, month = {7}, pages = {130032}, shortjournal = {Neurocomputing}, title = {A meta-reinforcement learning method for adaptive payload transportation with variations}, volume = {638}, year = {2025}, }
• (2025). Iterative compression towards in-distribution features in domain generalization. NEUCOM, 638, 130011. (www)

  @article{ , abstract = {Domain generalization seeks to train models that can generalize to unseen out-of-distribution (OOD) test domains. A typical strategy is to learn “invariant” features that are assumed to be identically distributed across all domains, in order to extract in-distribution features for downstream tasks. However, this hypothesis often fails in real-world scenarios, leading to feature shifts that degrade performance. To address this, we first theoretically decompose the generalization error in the feature space and prove that ensuring in-distribution features can provide a tighter error bound. We then propose a novel solution, Iterative In-Distribution Compression, which models the exaction of in-distribution features as a convergent iterative process. It adaptively measures the feature space to select the most similar domain for each test sample, and through iterative contraction mappings, it progressively reduces the feature distance, aligning the features with known domains. Extensive experiments on five benchmarks demonstrate the effectiveness of our approach, highlighting the advantages of in-distribution mapping for enhancing generalization performance.}, archive = {J_NEUCOM}, author = {Yizhou Jiang and Tianren Zhang and Yihan Li and Guanyu Chen and Feng Chen}, doi = {10.1016/j.neucom.2025.130011}, journal = {Neurocomputing}, month = {7}, pages = {130011}, shortjournal = {Neurocomputing}, title = {Iterative compression towards in-distribution features in domain generalization}, volume = {638}, year = {2025}, }
• (2025). Two-phase hybrid explainable machine learning model in the early diagnosis of motor and non-motor symptoms of parkinson’s disease. NEUCOM, 638, 130004. (www)

  @article{ , abstract = {Background The critical step in Parkinson’s disease (PD) detection is to categorize symptoms that may overlap with symptoms of other disorders, as PD is a slow-progressing and chronic disease. Therefore, it is necessary to develop an automated model to recognize and detect the symptoms of PD at an early stage. Methods This two-phase hybrid machine-learning model uses 59 features based on motor and non-motor symptoms. In phase-1, various ML classifiers with the best-optimized hyper-parameters were applied first to motor symptoms and then to non-motor symptoms and evaluated using various performance metrics. Afterward, the topmost features based on both symptoms were selected from the best classifiers chosen during phase-1 classification using explainable models (used as feature selectors). Lastly, these selected topmost features are applied to the neural network (NN) model for phase-2 classification. The two-phase approach was selected for its simplicity and effectiveness, as adding more phases would complicate the analysis without offering additional insights, risking overfitting and longer computational times. Results In phase-1, the random forest and extra trees classifiers effectively identified the topmost 22 motor and 11 non-motor features using explainable models. Furthermore, the NN model outperformed in phase-2 classification, achieving 99.40 % accuracy, 99.39 % F1-score, 99.47 % recall, and 99.32 % precision. Conclusion The topmost features of motor (constancy of rest tremor, facial, and bradykinesia) and non-motor (constipation, urine, and anxious mood) significantly impact PD diagnosis. These findings align with prior research, reinforcing their relevance in PD diagnosis. Hence, hybrid models can be beneficial to doctors in diagnosing the disease.}, archive = {J_NEUCOM}, author = {Nikita Aggarwal and Barjinder Singh Saini and Savita Gupta}, doi = {10.1016/j.neucom.2025.130004}, journal = {Neurocomputing}, month = {7}, pages = {130004}, shortjournal = {Neurocomputing}, title = {Two-phase hybrid explainable machine learning model in the early diagnosis of motor and non-motor symptoms of parkinson’s disease}, volume = {638}, year = {2025}, }
• (2025). Robot learning in the era of foundation models: A survey. NEUCOM, 638, 129963. (www)

  @article{ , abstract = {The proliferation of large language models (LLMs) has fueled a shift in robot learning from automation towards general embodied artificial intelligence (AI). Adopting foundation models together with traditional learning methods for robot learning has increasingly gained interest in the research community and shown potential for real-life application. However, there is little literature that comprehensively reviews the relatively new technologies combined with robotics. The purpose of this review is to systematically assess the state-of-the-art foundation models in robot learning and to identify future potential areas. Specifically, we first summarized the technical evolution of robot learning and identified the necessary preliminary preparations for foundation models, including the simulators, datasets, and foundation model framework. In addition, we focused on the following four mainstream areas of robot learning, including manipulation, navigation, task planning, and reasoning, and demonstrated how the foundation model can be adopted in the above scenarios. Furthermore, critical issues that are neglected in the current literature, including robot hardware and software decoupling, dynamic data, generalization performance in the presence of humans, etc., were discussed. This review highlights the state-of-the-art progress of foundation models in robot learning. Future research should focus on multimodal interaction, especially dynamics data, robotics-specific foundation models, AI alignment, etc.}, archive = {J_NEUCOM}, author = {Xuan Xiao and Jiahang Liu and Zhipeng Wang and Yanmin Zhou and Yong Qi and Shuo Jiang and Bin He and Qian Cheng}, doi = {10.1016/j.neucom.2025.129963}, journal = {Neurocomputing}, month = {7}, pages = {129963}, shortjournal = {Neurocomputing}, title = {Robot learning in the era of foundation models: A survey}, volume = {638}, year = {2025}, }
• (2025). Spatiotemporal semantic structural representation learning for image sequence prediction. NEUCOM, 637, 130159. (www)

  @article{ , abstract = {Image sequence prediction is a fundamental task in computer vision that neural networks predict what happens in next frames given by a sequence of images. Despite remarkable progress in recent years, it is still challenging for predictive models to learn robust representations avoiding blurry objects, due to several aspects. (i) Incomplete semantic structures. Spatial structures including textures, positions and shapes of potential semantics are not comprehensively modeled, where existing methods directly learn from raw pixels or limited structural features within specific categories. (ii) Absent structural correlation. Correlations between images are commonly modeled by recurrent and convolutional architectures in this direction, while leaving correlations among explicit structures lack of explorations. (iii) Non-selective structural fusion . Existing fusion of structures in this task focuses on the concatenation of data or intermediate features, resulting features being equally treated and representative ones are ignored. In order to address above problems, we propose a spatiotemporal semantic structural representation learning pipeline in this study. (i) For comprehensively spatial structural modeling, this pipeline starts by SAM-based segmenting potential semantic objects, then providing spatial structures including textures, positions and shapes. (ii) To realize structural correlation modeling, temporal self-attention modules are exploited in this pipeline to extract compressed temporal features in corresponding structures. (iii) Finally for fusion of structural features, spatiotemporal semantic structural representations are accessed by integrating compressed features through cross-attention based fusion modules. Extensive experiments are conducted on both KITTI, KTH and Sthv2 datasets, showing superior performance of our model and especially better visual quality of semantic parts.}, archive = {J_NEUCOM}, author = {Yechao Xu and Zhengxing Sun and Qian Li and Yunhan Sun and Yi Wang}, doi = {10.1016/j.neucom.2025.130159}, journal = {Neurocomputing}, month = {7}, pages = {130159}, shortjournal = {Neurocomputing}, title = {Spatiotemporal semantic structural representation learning for image sequence prediction}, volume = {637}, year = {2025}, }
• (2025). Anomalous sound detection using sound image and CTF-bilateral filter. NEUCOM, 637, 130151. (www)

  @article{ , abstract = {This paper introduces an innovative filter algorithm for anomalous sound detection (ASD) inspired by the bilateral filter, which outputs a weighted average of neighboring information at each pixel, smoothing the image and removing anomalies while preserving edges. Our method aims to eliminate background noise in sound signals, retaining only abnormal and prominent points, and improving the detection process to focus on these points' presence or the emergence of new anomalies. We present the Chroma-Time-Frequency (CTF) Bilateral Filter designed for ASD in non-stationary sound signals. Unlike traditional bilateral filters that use Gaussian chroma and distance factors, our CTF-Bilateral Filter incorporates additional time and frequency factors to better suit acoustic data represented by Log-Mel Spectrograms. To enhance ASD by minimizing the similarity between adjacent pixels, we introduce a Dual Window approach for applying the three weight factors. We further refine this approach with a Flat-Dual Window to better consider the time-frequency information of sound features. The residual image, derived from subtracting the background image from the raw image, undergoes feature processing to highlight prominent points. Experimental validation demonstrates that our method effectively filters out background sound, captures prominent points, improves the detection of subtle anomalies, and exhibits strong generalization capabilities.}, archive = {J_NEUCOM}, author = {Jiyu Lu and Ming Zhang and Ta Li}, doi = {10.1016/j.neucom.2025.130151}, journal = {Neurocomputing}, month = {7}, pages = {130151}, shortjournal = {Neurocomputing}, title = {Anomalous sound detection using sound image and CTF-bilateral filter}, volume = {637}, year = {2025}, }
• (2025). Leveraging hierarchical temporal importance sampling and adaptive noise modulation to enhance resilience in multi-agent task execution systems. NEUCOM, 637, 130134. (www)

  @article{ , abstract = {Multi-agent task execution systems (MATES) constitute a foundational paradigm for addressing multifarious challenges in dynamic environments, thereby necessitating resilient recovery mechanisms amidst adversarial conditions. This study introduces the Hierarchical Importance Sampling-based Multi-scale Spatial Temporal Synergetic Enhancement Framework (HIS-MSTSEF), engineered to fortify the robustness of MATES. HIS-MSTSEF employs a hierarchical importance sampling methodology that optimally discerns pivotal nodes and edges across multiple scales, thereby substantially augmenting the delineation of complex spatiotemporal interdependencies. Furthermore, the integration of an adaptive noise modulation mechanism within the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm engenders a dynamic equilibrium between exploration and exploitation during policy optimization, thereby enabling efficacious recovery strategies across a diverse array of attack scenarios. Comprehensive experiment evaluations and robustness analysis reveal that HIS-MSTSEF consistently surpasses competing algorithms, yielding resilience enhancements ranging from 30% to 56% across multiple attack scenarios. In summation, the experimental evidence substantiates that HIS-MSTSEF markedly attenuates performance volatility in multi-agent task execution systems and exhibits a pronounced aptitude for adaptive responses to an evolving attack scenarios, thereby accentuating its potential for deployment in high-stakes, dynamic operational environments.}, archive = {J_NEUCOM}, author = {Dingrui Xue and Yuqing Lin and Siwen Wei and Zhipeng Zhang and Wanlong Qi and Jiancheng Liu and Kang Sun and Jiangying Si and Kai Xu and Jiangfeng Hu and Xin Lai and Kexin Wang}, doi = {10.1016/j.neucom.2025.130134}, journal = {Neurocomputing}, month = {7}, pages = {130134}, shortjournal = {Neurocomputing}, title = {Leveraging hierarchical temporal importance sampling and adaptive noise modulation to enhance resilience in multi-agent task execution systems}, volume = {637}, year = {2025}, }
• (2025). Scene graph generation based on lightweight entity pair object detection and relation classification ensemble. NEUCOM, 637, 130130. (www)

  @article{ , abstract = {Scene Graph Generation (SGG) aims to automatically generate a semantic graph structure, enabling a deeper understanding and reasoning of visual scenes. It is widely used in scenarios such as autonomous driving, virtual reality and smart cities. Existing SGG models generally have problems such as complex structure, large number of parameters, and difficulty in deployment. At the same time, the bias in predicate relationship prediction caused by the long-tail distribution of training data affects the overall performance of SGG. In order to improve the limitations of the above methods, first, we propose a lightweight entity pair object detection method that can directly and efficiently decode subject-object pairs that may have predicate relations from images. Next, in order to alleviate the long-tail distribution, we propose a predicate relation classifier optimization method (PRCE) based on ensemble learning. The original training set is used to generate multiple new training sets based on varied sampling rules, and distinct predicate relationship classifiers are trained with the resulting classifiers integrated for final predicate relationship prediction. On the Visual Genome (VG) and OpenImageV6 datasets, by comparing with well-known methods, experiments show that our proposed model achieves very competitive performance. The code is available at: https://github.com/xhonghu/match-prce .}, archive = {J_NEUCOM}, author = {Hong-Xiang Hu and Xu-Hua Yang and Yu-Yong Zhao}, doi = {10.1016/j.neucom.2025.130130}, journal = {Neurocomputing}, month = {7}, pages = {130130}, shortjournal = {Neurocomputing}, title = {Scene graph generation based on lightweight entity pair object detection and relation classification ensemble}, volume = {637}, year = {2025}, }
• (2025). Adaptive blank compensation for few-shot image classification. NEUCOM, 637, 130127. (www)

  @article{ , abstract = {In recent years, few-shot learning has garnered significant attention from researchers. A notable trend in this domain involves the integration of feature distribution information from approximate base classes and the calibration of distributions stemming from a limited number of samples. However, the approximate classes found by existing methods are often not optimal. Through a limited number of samples in a novel class, approaching their class center is the key to determining the approximate base classes. We assume that the center of a novel class tends to be close to the center of the blank region between multiple approximate base classes. We verified this hypothesis and propose a calibration method based on it. With this method, a limited number of samples are compensated and calibrated in the direction close to the blank center of the approximate base class distribution. Finally, the more approximate base classes and more accurate distribution information can be obtained. Extensive experiments on miniImagenet, CUB, and CIFAR-FS datasets show that the proposed method achieves competitive performance. The code is available at https://github.com/DN-KID/ABC .}, archive = {J_NEUCOM}, author = {Jianfeng Xu and Baozhe Wang and Huan Wan and Pengxiang Su and Xin Wei}, doi = {10.1016/j.neucom.2025.130127}, journal = {Neurocomputing}, month = {7}, pages = {130127}, shortjournal = {Neurocomputing}, title = {Adaptive blank compensation for few-shot image classification}, volume = {637}, year = {2025}, }
• (2025). Self-supervised progressive learning for fault diagnosis under limited labeled data and varying conditions. NEUCOM, 637, 130126. (www)

  @article{ , abstract = {Insufficient labeled data and varying working conditions are two key challenges in exploring reliable machinery fault diagnosis methods for industrial production. With the capability of learning valuable representations from unlabeled data in a pretrain-finetune manner, self-supervised learning (SSL) has tried to address these challenges and achieved notable success. However, current SSL-based approaches heavily rely on the pre-defined pretext tasks whose rationality can only be evaluated from the performance in downstream tasks, resulting in the fuzzy and unstable fault diagnosis. Motivated by the human learning process which progresses from easy to difficult tasks, a self-supervised progressive learning (SSPL) framework is proposed in this paper to enhance the rationality of pretext tasks, thus improving machinery fault diagnosis. Inheriting the auto-encoder architecture, an additional predictor is introduced to measure the difficulty of tasks, thereby providing an easy-to-difficult guide for the subsequent progressive mask generation. Benefited from the dynamic cyclic training with the strategy of task difficulty measurement and the progressive mask generation module, the reasonable pretext task endows encoder with the ability to dual understand the input monitoring data. Comparative and ablation experiments on datasets have validated the superiority of the proposed SSPL framework, achieving high-precision and stable results under limited labeled data and varying working conditions. Further, an in-depth discussion has been implemented on the operational principle and practical effect of the model, demonstrating the feasible work of the proposed progressive approach in feature extraction.}, archive = {J_NEUCOM}, author = {Qiuyu Song and Lidong Yang and Xingxing Jiang and Zhongkui Zhu}, doi = {10.1016/j.neucom.2025.130126}, journal = {Neurocomputing}, month = {7}, pages = {130126}, shortjournal = {Neurocomputing}, title = {Self-supervised progressive learning for fault diagnosis under limited labeled data and varying conditions}, volume = {637}, year = {2025}, }
• (2025). Integrating neural networks with numerical schemes for dynamical systems: A review. NEUCOM, 637, 130122. (www)

  @article{ , abstract = {As scientific discovery becomes increasingly data-driven, integrating physics-based numerical methods with advanced machine learning (ML) techniques has brought new insight in the analysis of complex physical systems. This paper explores how this integrated approach overcomes the limitations of traditional first-principle methods and brute-force ML techniques to achieve a more precise solution to complex physical problems. Specifically, we review networks that combine classical numerical schemes with neural networks applied to various physical systems. These integrated methods with residual structures effectively adhere to system symmetries and conservation laws. This integration outperforms conventional data-driven techniques in robustness and predictive capability, even with smaller datasets, owing to its improved ability to capture complex physical patterns.}, archive = {J_NEUCOM}, author = {Jinsong Tang and Yunjin Tong and Lihua Chen and Shengze Cai and Shiying Xiong}, doi = {10.1016/j.neucom.2025.130122}, journal = {Neurocomputing}, month = {7}, pages = {130122}, shortjournal = {Neurocomputing}, title = {Integrating neural networks with numerical schemes for dynamical systems: A review}, volume = {637}, year = {2025}, }
• (2025). PFCC-net: A polarized fusion cross-correlation network for efficient and high-quality embedding in few-shot classification. NEUCOM, 637, 130111. (www)

  @article{ , abstract = {Constructing an effective embedding space is critical for metric-based few-shot learning. However, existing methods often struggle to balance embedding quality with computational cost. In this paper, we propose the Polarized Fusion Cross-Correlation Network (PFCC-Net), which leverages inter-image relationships to generate high-quality embeddings. The high-quality embedding depends on two core modules: the Parallel Path Fusion (PPF) module, designed to enhance multi-level embeddings while preserving rich semantic information, and the Polarized Hybrid Cross-Attention (PHCA) module, which performs polarization filtering on spatial and channel branches to extract salient features. Experimental results demonstrate that PFCC-Net achieves competitive results on four widely used few-shot classification benchmarks: miniImageNet, tieredImageNet, CUB-200-2011, and CIFAR-FS, while significantly reducing computational costs compared to ViT-based methods.}, archive = {J_NEUCOM}, author = {Qixuan Zhang and Yun Wei}, doi = {10.1016/j.neucom.2025.130111}, journal = {Neurocomputing}, month = {7}, pages = {130111}, shortjournal = {Neurocomputing}, title = {PFCC-net: A polarized fusion cross-correlation network for efficient and high-quality embedding in few-shot classification}, volume = {637}, year = {2025}, }
• (2025). Multi-view graph anomaly detection via subgraph anomaly augmentation. NEUCOM, 637, 130109. (www)

  @article{ , abstract = {Graph Anomaly Detection (GAD) aims to identify abnormal patterns whose structure and attributes deviate from those of the majority of nodes, and it has been widely applied in various real-world scenarios, such as social network analysis and financial fraud detection. However, most existing GAD methods primarily focus on detecting anomalies in a single view of graph structures, while in real-world scenarios, attributed graphs often possess multi-view graph structures. For instance, in social networks, users may perform various types of interactive actions such as liking, bookmarking, and commenting, thereby forming multi-view attributed graphs with a variety of graph structures. Most existing methods struggle to handle multi-view attributed graphs and fail to capture the underlying interactions between various views, leading to suboptimal performance. Therefore, how to take various types of views into account for anomaly detection remains a challenging task. Moreover, the sparsity of anomalous nodes also limits the performance of existing methods. In this paper, we propose a novel Multi-view Graph Anomaly Detection via subgraph anomaly augmentation, namely MulGAD. First, we fuse subgraph sampling from different views to capture the semantic interactions among them. Second, we utilize subgraph anomaly data augmentation to enhance the model’s perception of anomalies. Finally, we train our model by contrastive learning. Thus, MulGAD has the anomaly-aware ability to recognize rare anomaly patterns. Experiments show that MulGAD has excellent anomaly detection performance on four public datasets and maintains stable anomaly-aware ability across different ratios of anomalies.}, archive = {J_NEUCOM}, author = {Fu Lin and Yue Zhang and Xuexiong Luo and Mingkang Li and Zitong Wang and Enfeng Song}, doi = {10.1016/j.neucom.2025.130109}, journal = {Neurocomputing}, month = {7}, pages = {130109}, shortjournal = {Neurocomputing}, title = {Multi-view graph anomaly detection via subgraph anomaly augmentation}, volume = {637}, year = {2025}, }
• (2025). WAIT: Feature warping for animation to illustration video translation using GANs. NEUCOM, 637, 130108. (www)

  @article{ , abstract = {In this paper, we explore a new domain for video-to-video translation. Motivated by the availability of animation movies that are adopted from illustrated books for children, we aim to stylize these videos with the style of the original illustrations. Current state-of-the-art video-to-video translation models rely on having a video sequence or a single style image to stylize an input video. We introduce a new problem for video stylizing where an unordered set of images are used. This is a challenging task for two reasons: (i) we do not have the advantage of temporal consistency as in video sequences; (ii) it is more difficult to obtain consistent styles for video frames from a set of unordered images compared to using a single image. Most of the video-to-video translation methods are built on an image-to-image translation model, and integrate additional networks such as optical flow, or temporal predictors to capture temporal relations. These additional networks make the model training and inference complicated and slow down the process. To ensure temporal coherency in video-to-video style transfer, we propose a new generator network with feature warping layers which overcomes the limitations of the previous methods. We show the effectiveness of our method on three datasets both qualitatively and quantitatively. Code and pretrained models are available at https://github.com/giddyyupp/wait .}, archive = {J_NEUCOM}, author = {Samet Hicsonmez and Nermin Samet and Fidan Samet and Oguz Bakir and Emre Akbas and Pinar Duygulu}, doi = {10.1016/j.neucom.2025.130108}, journal = {Neurocomputing}, month = {7}, pages = {130108}, shortjournal = {Neurocomputing}, title = {WAIT: Feature warping for animation to illustration video translation using GANs}, volume = {637}, year = {2025}, }
• (2025). Multimodal sentiment analysis method based on image-text quantum transformer. NEUCOM, 637, 130107. (www)

  @article{ , abstract = {Multimodal sentiment analysis aims to recognize and interpret the diverse emotional information contained in different modal data. Although multimodal sentiment analysis methods have achieved significant results, they still have certain limitations in capturing the complex features of high-dimensional data. This paper proposes a multimodal sentiment analysis method utilizing the Image and Text Quantum Transformer (ITQT-MSA), which innovatively combines quantum computing and deep learning technologies to achieve more accurate emotion recognition and analysis. Firstly, a text-oriented parametric quantum circuit is designed that exploits quantum superposition and entanglement properties and combines the transformer model to achieve deep feature extraction from text data. Secondly, another image-oriented parametric quantum circuit is constructed and combined with the Visual Transformer (VIT) to adequately extract the emotional information in the image. Finally, effective alignment and the integration of features from both text and images are achieved by the designed cross-attention fusion mechanism. The experiments are performed on classical computers by designing quantum circuits within a simulated noisy environment, and the results show that the proposed method outperforms SOTA models.}, archive = {J_NEUCOM}, author = {Huaiguang Wu and Delong Kong and Lijie Wang and Daiyi Li and Jiahui Zhang and Yucan Han}, doi = {10.1016/j.neucom.2025.130107}, journal = {Neurocomputing}, month = {7}, pages = {130107}, shortjournal = {Neurocomputing}, title = {Multimodal sentiment analysis method based on image-text quantum transformer}, volume = {637}, year = {2025}, }
• (2025). Contrastive learning-based fuzzy support vector machine. NEUCOM, 637, 130101. (www)

  @article{ , abstract = {SVM faces significant challenges when dealing with noisy data, particularly in the context of complex data distributions and high levels of noise. Existing models, while offering some improvements in robustness, struggle with adapting to various noise types, handling extreme outliers, and overly relying on classification boundaries, which affects stability and classification accuracy. To address these limitations, this paper proposes a novel Fuzzy Support Vector Machine model based on contrastive learning, designed to enhance robustness to noise and improve generalization. The proposed model incorporates neighborhood structural information through contrastive learning, which refines the evaluation of slack variables and reduces the impact of noisy and outlier samples. Additionally, a redesigned fuzzy membership fusion mechanism is introduced, enabling more accurate handling of uncertain data. The experimental results on the benchmark dataset indicate that the average rankings of the accuracy and G-mean of CL-FSVM with linear and Gaussian kernels are below 2, which is superior to several excellent FSVM variants, demonstrating the effectiveness of this model.}, archive = {J_NEUCOM}, author = {Yunlong Gao and Junwen Jiang and Jinyan Pan and Bingjie Yuan and Haifeng Zhang and Qingyuan Zhu}, doi = {10.1016/j.neucom.2025.130101}, journal = {Neurocomputing}, month = {7}, pages = {130101}, shortjournal = {Neurocomputing}, title = {Contrastive learning-based fuzzy support vector machine}, volume = {637}, year = {2025}, }
• (2025). Mask-Q attention network for flare removal. NEUCOM, 637, 130100. (www)

  @article{ , abstract = {Lens flare is a common optical phenomenon that is typically undesirable because it significantly degrades image quality, thus affecting certain visual tasks. The existing main methods using CNN-based models and transformer-based models aimed at removing lens flare, however, perform poorly in removing large-scale flares because they lack the inductive bias for spatial equivariance and the ability to capture both global and local features simultaneously. This deficiency prevents networks from quickly pinpointing the locations of flares and from fully learning the contextual content necessary for repairing contaminated areas, and they also struggle with restoring uncontaminated regions. In this paper, we introduce the Mask-Q Attention Network, a multi-scale framework designed to address the flare removal problem. Our approach focuses on extracting both global and local features, leveraging ResBlock for local feature extraction and Mask-Q Attention for capturing global features. In the Mask-Q Attention, we enhance the localization capability of the attention mechanism by integrating the flare mask with the query vector (Q) in the self-attention process. The flare mask is obtained by binarizing the initial image. This integration with the flare mask effectively resolves the issue of the lack of spatial equivariance in Transformer blocks, providing the network with prior knowledge of the flare locations. Extensive experiments demonstrate MQANet’s superiority, achieving minimum gains of 0.5% PSNR and 0.6% SSIM on synthetic datasets, alongside a 0.1% SSIM improvement on real-world data. It also performs well in terms of the LPIPS metric.}, archive = {J_NEUCOM}, author = {Zihao Li and Junming Feng and Siyao Hao and Yuze Wang and Weibang Bai}, doi = {10.1016/j.neucom.2025.130100}, journal = {Neurocomputing}, month = {7}, pages = {130100}, shortjournal = {Neurocomputing}, title = {Mask-Q attention network for flare removal}, volume = {637}, year = {2025}, }
• (2025). Multi-path feature enhancement network for real-time semantic segmentation. NEUCOM, 637, 130099. (www)

  @article{ , abstract = {Convolutional Neural Networks (CNN) have been widely used in semantic segmentation. But it mainly focuses on extracting the local hierarchical information, and the ability of the global information extraction is insufficient. Transformer is good at extracting the global information, but it is computationally intensive and unfriendly to real-time semantic segmentation. In this work, we propose a Multi-Path Feature Enhancement Network (MPFENet) for real-time semantic segmentation. Its backbone is CNN and uses the Multi-Path Feature Enhancement (MPFE) module to enhance the global feature information by Transformer. We propose a Dual Pooling Transformer (DPFormer) to extract the salient and global information while reducing the computational effort. We further propose a Parallel Aggregation Asymmetric Pyramid Pooling Module (PAAPPM) to extract different scale and orientation information. Finally, we propose a Multi-Scale Feature Fusion Module (MSFFM) to fuse the features of different paths processed by feature enhancement. Numerous experiments on five datasets demonstrate that our MPFENet is effective and achieves satisfactory results. Specifically, MPFENet achieves 80.2% mIoU at 113.8 FPS on the Cityscapes dataset and 79.7% mIoU at 151.2 FPS on the CamVid dataset.}, archive = {J_NEUCOM}, author = {Siming Jia and Yongsheng Dong and Chongchong Mao and Lintao Zheng and Yaxin Li and Kang Liu}, doi = {10.1016/j.neucom.2025.130099}, journal = {Neurocomputing}, month = {7}, pages = {130099}, shortjournal = {Neurocomputing}, title = {Multi-path feature enhancement network for real-time semantic segmentation}, volume = {637}, year = {2025}, }
• (2025). A stochastic gradient tracking algorithm with adaptive momentum for distributed optimization. NEUCOM, 637, 130095. (www)

  @article{ , abstract = {In this paper, we study distributed optimization problems where each node owns a local convex cost function calculated as the average of multiple constituent functions, and multiple nodes collaborate to minimize the finite sum of these local functions. Reviewing existing work, distributed optimization methods with adaptive momentum that consider reducing computation costs have not yet been explored. To this aim, we propose a gradient tracking stochastic distributed optimization algorithm with adaptive momentum, called GTSADAM. GTSADAM combines the distributed adaptive momentum method for faster convergence with the variance reduction mechanism to reduce computation costs. We provide a convergence analysis indicating that, under certain step size conditions, GTSADAM achieves linear convergence in the mean to the exact optimal solution when each constituent function is strongly convex and smooth. Moreover, GTSADAM maintains the acceleration efficiency of adaptive momentum while minimizing computation costs, which is confirmed by numerical simulations, and its performance is better than that of existing methods.}, archive = {J_NEUCOM}, author = {Yantao Li and Hanqing Hu and Keke Zhang and Qingguo Lü and Shaojiang Deng and Huaqing Li}, doi = {10.1016/j.neucom.2025.130095}, journal = {Neurocomputing}, month = {7}, pages = {130095}, shortjournal = {Neurocomputing}, title = {A stochastic gradient tracking algorithm with adaptive momentum for distributed optimization}, volume = {637}, year = {2025}, }
• (2025). Synchronization of coupled neural networks via saturated impulsive correction involving state-dependent delay. NEUCOM, 637, 130093. (www)

  @article{ , abstract = {The synchronization problem of coupled neural networks is concerned, where the saturated impulsive correction involving state-dependent delay is taken into consideration in this paper. Such state-dependent delay makes the value of input delay dependent on historical information of system state, which implies that it is impossible to exactly know a priori how far the system state information in the history is required. Thus, the main challenge is that in the framework of state-dependent delay, the saturated nonlinearity is handled by a kind of modified linear differential inclusion methods involving a polyhedral set with expanded dimensions. It is worth mentioning that an optimization problem is established to acquire maximized estimation of attractive domain and the robustness of coupled neural networks with parameter uncertainties will be discussed in this paper. Finally, the theoretical analysis is demonstrated by two examples.}, archive = {J_NEUCOM}, author = {Cuiping Lu and Shuchen Wu and Xiaodi Li}, doi = {10.1016/j.neucom.2025.130093}, journal = {Neurocomputing}, month = {7}, pages = {130093}, shortjournal = {Neurocomputing}, title = {Synchronization of coupled neural networks via saturated impulsive correction involving state-dependent delay}, volume = {637}, year = {2025}, }
• (2025). Multi-view feature extraction based on dual contrastive heads. NEUCOM, 637, 130092. (www)

  @article{ , abstract = {Multi-view feature extraction can effectively address the problem of high dimensionality in multi-view data. Contrastive learning (CL), which is a popular self-supervised learning method, has recently attracted considerable attention. Most CL-based methods are only constructed at the sample level, which ignores the real structural information useful for feature extraction. In this study, we construct a structural-level contrastive loss, which promotes the consistency of the potential subspace structures in any two cross views in order to explore realistic and reliable structural information. On this basis, we propose a novel multi-view feature extraction method based on dual contrastive heads, which constructs the structural-level contrastive loss and integrates it into the sample-level CL-based method. In our method, the structural-level contrastive loss can help the sample-level contrastive loss extract discriminative features more effectively. Furthermore, the relationship between structural-level loss and mutual information, as well as the relationship between structural-level loss and probabilistic intra-class and inter-class scatter, are revealed, which provides the theoretical support for the excellent performance of our method. The final experiments on six real-world datasets demonstrate the superior performance of the proposed method compared to existing methods.}, archive = {J_NEUCOM}, author = {Hongjie Zhang and Ling Jing}, doi = {10.1016/j.neucom.2025.130092}, journal = {Neurocomputing}, month = {7}, pages = {130092}, shortjournal = {Neurocomputing}, title = {Multi-view feature extraction based on dual contrastive heads}, volume = {637}, year = {2025}, }
• (2025). DSINet: Dual semantic interaction and edge refinement network for salient object detection. NEUCOM, 637, 130091. (www)

  @article{ , abstract = {Contrast semantics and spatial structure semantics are known as the two types of semantics for salient object detection. However, few works have focused on the substantial impact of their complementary properties on the salient object detection task. Additionally, the edge details of salient objects are also a significant research issue. In this paper, we thoroughly explore these two problems and propose a dual semantic interaction and edge refinement network (DSINet) for salient object detection. In the feature extraction process of the encoder, we design a multi-scale context extraction module (MCE) and an attention feature enhancement module (AFE) to extract multi-scale context features from different levels of features and perform attention fusion enhancement to obtain rich global semantic features. To explore the impact of the complementary properties of the two semantics on the salient object detection task, we propose a dual-branch semantic interaction module (DBSI) to fully interact and integrate the complementary properties of contrast semantics and spatial structure semantics. This helps us obtain better saliency priors and guide the generation of more accurate saliency maps. Additionally, we propose an edge refinement fusion module (ERF) for the edge refinement of salient features. We further fuse the two semantic features in the decoding process to generate more refined saliency maps. Extensive experimental evaluations on five benchmarks show that the proposed DSINet outperforms 25 state-of-the-art salient object detection methods. Codes can be found at https://github.com/Karry-Wu/DSINet .}, archive = {J_NEUCOM}, author = {Jiazheng Wu and Zhenxue Chen and Qingqiang Guo and Chengyun Liu and Hanxiao Zhai}, doi = {10.1016/j.neucom.2025.130091}, journal = {Neurocomputing}, month = {7}, pages = {130091}, shortjournal = {Neurocomputing}, title = {DSINet: Dual semantic interaction and edge refinement network for salient object detection}, volume = {637}, year = {2025}, }
• (2025). Artificial intelligence-powered tuberculosis detection with complementary domain attention model. NEUCOM, 637, 130089. (www)

  @article{ , abstract = {Artificial intelligence-based X-ray image detection can significantly aid early tuberculosis (TB) detection. However, the varying distribution of X-ray image data across different hospitals has resulted in a decline in the model’s performance when transitioning to a new dataset. Domain adaptation techniques can effectively mitigate the impact of this issue. However, current domain adaptation methods align the entire image features between the source and target domains without explicitly focusing on regions containing transferable classification information across domains. Forced alignment of features across the entire image may lead to negative transfer. This paper proposes a complementary domain attention model (CDAM) for TB detection where the feature map is partitioned into domain-shared (DSH) and domain-specific (DSP) features. DSP features are complementary to DSH features. DSH features are responsible for mitigating the impact of domain gaps on classification. Consequently, they focus on areas containing classification information that can be transferred across domains. In contrast, the role of the DSP feature is to maximize the domain gap, concentrating its attention on areas rich in domain information. Given that the DSH and DSP features are complementary, when the DSP feature occupies domain-informative areas, it simultaneously encourages the DSH feature to focus more accurately on areas containing transferable classification information across domains, thereby enhancing classification performance. The objective of CDAM is to fully consider the importance of different regions within the feature map and mitigate negative transfer. The proposed method underwent domain adaptation experiments on the Shenzhen, Montgomery, and TBX11K datasets, achieving average accuracy, sensitivity, and specificity scores of 73.3%, 74.0%, and 72.7%, respectively. This result surpasses existing domain adaptation methods for TB data, providing evidence for the effectiveness and robustness of the proposed approach.}, archive = {J_NEUCOM}, author = {Zeyu Ding and Razali Yaakob and Azreen Azman and Siti Nurulain Mohd Rum and Norfadhlina Zakaria and Azree Shahril Ahmad Nazri}, doi = {10.1016/j.neucom.2025.130089}, journal = {Neurocomputing}, month = {7}, pages = {130089}, shortjournal = {Neurocomputing}, title = {Artificial intelligence-powered tuberculosis detection with complementary domain attention model}, volume = {637}, year = {2025}, }
• (2025). GDHS: An efficient hybrid sampling method for multi-class imbalanced data classification. NEUCOM, 637, 130088. (www)

  @article{ , abstract = {Multi-class imbalanced data has received increasing attention from the imbalanced learning community. Numerous methods have emerged in recent years, each of them exhibits superiority in certain scenarios. However, these methods usually confronted with the challenge of the complicated correlations among multiple classes, and the intractable data overlapping between classes poses additional difficulties for modeling the interrelations between classes. To this end, this paper proposes an efficient hybrid sampling method called GDHS for multi-class imbalanced data classification. It considers both the learning difficulty and the generalization potential of the data for minority oversampling. Moreover, it also considers the overlapping problems between classes, introducing majority cleaning strategies to enhance the visibility of each class as well as learning performance. To be specific, it proposes three cleaning strategies for handling the class overlapping problem: (1) self-inner class overlapping information based majority cleaning, (2) global overlapping information based majority cleaning and (3) balanced majority cleaning. Numerical experiments over 20 data sets demonstrate the superiority of the proposed method in terms of mGM and MAUC compared with 12 state-of-the-art methods. The implement of the proposed GDHS in programming language Python is available at https://github.com/ytyancp/GDHS .}, archive = {J_NEUCOM}, author = {Yuanting Yan and Yan Lv and Shuangyue Han and Chengjin Yu and Peng Zhou}, doi = {10.1016/j.neucom.2025.130088}, journal = {Neurocomputing}, month = {7}, pages = {130088}, shortjournal = {Neurocomputing}, title = {GDHS: An efficient hybrid sampling method for multi-class imbalanced data classification}, volume = {637}, year = {2025}, }
• (2025). From classical approach to deep-learning: A review on underwater target segmentation with sonar image. NEUCOM, 637, 130087. (www)

  @article{ , abstract = {Sonar image segmentation is a crucial technology in the process of ocean exploration, which is used to separate the foreground including target region, shadow region and background from sonar images. As a key approach for underwater target detection and recognition, sonar image segmentation algorithms have received extensive attention from researchers. Compared with optical image, sonar images have some unique characteristics, such as widespread noise sources and uneven intensity distribution, so it is necessary to design special segmentation algorithms. In the past, researchers generally tended to use traditional image processing algorithms in this field. As deep-learning based image segmentation algorithms show great potential gradually, research based on this technical route in the field of sonar image analysis has increased. To be best of our knowledge, there are almost few review papers on sonar image segmentation technology currently. In this paper, we analyze the research progress in this field from the last century to the present comprehensively through cluster analysis. Moreover, we conduct a detailed classification and comparison between traditional sonar image segmentation methods and that based on deep-learning method. From above comparison results, it can be seen that although segmentation methods based on deep-learning are becoming more and more common, in the field of sonar segmentation, due to the lack of data samples and other factors, traditional analysis based methods still have a significant impact. Finally, we show the multi-modal dataset based on sonar and camera collected in the experimental pool, summarize the current challenges, and propose some helpful research directions to be studied in the future.}, archive = {J_NEUCOM}, author = {Wenyu Cai and Jifeng Zhu and Meiyan Zhang}, doi = {10.1016/j.neucom.2025.130087}, journal = {Neurocomputing}, month = {7}, pages = {130087}, shortjournal = {Neurocomputing}, title = {From classical approach to deep-learning: A review on underwater target segmentation with sonar image}, volume = {637}, year = {2025}, }
• (2025). Topology analysis of EEG-based functional brain network after stroke. NEUCOM, 637, 130086. (www)

  @article{ , abstract = {Stroke is a prevalent neurological condition that can lead to various physiological dysfunctions. The functional brain network (FBN) built using electroencephalogram (EEG) can serve as a tool to examine the functional associations between different brain regions and the capability and efficiency of the brain to process information following a stroke. However, current research on stroke primarily emphasizes the functional connections between distinct brain regions and the number of nodes and connections in the brain network, disregarding the topological properties of the brain network. To solve this problem, this paper proposes a framework for analyzing the topological properties of EEG-based FBN after stroke, utilizing directed persistent homology (DPH) approach. First, the coupling relationship between EEG signals of various brain regions is employed to construct connectivity matrices for multiple FBNs. Following this, DPH is applied to compute the topological invariants that correspond to the topological properties of the connectivity matrices. Finally, persistent entropy (PE) and persistent norm (PN) are utilized to further quantify the topological invariants. These topological properties and quantitative indicators offer significant insights into FBNs, which can aid in the examination of alterations in FBNs across varying conditions. The experiments conducted to analyze the topological properties of distinct FBNs in both stroke and healthy subjects demonstrated the substantial impact of stroke on attentional, sensorimotor, and default mode networks. Furthermore, stroke patients exhibited lower PE and PN values than healthy subjects, implying that stroke impairs both the exchange of information between brain regions and physical control, as well as disrupting spontaneous brain activity. Our study provides a preliminary exploration of the use of DPH in the topology analysis of EEG-based FBN after stroke, offering a novel approach for the quantification of FBN features.}, archive = {J_NEUCOM}, author = {Xugang Xi and Zhuyao Fan and Ting Wang and Lihua Li and Jian Yang}, doi = {10.1016/j.neucom.2025.130086}, journal = {Neurocomputing}, month = {7}, pages = {130086}, shortjournal = {Neurocomputing}, title = {Topology analysis of EEG-based functional brain network after stroke}, volume = {637}, year = {2025}, }
• (2025). SRAD: A spatially-aware reconstruction network with anomaly suppression for multi-class anomaly detection. NEUCOM, 637, 130085. (www)

  @article{ , abstract = {Reconstruction-based methods have achieved remarkable outcomes in unsupervised image anomaly detection by training separate models for different categories. However, when it comes to a practical unified model, these methods often face the “identical shortcut” problem, where both normal and abnormal samples can be recovered well, leading to failure in anomaly detection. To address this problem, we propose a S patially-aware R econstruction network with anomaly suppression for multi-class A nomaly D etection (SRAD). Firstly, we propose a Spatially-aware Channel Convolutional (SCC) neural network, which replaces general convolution with channel convolution and incorporates a Spatial Information Fusion (SIF) block during encoding and decoding. The SIF block is proposed to allow the model to capture rich feature representations while avoiding overemphasizing details. Secondly, to prevent the model from learning identical reconstruction, we propose an Anomaly Suppression Feedback Learning (ASFL) strategy. The ASFL strategy encourages the model to accurately reconstruct normal samples while inhibiting the reconstruction of abnormal samples through a feedback mechanism. Experiments on MVTec AD, VisA and BTAD datasets demonstrate the clear superiority of SRAD compared to previous state-of-the-art unified models, e.g., achieving 98.2% I-AUROC and 97.6% P-AUROC on multi-class MVTec AD dataset. Furthermore, SRAD exhibits a high frame rate of 51 FPS on a 2080 Ti GPU, making it potential for practical applications.}, archive = {J_NEUCOM}, author = {Shuyun Li and Zhi Li and Rongxiang Wang and Weidong Wang and Long Zheng and Yu Lu}, doi = {10.1016/j.neucom.2025.130085}, journal = {Neurocomputing}, month = {7}, pages = {130085}, shortjournal = {Neurocomputing}, title = {SRAD: A spatially-aware reconstruction network with anomaly suppression for multi-class anomaly detection}, volume = {637}, year = {2025}, }
• (2025). Echo state network with a non-convex penalty for nonlinear time series prediction. NEUCOM, 637, 130084. (www)

  @article{ , abstract = {Echo state networks (ESNs) with large reservoirs have been widely used in nonlinear time series prediction. However, over-large reservoirs will lead to ill-conditioned solutions when the output weights of ESNs are calculated by solving a linear regression problem. To address this issue, we propose an improved ESN with a non-convex penalty (NCP-ESN) for nonlinear time series prediction. The main idea of NCP-ESN is that an adjustable log penalty with nonconvex characteristics is introduced to the loss function for generating unbiased and sparse solutions when optimizing the output weights of the network. Meanwhile, a learning method with two-stage optimization is developed for the optimal output weights by combining the coordinate descent algorithm with the generalized inverse method. Finally, two simulation sequences and two real sequences are used to test the performance of the proposed NCP-ESN on time series prediction. Experimental results have shown the better performance of the proposed NCP-ESN compared with some regularized ESNs.}, archive = {J_NEUCOM}, author = {Wenting Wang and Fanjun Li and Qianwen Liu}, doi = {10.1016/j.neucom.2025.130084}, journal = {Neurocomputing}, month = {7}, pages = {130084}, shortjournal = {Neurocomputing}, title = {Echo state network with a non-convex penalty for nonlinear time series prediction}, volume = {637}, year = {2025}, }
• (2025). SiameseDuo++: Active learning from data streams with dual augmented siamese networks. NEUCOM, 637, 130083. (www)

  @article{ , abstract = {Data stream mining, also known as stream learning, is a growing area which deals with learning from high-speed arriving data. Its relevance has surged recently due to its wide range of applicability, such as, critical infrastructure monitoring, social media analysis, and recommender systems. The design of stream learning methods faces significant research challenges; from the nonstationary nature of the data (referred to as concept drift) and the fact that data streams are typically not annotated with the ground truth, to the requirement that such methods should process large amounts of data in real-time with limited memory. This work proposes the SiameseDuo++ method, which uses active learning to automatically select instances for a human expert to label according to a budget. Specifically, it incrementally trains two siamese neural networks which operate in synergy, augmented by generated examples. Both the proposed active learning strategy and augmentation operate in the latent space. SiameseDuo++ addresses the aforementioned challenges by operating with limited memory and limited labelling budget. Simulation experiments show that the proposed method outperforms strong baselines and state-of-the-art methods in terms of learning speed and/or performance. To promote open science we publicly release our code and datasets.}, archive = {J_NEUCOM}, author = {Kleanthis Malialis and Stylianos Filippou and Christos G. Panayiotou and Marios M. Polycarpou}, doi = {10.1016/j.neucom.2025.130083}, journal = {Neurocomputing}, month = {7}, pages = {130083}, shortjournal = {Neurocomputing}, title = {SiameseDuo++: Active learning from data streams with dual augmented siamese networks}, volume = {637}, year = {2025}, }
• (2025). SMSBPR: A symmetric multi-pairwise preferences and similarity based BPR method for recommendation with implicit feedback. NEUCOM, 637, 130082. (www)

  @article{ , abstract = {The Bayesian Personalized Ranking (BPR) method is a classical pairwise preference learning algorithm designed for implicit feedback. However, it assumes uniform user preferences for all non-interacted items, overlooking the diversity of user preferences. To address this, the Multi-pairwise Item preferences and Similarity based Bayesian Personalized Ranking (MISBPR) method (Zeng et al., 2023) further subdivides user preferences for items and proposes a multiple pairwise item preference assumption, which learns the diversity of user preferences for items. However, users and items are two entities that naturally appear in pairs in recommendation systems, solely considering preference relationships among items is insufficient. Therefore, we classify users into different preference levels for each given item, thereby capturing more granular user–item interactions. Based on this, we propose a Symmetric Multi-pairwise preference and Similarity based BPR (SMSBPR) method. SMSBPR further subdivides the preferences of users and items through an improved similarity computation method. By integrating our refined multiple pairwise preference assumption, SMSBPR delicately learns the user–item preference relationships and captures more preference diversity. Additionally, we develop an efficient learning algorithm based on stochastic gradient descent to optimize the proposed method. Empirical studies on six real-world datasets demonstrate that SMSBPR generally surpasses eleven state-of-the-art methods in terms of six evaluation metrics, highlighting its superiority in learning greater user/item preference diversity. The source code for implementing our method can be accessed publicly at: https://github.com/wjy/smsbpr .}, archive = {J_NEUCOM}, author = {Liang Zeng and Bilian Chen and Jianyi Wu and Langcai Cao}, doi = {10.1016/j.neucom.2025.130082}, journal = {Neurocomputing}, month = {7}, pages = {130082}, shortjournal = {Neurocomputing}, title = {SMSBPR: A symmetric multi-pairwise preferences and similarity based BPR method for recommendation with implicit feedback}, volume = {637}, year = {2025}, }
• (2025). PUAL: A classifier on trifurcate positive-unlabelled data. NEUCOM, 637, 130080. (www)

  @article{ , abstract = {Positive-unlabelled (PU) learning aims to train a classifier using the data containing only labelled-positive instances and unlabelled instances. However, existing PU learning methods are generally hard to achieve satisfactory performance on trifurcate data, where the positive instances distribute on both sides of the negative instances. To address this issue, firstly we propose a PU classifier with asymmetric loss (PUAL), by introducing a structure of asymmetric loss on positive instances into the objective function of the global and local learning classifier. Then we develop a kernel-based algorithm to enable PUAL to obtain non-linear decision boundary. We show that, through experiments on both simulated and real-world datasets, PUAL can achieve satisfactory classification on trifurcate data.}, archive = {J_NEUCOM}, author = {Xiaoke Wang and Xiaochen Yang and Rui Zhu and Jing-Hao Xue}, doi = {10.1016/j.neucom.2025.130080}, journal = {Neurocomputing}, month = {7}, pages = {130080}, shortjournal = {Neurocomputing}, title = {PUAL: A classifier on trifurcate positive-unlabelled data}, volume = {637}, year = {2025}, }
• (2025). Finite-time neuro-optimal control for constrained nonlinear systems through reinforcement learning. NEUCOM, 637, 130079. (www)

  @article{ , abstract = {This paper investigates the problem of reinforcement learning-based finite-time neuro-optimal control for constrained nonlinear systems with dynamic asymmetric constraints and input saturation. To begin with, based on the definition of finite time, an optimal backstepping technique is used to build the Sub-Actor-Critic framework, under which the tracking error and system variables are stays within the bounds imposed by the asymmetric time-varying constraint. Then, identifier neural networks are utilized to tackle the asymmetric saturation constraints and estimate unknown nonlinearities. Furthermore, the policy iteration algorithm is introduced to update each subsystem’s control input and value function, simultaneously solving the Hamilton–Jacobi–Bellman equation. In addition, all signals in the closed-loop system are guaranteed to be bounded within finite time via Lyapunov stability analysis. Finally, simulation results validate the proposed method’s effectiveness.}, archive = {J_NEUCOM}, author = {Jinshan Bian and Hongbing Xia and Jun Yi and Chaoxu Mu and Chenyi Si}, doi = {10.1016/j.neucom.2025.130079}, journal = {Neurocomputing}, month = {7}, pages = {130079}, shortjournal = {Neurocomputing}, title = {Finite-time neuro-optimal control for constrained nonlinear systems through reinforcement learning}, volume = {637}, year = {2025}, }
• (2025). Graph traverse reference network for sign language corpus retrieval in the wild. NEUCOM, 637, 130077. (www)

  @article{ , abstract = {Sign languages are the primary languages of the deaf community as well as hearing individuals who are unable to speak, which engage the visual-manual modality to convey meanings. In recent years, there has been an explosive growth of sign language videos available from video streaming and social media service platforms. Given the size of these corpora, sign language users often face significant challenges in effectively acquiring the information they need. Therefore, we propose a novel deep learning architecture, namely Graph Traverse Reference Network (GTRN), allowing visual signing queries to retrieve relevant sign language videos (documents) from a large corpus. GTRN introduces a traverse graph, which provides coarse-to-fine reference information in a hierarchical manner from frame-level to body-part-level observations. A reference-based attention is devised to obtain the embedding for a visual input of each level, which allows the computations to be allocated and processed at difference locations regarding local devices and central servers. A contrastive learning strategy optimizes GTRN in pursuit of a joint latent space for the queries and the documents by their meanings. Moreover, GTRN is compatible to leverage existing general visual representation foundation models, by which their resulted embeddings are used as the frame-level reference of GTRN. To the best of our knowledge, it is one of the first studies on using visual signing queries for retrieving sign language videos in a real-world setting and comprehensive experiments were conducted which demonstrated the effectiveness of our proposed method.}, archive = {J_NEUCOM}, author = {Kun Hu and Fengxiang He and Adam Schembri and Zhiyong Wang}, doi = {10.1016/j.neucom.2025.130077}, journal = {Neurocomputing}, month = {7}, pages = {130077}, shortjournal = {Neurocomputing}, title = {Graph traverse reference network for sign language corpus retrieval in the wild}, volume = {637}, year = {2025}, }
• (2025). Fusion-restoration model for industrial multimodal anomaly detection. NEUCOM, 637, 130073. (www)

  @article{ , abstract = {Industrial anomaly detection based on multimodal data is receiving increasing attention. The application of the feature mapping paradigm represents a prevailing trend. However, the existing feature mapping method is limited by the lack of multimodal fusion, which hinders the comprehensive interaction between RGB and point cloud features. In this paper, we introduce a novel feature reconstruction paradigm called F usion- R estoration M odel (FRM) to ameliorate this problem. A fusion encoder integrates the information of two domains into a fusion embedding. Then, a pair of decoupled decoders independently restore embeddings of the corresponding domains from the fusion embedding. FRM learns nominal feature reconstruction from anomaly-free training samples and detects and localizes anomalies based on the reconstruction residuals in the inference phase. A joint loss that constrains both direction and magnitude is used to enhance the robustness of the reconstruction. Additionally, a semi-frozen training strategy is designed to adapt the batch normalization parameters of the 3D feature extractor to the target industrial dataset. Extensive experiments show that our method achieves effective and efficient multimodal anomaly detection on the MVTec 3D-AD dataset.}, archive = {J_NEUCOM}, author = {Jiaxun Wang and Yanchang Niu and Biqing Huang}, doi = {10.1016/j.neucom.2025.130073}, journal = {Neurocomputing}, month = {7}, pages = {130073}, shortjournal = {Neurocomputing}, title = {Fusion-restoration model for industrial multimodal anomaly detection}, volume = {637}, year = {2025}, }
• (2025). 3DAxisPrompt: Promoting the 3D grounding and reasoning in GPT-4o. NEUCOM, 637, 130072. (www)

  @article{ , abstract = {Multimodal Large Language Models (MLLMs) exhibit impressive capabilities across a variety of tasks, especially when equipped with carefully designed visual prompts. However, existing studies primarily focus on logical reasoning and visual understanding, while the capability of MLLMs to operate effectively in 3D vision remains an ongoing area of exploration. In this paper, we introduce a novel visual prompting method, called 3DAxisPrompt, to elicit the 3D understanding capabilities of MLLMs in real-world scenes. More specifically, our method leverages the 3D coordinate axis and masks generated from the Segment Anything Model (SAM) to provide explicit geometric priors to MLLMs and then extend their impressive 2D grounding/reasoning ability to real-world 3D scenarios. Besides, we first provide a thorough investigation of the potential visual prompting formats and conclude our findings to reveal the potential and limits of 3D understanding capabilities in GPT-4o, as a representative of MLLMs. Finally, we build evaluation environments with four datasets, i.e. ScanRefer, ScanNet, FMB, and nuScene datasets, covering various 3D tasks. Based on this, we conduct extensive quantitative and qualitative experiments, which demonstrate the effectiveness of the proposed method. Overall, our study reveals that MLLMs, with the help of 3DAxisPrompt, can effectively perceive an object’s 3D position in real-world scenarios. Nevertheless, a single prompt engineering approach does not consistently achieve the best outcomes for all 3D tasks. This study highlights the feasibility of leveraging MLLMs for 3D vision grounding/reasoning with prompt engineering techniques.}, archive = {J_NEUCOM}, author = {Dingning Liu and Cheng Wang and Peng Gao and Renrui Zhang and Xinzhu Ma and Yuan Meng and Zhihui Wang}, doi = {10.1016/j.neucom.2025.130072}, journal = {Neurocomputing}, month = {7}, pages = {130072}, shortjournal = {Neurocomputing}, title = {3DAxisPrompt: Promoting the 3D grounding and reasoning in GPT-4o}, volume = {637}, year = {2025}, }
• (2025). HB-net: Holistic bursting cell cluster integrated network for occluded multi-objects recognition. NEUCOM, 637, 130071. (www)

  @article{ , abstract = {Within the realm of image recognition, a specific category of multi-label classification (MLC) challenges arises when objects within the visual field may occlude one another, demanding simultaneous identification of both occluded and occluding objects. While traditional convolutional neural networks (CNNs) address these tasks, they are often bulky and achieve only moderate accuracy. To overcome this limitation, this paper introduces HB-net, a novel integrated network framework inspired by the Holistic Bursting (HB) cell from cutting-edge neural science research. Built upon the foundation of HB cell clusters, HB-net is designed to address the intricate task of simultaneously recognizing multiple occluded objects within images. The framework incorporates various Bursting cell cluster structures along with an evidence accumulation mechanism to enhance performance. Testing on multiple datasets, including digits and letters, shows that models incorporating the HB framework achieve a significant 2.98% improvement in recognition accuracy compared to models without the HB framework (1.0298 times, p=0.0499). Although in high-noise settings, standard CNNs exhibit slightly greater robustness when compared to HB-net models, the models that combine the HB framework and EA mechanism achieve a comparable level of accuracy and resilience to ResNet50, despite having only three convolutional layers and approximately 1 / 30 of the parameters. These findings of this study offer valuable insights for improving computer vision algorithms. The essential code is provided at https://github.com/d-lab438/hb-net.git .}, archive = {J_NEUCOM}, author = {Xudong Gao and Xiaoguang Gao and Jia Rong and Xiaowei Chen and Xiang Liao and Jun Chen}, doi = {10.1016/j.neucom.2025.130071}, journal = {Neurocomputing}, month = {7}, pages = {130071}, shortjournal = {Neurocomputing}, title = {HB-net: Holistic bursting cell cluster integrated network for occluded multi-objects recognition}, volume = {637}, year = {2025}, }
• (2025). ConceptPsy: A comprehensive benchmark suite for hierarchical psychological concept understanding in LLMs. NEUCOM, 637, 130070. (www)

  @article{ , abstract = {The safe and efficient deployment of Large Language Models (LLMs) in psychology requires rigorous evaluation frameworks, which current benchmarks fail to provide. Current Massive Multitask Language Understanding (MMLU) benchmarks suffer from two critical limitations: (1) coarse-grained evaluation, lacking systematic assessments of conceptual understanding, and (2) low concept coverage rate, covering only a small fraction of essential psychological concepts. For example, in the widely-used MMLU benchmarks, CMMLU covers only 59% of the necessary chapter-level concepts in psychology, while the C-EVAL benchmark lacks a specific psychology category. Both fall short in their coverage of psychological concepts. We address these limitations through ConceptPsy, introducing three technical innovations: (1) fine-grained labeling for conceptual hierarchy evaluation, rather than only subject-level average scores, (2) a comprehensive benchmark spanning 12 core psychological subjects and 1,383 manually collected concepts, and (3) a novel concept coverage metric that identifies evaluation gaps. The evaluation results indicate that although some models achieve high average scores, they may underperform on specific concepts. For instance, Qwen2.5-72B-Instruct achieved an average performance of 88%, yet scored only around 70% on certain individual concepts. This fine-grained evaluation offers valuable insights for targeted refinement. We publish our data at: https://huggingface.co/datasets/ConceptPsy/ConceptPsy .}, archive = {J_NEUCOM}, author = {Junlei Zhang and Hongliang He and Lizhi Ma and Nirui Song and Shuyuan He and Shuai Zhang and Huachuan Qiu and Zhanchao Zhou and Anqi Li and Yong Dai and Renjun Xu and Zhenzhong Lan}, doi = {10.1016/j.neucom.2025.130070}, journal = {Neurocomputing}, month = {7}, pages = {130070}, shortjournal = {Neurocomputing}, title = {ConceptPsy: A comprehensive benchmark suite for hierarchical psychological concept understanding in LLMs}, volume = {637}, year = {2025}, }
• (2025). An adversarial contrastive learning based cross-modality zero-watermarking scheme for DIBR 3D video copyright protection. NEUCOM, 637, 130068. (www)

  @article{ , abstract = {Copyright protection of depth image-based rendering (DIBR) videos has raised significant concerns due to their increasing popularity. Zero-watermarking, emerging as a powerful tool to protect the copyright of DIBR 3D videos, mainly relies on traditional feature extraction methods, thus necessitating improvements in robustness against complex geometric attacks and its ability to strike a balance between robustness and distinguishability. This paper presents a novel zero-watermarking scheme based on cross-modality feature fusion within a contrastive learning framework. Our approach integrates complementary information from 2D frames and depth maps using a cross-modality attention feature fusion mechanism to obtain discriminative features. Moreover, our features achieve a better trade-off between robustness and distinguishability by leveraging a designed contrastive learning strategy with an adversarial distortion simulator. Experimental results demonstrate our remarkable performance by reducing the false negative rates to around 0.2% when the false positive rate is equal to 0.5%, which is superior to the state-of-the-art zero-watermarking methods.}, archive = {J_NEUCOM}, author = {Xiyao Liu and Qingyu Dang and Huiyi Wang and Xiaoheng Deng and Xunli Fan and Cundian Yang and Zhihong Chen and Hui Fang}, doi = {10.1016/j.neucom.2025.130068}, journal = {Neurocomputing}, month = {7}, pages = {130068}, shortjournal = {Neurocomputing}, title = {An adversarial contrastive learning based cross-modality zero-watermarking scheme for DIBR 3D video copyright protection}, volume = {637}, year = {2025}, }
• (2025). AMSFormer: A transformer with adaptive multi-scale partitioning and multi-level spectral filtering for time-series forecasting. NEUCOM, 637, 130067. (www)

  @article{ , abstract = {Time series forecasting is essential in numerous real-world scenarios, yet it remains a challenging task due to complex temporal dependencies and the coexistence of multi-scale, periodic, and non-periodic patterns. In this paper, we present an innovative approach called the Adaptive Multi-Scale Transformer (AMSFormer) to tackle these challenges. AMSFormer integrates both frequency-domain and time-domain information to model local and global features collaboratively. By leveraging Fourier transforms, AMSFormer adaptively partitions time series data into patches that align with the data’s intrinsic patterns, enabling dynamic and efficient processing. A convolutional attention mechanism is employed to extract fine-grained local features at multiple scales while maintaining low computational overhead. For global feature extraction, AMSFormer utilizes a hierarchical frequency-domain filter to isolate key periodic components and suppress noise, enhancing the stability and accuracy of global pattern modeling. Extensive experiments on several real-world benchmark datasets demonstrate that AMSFormer consistently outperforms state-of-the-art models, highlighting its robust generalization ability and wide applicability across various forecasting tasks.}, archive = {J_NEUCOM}, author = {Honghao Liu and Yining Diao and Ke Sun and Zhaolin Wan and Zhiyang Li}, doi = {10.1016/j.neucom.2025.130067}, journal = {Neurocomputing}, month = {7}, pages = {130067}, shortjournal = {Neurocomputing}, title = {AMSFormer: A transformer with adaptive multi-scale partitioning and multi-level spectral filtering for time-series forecasting}, volume = {637}, year = {2025}, }
• (2025). Dstsa-gcn: Advancing skeleton-based gesture recognition with semantic-aware spatio-temporal topology modeling. NEUCOM, 637, 130066. (www)

  @article{ , abstract = {Graph convolutional networks (GCNs) have emerged as a powerful tool for skeleton-based action and gesture recognition, thanks to their ability to model spatial and temporal dependencies in skeleton data. However, existing GCN-based methods face critical limitations: (1) they lack effective spatio-temporal topology modeling that captures dynamic variations in skeletal motion, and (2) they struggle to model multiscale structural relationships beyond local joint connectivity. To address these issues, we propose a novel framework called Dynamic Spatial-Temporal Semantic Awareness Graph Convolutional Network (DSTSA-GCN). DSTSA-GCN introduces three key modules: Group Channel-wise Graph Convolution (GC-GC), Group Temporal-wise Graph Convolution (GT-GC), and Multi-Scale Temporal Convolution (MS-TCN). GC-GC and GT-GC operate in parallel to independently model channel-specific and frame-specific correlations, enabling robust topology learning that accounts for temporal variations. Additionally, both modules employ a grouping strategy to adaptively capture multiscale structural relationships. Complementing this, MS-TCN enhances temporal modeling through group-wise temporal convolutions with diverse receptive fields. Extensive experiments demonstrate that DSTSA-GCN significantly improves the topology modeling capabilities of GCNs, achieving state-of-the-art performance on benchmark datasets for gesture and action recognition, including SHREC’17 Track, DHG-14/28, NTU-RGB+D, NTU-RGB+D-120 and NW-ULCA. The code will be publicly available https://hucui2022.github.io/dstsa_gcn/ .}, archive = {J_NEUCOM}, author = {Hu Cui and Renjing Huang and Ruoyu Zhang and Tessai Hayama}, doi = {10.1016/j.neucom.2025.130066}, journal = {Neurocomputing}, month = {7}, pages = {130066}, shortjournal = {Neurocomputing}, title = {Dstsa-gcn: Advancing skeleton-based gesture recognition with semantic-aware spatio-temporal topology modeling}, volume = {637}, year = {2025}, }
• (2025). GlobalLight: Exploring global influence in multi-agent deep reinforcement learning for large-scale traffic signal control. NEUCOM, 637, 130065. (www)

  @article{ , abstract = {By treating each intersection as an intelligent agent, multi-agent deep reinforcement learning (MADRL) offers a promising solution to adaptive traffic signal control (ATSC) in complex urban environments. However, existing approaches often emphasize the interactions between adjacent intersections while overlooking the global influence of distant relationships. This oversight limits their scalability to small-scale traffic networks, reducing their effectiveness in real-world urban transportation systems. In this paper, we propose GlobalLight , a novel MADRL-based traffic signal control method that addresses these challenges by exploring and exploiting global influence in traffic networks. We first propose a multidimensional feature extraction module via a multi-head graph attention network, which captures the mutual influence among locally adjacent intersections. Then we design a similarity mining module with two loss functions to analyze node embeddings in the representation space, uncovering latent relationships across distant intersections in the global traffic network. Finally, GlobalLight enables similar intersections to share policy parameters for decision-making within an effective MADRL framework. Our method simultaneously considers local dependencies between adjacent intersections and global traffic flow influence, enhancing scalability and decision efficiency for ATSC in city-level larger-scale traffic systems. Experimental evaluations on both synthetic and real-world traffic networks, encompassing up to 1000 of intersections, demonstrate that our method significantly outperforms SOTA approaches across multiple performance metrics, particularly in large-scale traffic scenarios.}, archive = {J_NEUCOM}, author = {Yilin Liu and Jintao Liang and Yifeng Zhang and Ping Gong and Guiyang Luo and Quan Yuan and Jinglin Li}, doi = {10.1016/j.neucom.2025.130065}, journal = {Neurocomputing}, month = {7}, pages = {130065}, shortjournal = {Neurocomputing}, title = {GlobalLight: Exploring global influence in multi-agent deep reinforcement learning for large-scale traffic signal control}, volume = {637}, year = {2025}, }
• (2025). DQT-CALF: Content adaptive neural network based in-loop filter in VVC using dual query transformer. NEUCOM, 637, 130064. (www)

  @article{ , abstract = {As auxiliary inputs for the neural network-based in-loop filter (NNLF) in the versatile video coding (VVC), the prediction frame, partition map and quantization parameter (QP) map are effectively used for enhancing the reconstruction frame, i.e., main input. The prediction frame and partition map play a crucial role in reducing compression artifacts, while the QP map distinguishes between inputs with different QPs. However, direct concatenation of the auxiliary inputs with the reconstruction frame may not fully leverage their advantages potentially causing conflicting effects. In this paper, we propose a content adaptive NNLF in VVC using dual query transformer (DQT), named DQT-CALF. We adopt DQT based on a dual query mechanism to effectively fuse low-frequency global features and high-frequency local features, thereby learning rich feature representations. The dual query mechanism enhances comprehensiveness and representation ability of the model, reduces the risk of information loss, and is suitable for fusion of data multiple types. DQT-CALF mainly includes three parts of feature extraction (head), feature enhancement (body), and reconstruction (tail). In feature extraction, we use a parallel network structure for the main and auxiliary inputs and assign feature maps to different weights according to the richness of the input information. For the auxiliary inputs, we do not directly concatenate them with the reconstruction frame, but generate a spatial attention map based on them to enhance the reconstruction frame. In feature enhancement, we design a multi-type feature fusion module that divides the input tensor into the low-frequency and high-frequency features from the frequency viewpoint and the local and global features from the spatial viewpoint. The two groups of features are processed separately and are mutually transformed by the high-frequency local feature generation and low-frequency global feature generation, respectively. In reconstruction, we reconstruct the frame using a 3x3 convolution layer and a pixel-shuffle layer with a long skip connection. Experimental results demonstrate that DQT-CALF achieves average BD rate gains of {8.79% (Y), 22.09% (U), 22.99% (V)} and {9.68% (Y), 22.27% (U), 22.52% (V)} over the VTM-11.0_NNVC-3.0 anchor under all intra (AI) and random access (RA) configurations, respectively.}, archive = {J_NEUCOM}, author = {Yunfeng Liu and Cheolkon Jung}, doi = {10.1016/j.neucom.2025.130064}, journal = {Neurocomputing}, month = {7}, pages = {130064}, shortjournal = {Neurocomputing}, title = {DQT-CALF: Content adaptive neural network based in-loop filter in VVC using dual query transformer}, volume = {637}, year = {2025}, }
• (2025). Fraud detection in multi-relation graph: Contrastive learning on feature and structural levels. NEUCOM, 637, 130063. (www)

  @article{ , abstract = {Fraud detection has emerged as a significant area of study, primarily due to its considerable impact on real-world applications. Despite the effectiveness of existing methods for fraud detection, they have not adequately addressed two key challenges: fraudulent camouflage and class imbalance. To tackle these challenges, we propose a novel model called Contrastive Learning on Feature and Structural Levels in Graph Neural Networks (CLFS-GNN) to effectively tackle these challenges. Our model incorporates an innovative neighbor nodes selection module that considers both feature and structural similarity between central nodes and their neighbor nodes, effectively reducing interference from fraudulent nodes by selecting highly similar neighbor nodes. Additionally, it employs an intra- and inter-graph message aggregation module with attention mechanisms to enhance the value of aggregated neighbor node information, thereby improving fraud detection performance. Furthermore, the algorithm incorporates contrastive learning to pull similar nodes closer and push dissimilar nodes further apart, mitigating class imbalance effects and achieving superior performance. Extensive experimental results show that this model outperforms the state-of-the-art GNN-based fraud detection on the Yelp and Amazon benchmark datasets.}, archive = {J_NEUCOM}, author = {Jiangnan Tang and Huanhuan Gu and Darko B. Vuković and Guandong Xu and Youquan Wang and Haicheng Tao and Jie Cao}, doi = {10.1016/j.neucom.2025.130063}, journal = {Neurocomputing}, month = {7}, pages = {130063}, shortjournal = {Neurocomputing}, title = {Fraud detection in multi-relation graph: Contrastive learning on feature and structural levels}, volume = {637}, year = {2025}, }
• (2025). AGQB-ViT: Adaptive granularity quantizer with bias for post-training quantization of vision transformers. NEUCOM, 637, 130061. (www)

  @article{ , abstract = {Despite the significant achievements of Vision Transformers (ViTs) in various computer vision tasks, these models are often large and computationally complex, making them unsuitable for resource-constrained devices. Quantizing ViTs converts model parameters from high-precision to low-precision formats, which significantly reduces computational complexity and memory requirements, thereby enhancing the performance and efficiency of ViTs in limited-resource environments. In the post-training quantization (PTQ) process of ViTs, the logarithmic quantizer is an effective method for components with power-law distribution characteristics, such as Softmax and GELU. However, when further exploring quantization scenarios at 4 bits or lower, the traditional log-based quantizer suffers from significant accuracy loss due to inflexibility in base selection and suboptimal quantization efficiency. To address this issue, we propose an adaptive granularity quantizer with bias, termed AGQB. It conducts adaptive granularity optimization for both the logarithmic base and bias according to different activation distributions and quantization bit-width requirements. Furthermore, we implemented a three-stage optimization process, setting block reconstruction as the learning objective to minimize the error before and after quantization. Extensive experimental results show that AGQB-ViT can effectively quantize various ViT models and outperforms previous methods on multiple computer vision tasks. In particular, when the ViTs model is quantized to 3 bits, we achieve an average accuracy improvement of 2.13% in image classification tasks relative to existing state-of-the-art PTQ methods. The related code is available at https://github.com/kkkyq/AGQB-ViT .}, archive = {J_NEUCOM}, author = {Ying Huo and Yongqiang Kang and Dawei Yang and Jiahao Zhu}, doi = {10.1016/j.neucom.2025.130061}, journal = {Neurocomputing}, month = {7}, pages = {130061}, shortjournal = {Neurocomputing}, title = {AGQB-ViT: Adaptive granularity quantizer with bias for post-training quantization of vision transformers}, volume = {637}, year = {2025}, }
• (2025). Robust attack-aware spread spectrum watermarking in real scenes. NEUCOM, 637, 130060. (www)

  @article{ , abstract = {Digital image watermarking hides copyright information in digital images and allows for its extraction when necessary to confirm ownership. Imperceptibility and robustness, the key indices of watermarking performance, constrain each other and are affected by the embedding location and weight. Existing techniques resolve this constraint by prioritizing imperceptibility while endeavoring to maximize robustness, which still entails a risk of missing higher visual quality or a failure in watermark extraction in real scenes. When the spread spectrum scheme is applied to embed watermarks in the discrete cosine transform domain, the embedding location and weight are not well calibrated. The embedding location is heuristically selected from among the divisions of the frequency domain, and the determination of the embedding weight relies on the predetermined imperceptibility. We address the issue between imperceptibility and robustness with the opposite strategy, prioritizing robustness while maximizing imperceptibility, and propose the attack-aware spread spectrum watermarking (ASSW) algorithm. ASSW takes prior attacks into consideration when determining the embedding location and weight with three goals: ensuring the stability, small modulus and small weight of the feature vector of the embedding location. Our experiments indicate that, with a carefully calibrated embedding location and weight, ASSW achieves greater imperceptibility and robustness than the state-of-the-art methods both on average and individual images.}, archive = {J_NEUCOM}, author = {Huijuan Guo and Baoning Niu and Ying Huang and Hu Guan and Peng Zhao}, doi = {10.1016/j.neucom.2025.130060}, journal = {Neurocomputing}, month = {7}, pages = {130060}, shortjournal = {Neurocomputing}, title = {Robust attack-aware spread spectrum watermarking in real scenes}, volume = {637}, year = {2025}, }
• (2025). One multimodal plugin enhancing all: CLIP-based pre-training framework enhancing multimodal item representations in recommendation systems. NEUCOM, 637, 130059. (www)

  @article{ , abstract = {With advances in multimodal pre-training, more efforts focus on integrating it into recommendation models. Current methods mainly focus on utilizing multimodal pre-training models to obtain multimodal representations of items and designing specific model architectures for downstream tasks. However, these methods often neglect the suitability of multimodal representations for recommendation systems since the pre-training is not conducted on recommendation datasets, making the directly obtained representations potentially suboptimal due to semantic biases from domain discrepancy and noise interference. Furthermore, collaborative information, a key element in recommendation systems, significantly impacts the effectiveness of recommendation models, but existing advanced multimodal pre-training models (e.g., CLIP) are unable to capture the collaborative information of items. To bridge the gap between multimodal pre-training models and recommendation systems, we propose a novel multimodal pre-training framework C LIP-based P re-training M ulti M odal (CPMM) item representations model for recommendation. First, the representations of images, text, and IDs are mapped to a new low-dimensional contrastive representation space for alignment and semantic enhancement, ensuring the consistency and robustness of the multimodal contrastive representation (MCR). A contrastive learning approach is designed to regulate the inter-modal distances, mitigating the impact of noise on recommendation performance. Finally, modeling of the first-order similarities of the items is conducted, thereby integrating the collaborative information of the items into the multimodal contrastive representations. Extensive experiments on Amazon benchmark datasets (Beauty, Toys, Tools) validate CPMM’s effectiveness across three core recommendation tasks: sequential recommendation, collaborative filtering, and click-through rate prediction.}, archive = {J_NEUCOM}, author = {Minghao Mo and Weihai Lu and Qixiao Xie and Zikai Xiao and Xiang Lv and Hong Yang and Yanchun Zhang}, doi = {10.1016/j.neucom.2025.130059}, journal = {Neurocomputing}, month = {7}, pages = {130059}, shortjournal = {Neurocomputing}, title = {One multimodal plugin enhancing all: CLIP-based pre-training framework enhancing multimodal item representations in recommendation systems}, volume = {637}, year = {2025}, }
• (2025). Structure modeling activation free fourier network for spacecraft image denoising. NEUCOM, 637, 130058. (www)

  @article{ , abstract = {Spacecraft image denoising is a crucial fundamental technology closely related to aerospace research. However, the existing deep learning-based image denoising methods are primarily designed for natural image and fail to adequately consider the characteristics of spacecraft image (e.g. low-light conditions, repetitive periodic structures), resulting in suboptimal performance in the spacecraft image denoising task. To address the aforementioned problems, we propose a Structure modeling Activation Free Fourier Network (SAFFN), which is an efficient spacecraft image denoising method including Structure Modeling Block (SMB) and Activation Free Fourier Block (AFFB). We present SMB to effectively extract edge information and model the structure for better identification of spacecraft components from dark regions in spacecraft noise image. We present AFFB and utilize an improved Fast Fourier block to extract repetitive periodic features and long-range information in noisy spacecraft image. Extensive experimental results demonstrate that our SAFFN performs competitively compared to the state-of-the-art methods on spacecraft noise image datasets. The codes are available at: https://github.com/shenduke/SAFFN .}, archive = {J_NEUCOM}, author = {Jingfan Yang and Hu Gao and Ying Zhang and Bowen Ma and Depeng Dang}, doi = {10.1016/j.neucom.2025.130058}, journal = {Neurocomputing}, month = {7}, pages = {130058}, shortjournal = {Neurocomputing}, title = {Structure modeling activation free fourier network for spacecraft image denoising}, volume = {637}, year = {2025}, }
• (2025). Design of an incremental extended kalman filter group for a class of multiplicatively decomposed strongly nonlinear systems considering hidden state variables. NEUCOM, 637, 130057. (www)

  @article{ , abstract = {The performance of nonlinear filters, such as the EKF, UKF, and CKF, often degrades as system nonlinearity increases. This degradation primarily stems from the finite approximations inherent in these methods — such as the linear expansion in the EKF, the unscented transformation (UT) in the UKF, and the cubature transformation in the CKF — which introduce truncation errors. To address this issue, this paper proposes an improved filtering method based on hidden variables for a class of multiplicatively decomposed strongly nonlinear systems. The proposed method treats the nonlinear functions in the multiplicatively decomposed state model as hidden variables. This approach avoids linear Taylor expansions or UT transformations of the nonlinear state transition functions, thereby significantly reducing truncation errors. Furthermore, the introduction of hidden variables effectively reduces the nonlinearity of the measurement model. For the multiplicative nonlinear functions in the measurement model, stepwise linear Taylor expansions (for the EKF) or UT transformations (for the UKF) are applied to further reduce truncation errors during the update process. By combining prediction improvements and update optimizations, this study develops a stepwise Kalman filter framework (RTSKF). Experimental results demonstrate that the proposed method significantly enhances both the accuracy and robustness of state estimation in strongly nonlinear systems.}, archive = {J_NEUCOM}, author = {Haiyang Zhang and Chenglin Wen}, doi = {10.1016/j.neucom.2025.130057}, journal = {Neurocomputing}, month = {7}, pages = {130057}, shortjournal = {Neurocomputing}, title = {Design of an incremental extended kalman filter group for a class of multiplicatively decomposed strongly nonlinear systems considering hidden state variables}, volume = {637}, year = {2025}, }
• (2025). Efficient image and video processing via symmetric inertial proximal ADMM with RPCA model. NEUCOM, 637, 130054. (www)

  @article{ , abstract = {This study proposes a novel symmetric inertial approximation alternating direction method of multiplier (SIPADMM) to address the significant challenges in solving separable two-block nonconvex nonsmooth optimization problems with linear constraints. The primary contribution of this work lies in the development of a unique algorithmic framework that successfully integrates the strengths of proximal and inertial methods while introducing a symmetric treatment of the Lagrange multiplier. A distinctive feature of our approach is the innovative utilization of numerical information from the previous three iteration steps in the proximal method, which significantly enhances the algorithm’s performance in handling nonconvex nonsmooth optimization problems. Furthermore, for the robust principal component analysis model, leveraging the properties of the hyperbolic tangent function, we have introduced a fresh nonconvex approximation function for the rank function, leading to the proposal of a new variant of the nonconvex model. Under carefully established assumptions, we obtain the convergence of the proposed algorithm through the construction of a specially designed auxiliary function H . Our theoretical analysis demonstrates that if the auxiliary function H satisfies the Kurdyka–Łojasiewicz inequality, every bounded sequence generated by the algorithm converges to a critical point of the minimization problem. In order to enhance the credibility of our approach and model in real-world scenarios, the proposed algorithm is utilized across various practical problems, especially in dealing with shadow and highlight removal in face images, and front–background separation tasks in videos. Experimental results consistently demonstrate the effectiveness of our algorithm in handling nonconvex approximation models based on rank functions.}, archive = {J_NEUCOM}, author = {Zhi-Bin Zhu and Ying Liu and Jia-Qi Huang and Yue-Hong Ding}, doi = {10.1016/j.neucom.2025.130054}, journal = {Neurocomputing}, month = {7}, pages = {130054}, shortjournal = {Neurocomputing}, title = {Efficient image and video processing via symmetric inertial proximal ADMM with RPCA model}, volume = {637}, year = {2025}, }
• (2025). SAC: Collaborative learning of structure and content features for android malware detection framework. NEUCOM, 637, 130053. (www)

  @article{ , abstract = {With the rapid development of Internet of Things (IoT) technology, Android devices have increasingly become primary targets for malware attacks. Although significant research has been conducted in the field of malware detection, existing methods still face challenges when dealing with complex samples. In particular, a more comprehensive analysis is required in the domain of feature extraction. To enhance the accuracy of malware detection, we propose the SAC framework. This method utilizes Dalvik Executable (DEX) files as the data source and achieves deep integration of multi-view features by collaboratively modeling image and graph data types. Specifically, to accurately capture the local features of malware and improve the identification of critical behavioral patterns, we designed a task-oriented convolutional neural network (CNN) named IFNeXt, which integrates visualization analysis with an inverted bottleneck structure. Furthermore, we introduced a dual-channel graph convolutional network (GCN) that models the hierarchical structure of bytecode as a directed graph, capturing the co-occurrence relationships and semantic similarities between method calls. This approach enables a deeper exploration of the global structural features of malware. The SAC framework fully leverages the complementary advantages of image and graph data structures, providing a more comprehensive characterization of malware features from both content and structural perspectives. Experimental results demonstrate that our method achieves a detection accuracy of 99.43% on multiple real-world public datasets, significantly outperforming existing state-of-the-art detection techniques. This indicates the potential and innovation of our approach in enhancing the security of the Android platform.}, archive = {J_NEUCOM}, author = {Jin Yang and Huijia Liang and Hang Ren and Dongqing Jia and Xin Wang}, doi = {10.1016/j.neucom.2025.130053}, journal = {Neurocomputing}, month = {7}, pages = {130053}, shortjournal = {Neurocomputing}, title = {SAC: Collaborative learning of structure and content features for android malware detection framework}, volume = {637}, year = {2025}, }
• (2025). MLC-DTA: Drug-target affinity prediction based on multi-level contrastive learning and equivariant graph neural networks. NEUCOM, 637, 130052. (www)

  @article{ , abstract = {With the development of computer-aided drug design in the field of pharmaceutical research, drug-target affinity (DTA) prediction is of great significance for compound screening and drug development. Recent studies have widely adopted deep learning techniques for DTA prediction, focusing on feature extraction from sequences and graph structures. Despite progress, these methods often overlook interaction relationships at the network level. Moreover, graph-based molecular representation methods relying on graph neural networks (GNNs) fail to incorporate 3D molecular structural information, limiting their potential for DTA prediction. To overcome these issues, we propose MLC-DTA, a computational method specifically designed for drug-target affinity prediction tasks. MLC-DTA uses equivariant graph neural networks to extract the structural features of drugs and targets, retaining the geometric equivariance of proteins while capturing the specific structural information of drug molecules. In addition, it integrates the interaction relationships between drugs and targets to obtain the network-level features, understanding the DTA interactions from multiple perspectives of molecules and networks. Finally, a contrastive learning strategy is introduced to maximize mutual information at both the molecular and network levels, thereby improving the prediction performance. Comparative experiments and case analyses on two datasets show that MLC-DTA has a significant improvement in accuracy.}, archive = {J_NEUCOM}, author = {Mengxin Zheng and Guicong Sun and Yongxian Fan}, doi = {10.1016/j.neucom.2025.130052}, journal = {Neurocomputing}, month = {7}, pages = {130052}, shortjournal = {Neurocomputing}, title = {MLC-DTA: Drug-target affinity prediction based on multi-level contrastive learning and equivariant graph neural networks}, volume = {637}, year = {2025}, }
• (2025). What learns next: Learning intents guided dual contrastive learning model for online course recommendation. NEUCOM, 637, 130051. (www)

  @article{ , abstract = {Course recommender systems play a crucial role in assisting learners from diverse backgrounds in selecting suitable courses from various options to achieve their educational pursuits. However, these systems face challenges in accuracy and robustness due to partial learning preference modeling and unconventional long sequences sparsity. This paper proposes a learning intents guided dual contrastive learning model for online course recommendation to address these issues. Grounded in self-determination theory, the proposed model incorporates both intrinsic motivation (long-term learning intent) and extrinsic motivation (short-term learning intent) to characterize learners’ decision-making processes. Additionally, a dual contrastive learning module, comprising self-contrastive learning and last-course guided supervised contrastive learning, is designed to mitigate the sparsity issue and enhance the robustness of sequence representation learning. Extensive experiments demonstrate the effectiveness of the model compared to several state-of-the-art methods. Furthermore, the model exhibits excellent performance for learners with varying learning sequences.}, archive = {J_NEUCOM}, author = {Weiqiang Yao and Xiaohuan Hu}, doi = {10.1016/j.neucom.2025.130051}, journal = {Neurocomputing}, month = {7}, pages = {130051}, shortjournal = {Neurocomputing}, title = {What learns next: Learning intents guided dual contrastive learning model for online course recommendation}, volume = {637}, year = {2025}, }
• (2025). DASC: Learning discriminative latent space for video clustering. NEUCOM, 637, 130050. (www)

  @article{ , abstract = {In recent years, significant advancements have been made in video analysis technologies. However, most existing methods are primarily designed for supervised learning, particularly in video classification. Accurately labeling video data is often time-consuming and labor-intensive, making large-scale annotation challenging. As a result, most of the available video data remain in an unsupervised or weakly supervised state. This situation underscores the urgent need to develop efficient methods for unsupervised video data analysis, with a particular emphasis on video clustering techniques, which can effectively alleviate the high cost and labor intensity of video data annotation by automatically grouping similar videos, thus reducing the reliance on manual labeling. This significantly enhances the efficiency and scalability of video analysis. In this paper, we propose a deep aggregation subspace clustering (DASC) network, designed to learn a video-level self-representation matrix in an end-to-end manner, without the need for any labeled data, thus operating in an unsupervised learning environment. Specifically, DASC consists of four main components: auto-encoder backbone, video modeling module (VMM), self-representation module (SrM) and feature recovered module (FRM). A frame-level latent space is first established by utilizing the auto-encoder backbone. Then, a video-level latent space is established by constructing the VMM. Next, the video-level self-representation matrix is learned in the latent space by using the SrM. Finally, the video-level latent feature will be restored to frame-level features using the FRM. Experimental results on multiple benchmark datasets demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Jiaxin Lin and Xizhan Gao and Zhihan Zhang and Haotian Deng}, doi = {10.1016/j.neucom.2025.130050}, journal = {Neurocomputing}, month = {7}, pages = {130050}, shortjournal = {Neurocomputing}, title = {DASC: Learning discriminative latent space for video clustering}, volume = {637}, year = {2025}, }
• (2025). Adaptive symbiotic graph convolutional network. NEUCOM, 637, 130049. (www)

  @article{ , abstract = {Within the intrinsic fabric of graph and network data, the latent reciprocity between network nodes forms a profound symbiotic relationship. Traditional graph neural networks struggle to adaptively parse and integrate multi-source features when learning these symbiotic relationships. To address this challenge, we propose a novel Adaptive Symbiotic Graph Convolutional Network (ASGCN) for semi-supervised node classification tasks. First, The initial contribution of this investigation is the introduction of a multi-scale feature convolution module, which enables the extraction of hierarchical features at varying scales and the construction of k-nearest neighbor graphs. This facilitates the deepening of the symbiotic features of nodes. Second, a symbiotic co-convolution module is put forth as a means of reinforcing the profound interdependence inherent to symbiotic relationships. Finally, an adaptive dynamic feature selection mechanism is introduced to flexibly respond to data characteristics, effectively identifying and fusing the most influential features in the processing information flow. Experimental results demonstrate that ASGCN exhibits significant advantages in deeply analyzing and integrating the intrinsic attributes of nodes with graph structural relationships, thereby improving performance in node classification tasks.}, archive = {J_NEUCOM}, author = {Lin Zhou and Yuzhi Xiao and Zhonglin Ye and Haixing Zhao and Zhen Liu}, doi = {10.1016/j.neucom.2025.130049}, journal = {Neurocomputing}, month = {7}, pages = {130049}, shortjournal = {Neurocomputing}, title = {Adaptive symbiotic graph convolutional network}, volume = {637}, year = {2025}, }
• (2025). MetaGen: A framework for metaheuristic development and hyperparameter optimization in machine and deep learning. NEUCOM, 637, 130046. (www)

  @article{ , abstract = {Hyperparameter optimization is a pivotal step in enhancing model performance within machine learning. Traditionally, this challenge is addressed through metaheuristics, which efficiently explore large search spaces to uncover optimal solutions. However, implementing these techniques can be complex without adequate development tools, which is the primary focus of this paper. Hence, we introduce MetaGen , a novel Python package designed to provide a comprehensive framework for developing and evaluating metaheuristic algorithms. MetaGen follows best practices in Python design, ensuring minimalistic code implementation, intuitive comprehension, and full flexibility in solution representation. The package defines two distinct user roles: Developers, responsible for algorithm implementation for hyperparameter optimization, and Solvers, who leverage pre-implemented metaheuristics to address optimization problems. Beyond algorithm implementation, MetaGen facilitates benchmarking through built-in test functions, ensuring standardized performance comparisons. It also provides automated reporting and visualization tools to analyze optimization progress and outcomes effectively. Furthermore, its modular design allows distribution and integration into existing machine learning workflows. Several illustrative use cases are presented to demonstrate its adaptability and efficacy. The package, along with code, a user manual, and supplementary materials, is available at: https://github.com/Data-Science-Big-Data-Research-Lab/MetaGen .}, archive = {J_NEUCOM}, author = {David Gutiérrez-Avilés and Manuel Jesús Jiménez-Navarro and José Francisco Torres and Francisco Martínez-Álvarez}, doi = {10.1016/j.neucom.2025.130046}, journal = {Neurocomputing}, month = {7}, pages = {130046}, shortjournal = {Neurocomputing}, title = {MetaGen: A framework for metaheuristic development and hyperparameter optimization in machine and deep learning}, volume = {637}, year = {2025}, }
• (2025). Prescribed performance event-triggered optimal control of nonlinear multi-input systems. NEUCOM, 637, 130044. (www)

  @article{ , abstract = {This article proposes a prescribed performance reinforcement learning control (PPRLC) based on event-triggered mechanism for nonlinear multi-input systems, where target error is constrained to a bounded set. Firstly, the constrained optimal control problem is reformulated as an unconstrained stationary optimal problem by using prescribed performance function. Then, the event-triggered mechanism (ETM) is integrated to save communication resources and reduce data transmission volume. In order to study the solution of the Hamilton-Jacobi-Bellman equation (HJB), we use a reinforcement learning (RL) algorithm based on the single-critic neural network (NN) and introduce a new adaptive law to update the NN weights. Based on the Lyapunov function, the convergence of weights and the closed-loop stability of the system are confirmed. Finally, the correctness and effectiveness of the method are proved by a numerical simulation example.}, archive = {J_NEUCOM}, author = {Yu Tang and Yongfeng Lv and Jun Zhao and Long Jian and Linwei Li}, doi = {10.1016/j.neucom.2025.130044}, journal = {Neurocomputing}, month = {7}, pages = {130044}, shortjournal = {Neurocomputing}, title = {Prescribed performance event-triggered optimal control of nonlinear multi-input systems}, volume = {637}, year = {2025}, }
• (2025). Trend-aware spatio-temporal fusion graph convolutional network with self-attention for traffic prediction. NEUCOM, 637, 130040. (www)

  @article{ , abstract = {Current traffic prediction methods often extract insufficient road network information, have difficulty in mining long-term temporal dependencies, and cause model performance decline due to uneven data distribution. To address the issues above, we propose a novel Spatio-Temporal Fusion Graph Convolutional Network with Trend-Aware(STFTGCN) for traffic prediction. It consists of Spatio Temporal Embedding, Spatio-Temporal Synchronous Graph Convolution, Temporal-Attention Module, and Trend-Aware Forecasting. By constructing and thus aggregating Spatial Distance Graph, Road Connection Graph, Geographic Correlation Graph, and the proposed Lagged Correlation Graph, the hidden information in the road network is fully extracted. Then, Multi-layer Spatio-Temporal Synchronous Graph Convolution captures local spatio-temporal correlations, while a sandwich structure combined Temporal Self-Attention and Temporal Trend-Aware Multi-Head Self-Attention effectively extracts long-term dependencies and responds to local traffic fluctuations. The Trend-Aware transformations method overcome uneven data distribution, improving node relationship matching and capturing dynamic traffic changes. Experiments results on real-world datasets (PEMS03, PEMS04, PEMS07, PEMS08, PEMS-BAY and METR-LA) demonstrate that the proposed STFTGCN outperforms baseline models, validating its effectiveness and practicality.}, archive = {J_NEUCOM}, author = {Xiongtao Zhang and Lijie Pan and Qing Shen and Zhenfang Liu and Jungang Lou and Yunliang Jiang}, doi = {10.1016/j.neucom.2025.130040}, journal = {Neurocomputing}, month = {7}, pages = {130040}, shortjournal = {Neurocomputing}, title = {Trend-aware spatio-temporal fusion graph convolutional network with self-attention for traffic prediction}, volume = {637}, year = {2025}, }
• (2025). Multi-type social patterns-based graph learning. NEUCOM, 637, 130039. (www)

  @article{ , abstract = {Capturing the topology exhibited in social graphs is of paramount importance to graph learning, since node representations are heavily reliant on their neighbors in graph neural networks. Conventional methods often compute node representations by aggregating features of their multi-hop neighbors, while overlooking those subgraph structures, such as those multi-type social patterns that contain unique and rich semantic information. However, modeling all the multi-type social patterns in graphs is non-trivial, due to the expensive computational cost incurred by the involvement of multiple nodes in these patterns. In addition, the multi-type social pattern usually exhibits a long-tailed distribution over graph nodes, which makes it difficult for graph neural networks to accurately calibrate the pattern information for each node. In this paper, we propose a novel multi-type social patterns-based graph learning method (MORAL), which adaptively learns the multi-type social patterns in the graph through a structural attention mechanism, and balances its distribution state to improve the accuracy of low-degree nodes. Experimental results show that MORAL significantly improves the performance compared with other baselines, particularly on large graphs.}, archive = {J_NEUCOM}, author = {Shuo Yu and Zhuoyang Han and Feng Ding and Huafei Huang and Renqiang Luo and Guoqing Han and Feng Xia}, doi = {10.1016/j.neucom.2025.130039}, journal = {Neurocomputing}, month = {7}, pages = {130039}, shortjournal = {Neurocomputing}, title = {Multi-type social patterns-based graph learning}, volume = {637}, year = {2025}, }
• (2025). All for one, and one for all: UrbanSyn dataset, the third musketeer of synthetic driving scenes. NEUCOM, 637, 130038. (www)

  @article{ , abstract = {We introduce UrbanSyn, a photorealistic dataset acquired through semi-procedurally generated synthetic urban driving scenarios. Developed using high-quality geometry and materials, UrbanSyn provides pixel-level ground truth, including depth, semantic segmentation, and instance segmentation with object bounding boxes and occlusion degree. It complements GTAV and Synscapes datasets to form what we coin as the ’Three Musketeers’. We demonstrate the value of the Three Musketeers in unsupervised domain adaptation for image semantic segmentation. Results on real-world datasets, Cityscapes, Mapillary Vistas, and BDD100K, establish new benchmarks, largely attributed to UrbanSyn. We make UrbanSyn openly and freely accessible.}, archive = {J_NEUCOM}, author = {Jose L. Gómez and Manuel Silva and Antonio Seoane and Agnés Borràs and Mario Noriega and German Ros and Jose A. Iglesias-Guitian and Antonio M. López}, doi = {10.1016/j.neucom.2025.130038}, journal = {Neurocomputing}, month = {7}, pages = {130038}, shortjournal = {Neurocomputing}, title = {All for one, and one for all: UrbanSyn dataset, the third musketeer of synthetic driving scenes}, volume = {637}, year = {2025}, }
• (2025). Tensor-train networks for learning predictive modeling of multidimensional data. NEUCOM, 637, 130037. (www)

  @article{ , abstract = {In this work, we firstly apply Tensor-Train (TT) networks to construct a compact representation of the classical Multilayer Perceptron, representing a reduction of up to 95% of the coefficients. A comparative analysis between tensor model and standard multilayer neural networks is also carried out in the context of prediction of the Mackey–Glass noisy chaotic time series and NASDAQ index. We show that the weights of a multidimensional regression model can be learned by means of TT network and the optimization of TT weights is more robust to the impact of coefficient initialization and hyper-parameter setting. Furthermore, an efficient algorithm based on alternating least squares has been proposed for approximating the weights in TT-format with a reduction of computational calculus, providing a much faster convergence than the well-known adaptive learning-method algorithms, widely applied for optimizing neural networks.}, archive = {J_NEUCOM}, author = {Michele Nazareth da Costa and Romis Attux and Andrzej Cichocki and João M.T. Romano}, doi = {10.1016/j.neucom.2025.130037}, journal = {Neurocomputing}, month = {7}, pages = {130037}, shortjournal = {Neurocomputing}, title = {Tensor-train networks for learning predictive modeling of multidimensional data}, volume = {637}, year = {2025}, }
• (2025). Competitive nonlinear layered spiking neural p system for solving classification problems. NEUCOM, 637, 130036. (www)

  @article{ , abstract = {Spiking neural P systems (SN P systems) are a class of membrane computing models that abstract the mechanism of spiking neurons. SN P system has been used in various engineering applications. The flexible structure of the SN P system allows it to be used for designing machine learning algorithms without the need for overly simplified neurons as in neural networks. In this paper, a novel SN P system called competitive nonlinear layered spiking neural P system (CNLSN P system) is proposed for solving classification problems. Experiments on benchmark datasets show that the recognition accuracy of the CNLSN P system is improved by 1.5%–2.5% compared to the layered spiking neural P system (LSN P system). Based on the CNLSN P system, a new class of deep learning model called the ConvCNLSNP model is developed. Experiments on benchmark datasets show that the ConvCNLSNP model reduces the time consumption by 85%–98% while maintaining recognition accuracy comparable to CNNs.}, archive = {J_NEUCOM}, author = {Hai Nan and Hongji Chen and Ping Guo and Chunmei Liao and S.M. Ahanaf Tahmid}, doi = {10.1016/j.neucom.2025.130036}, journal = {Neurocomputing}, month = {7}, pages = {130036}, shortjournal = {Neurocomputing}, title = {Competitive nonlinear layered spiking neural p system for solving classification problems}, volume = {637}, year = {2025}, }
• (2025). Class-view graph knowledge distillation: A new idea for learning MLPs on graphs. NEUCOM, 637, 130035. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs), while effective for processing non-Euclidean structured data, suffer from computationally intensive neighbor fetching, which hinders their deployment in low-latency applications. Cross-architecture graph knowledge distillation (KD), which trains high-performance Multi-layer Perceptrons (MLPs) to emulate teacher GNNs, offers a promising solution. However, existing GNN-MLP distillation methods rely on the sample-view KD paradigm, where student models directly mimic the teacher’s parameter space. Given the fundamentally different architectures and parameter spaces of GNNs and MLPs, this direct mimicry approach limits effective knowledge transfer. Inspired by the inherent properties of MLPs, we propose a novel class-view KD paradigm for GNN-MLP distillation. Unlike sample-view KD, our method guides student MLPs to generate more discriminative parameter configurations within their own parameter space while preserving the similarity of prediction distributions with the teacher, rather than directly imitating the teacher’s parameter configurations. Extensive experiments on public benchmark datasets demonstrate that class-view KD outperforms sample-view KD across various evaluation metrics and can be seamlessly integrated into existing GNN-MLP distillation methods to improve performance without additional computational cost. The code is available at https://github.com/xsk160/Class-View-Graph-KD .}, archive = {J_NEUCOM}, author = {Yingjie Tian and Shaokai Xu and Muyang Li}, doi = {10.1016/j.neucom.2025.130035}, journal = {Neurocomputing}, month = {7}, pages = {130035}, shortjournal = {Neurocomputing}, title = {Class-view graph knowledge distillation: A new idea for learning MLPs on graphs}, volume = {637}, year = {2025}, }
• (2025). A novel approach based on clustering and optimized ensemble deep learning for energy consumption forecasting in ethiopia. NEUCOM, 637, 130027. (www)

  @article{ , abstract = {Predicting energy consumption accurately is crucial for optimizing energy management strategies and achieving sustainability goals. Traditional methods often struggle with the complexity of energy consumption patterns, particularly in developing regions such as Ethiopia, where unique challenges exist. This study proposes an ensemble deep learning approach that integrates multiple models to enhance prediction accuracy. Additionally, as a previous step, a clustering process has been applied to discover different groups of customers. Our method combines deep learning architectures, including Gated Recurrent Units, Long Short-Term Memory, and Convolutional Neural Networks, within an optimized ensemble with weights computed with the Coronavirus Optimization Algorithm. This approach aims to leverage the strengths of each model to produce robust and reliable predictions. We demonstrate that our ensemble approach yields competitive results, outperforming individual models within the ensemble. By integrating diverse models, our framework captures nuanced patterns in energy consumption data more effectively, contributing to improved prediction accuracy. Furthermore, we validate the effectiveness of our approach using three distinct datasets from Ethiopia for three different customer clusters. These datasets represent different regions and consumption profiles within the country, ensuring the robustness and generalizability of our proposed methodology.}, archive = {J_NEUCOM}, author = {E. Tefera Habtemariam and M. Martínez-Ballesteros and A. Troncoso and F. Martínez-Álvarez}, doi = {10.1016/j.neucom.2025.130027}, journal = {Neurocomputing}, month = {7}, pages = {130027}, shortjournal = {Neurocomputing}, title = {A novel approach based on clustering and optimized ensemble deep learning for energy consumption forecasting in ethiopia}, volume = {637}, year = {2025}, }
• (2025). Counterfactual experience augmented off-policy reinforcement learning. NEUCOM, 637, 130017. (www)

  @article{ , abstract = {Reinforcement learning control algorithms face significant challenges due to out-of-distribution and inefficient exploration problems. While model-based reinforcement learning enhances the agent’s reasoning and planning capabilities by constructing virtual environments, training such virtual environments can be very complex. In order to build an efficient inference model and enhance the representativeness of learning data, we propose the Counterfactual Experience Augmentation (CEA) algorithm. CEA leverages variational autoencoders to model the dynamic patterns of state transitions and introduces randomness to model non-stationarity. This approach focuses on expanding the learning data in the experience pool through counterfactual inference and performs exceptionally well in environments that follow the bisimulation assumption. Environments with bisimulation properties are usually represented by discrete observation and action spaces, we propose a sampling method based on maximum kernel density estimation entropy to extend CEA to various environments. By providing reward signals for counterfactual state transitions based on real information, CEA constructs a complete counterfactual experience to alleviate the out-of-distribution problem of the learning data, and outperforms general SOTA algorithms in environments with difference properties. Finally, we discuss the similarities, differences and properties of generated counterfactual experiences and real experiences. The code is available at https://github.com/Aegis1863/CEA .}, archive = {J_NEUCOM}, author = {Sunbowen Lee and Yicheng Gong and Chao Deng}, doi = {10.1016/j.neucom.2025.130017}, journal = {Neurocomputing}, month = {7}, pages = {130017}, shortjournal = {Neurocomputing}, title = {Counterfactual experience augmented off-policy reinforcement learning}, volume = {637}, year = {2025}, }
• (2025). Centroid-based contrastive consistency learning for transferable deepfake detection. NEUCOM, 637, 130009. (www)

  @article{ , abstract = {Previous research efforts in deepfake detection mainly concentrated on identifying and differentiating artifacts discernible to humans. Those methods have left a bias in learned models, as they tend to concentrate on the disparities between forged and natural regions from the perspective of a single sample while overlooking consistency within categories from the perspective of the entire sample set, which remains crucial across various real-world applications. Therefore, inspired by contrastive learning, we tackle the deepfake detection problem by learning the invariant representations of both categories. Our proposed method, termed Centroid-based Contrastive Consistency Learning (C3L) method, integrates constraints on representations at both the data preprocessing and feature extraction stages. Specifically, during data preprocessing, we consider both temporal relationships within videos and the latent relationships within synthesis data. We introduce a novel Positive Enhancement Module (PEM) designed to characterize natural and forged samples in a facial semantically irrelevant way, thereby guiding a task-oriented positive pair contrasting strategy. In addition, at the feature extraction stage, we introduce the Margin Feature Simulation Module (MFSM), which leverages the centroid of the natural category to simulate marginal features for both categories. Subsequently, we employ the Supervised Contrastive Margin Loss (SCML), utilizing simulated features to emphasize differences at decision boundaries and optimize the learning process. The effectiveness and robustness of the proposed method have been demonstrated through extensive experiments.}, archive = {J_NEUCOM}, author = {Ruiqi Zha and Zhichao Lian and Qianmu Li}, doi = {10.1016/j.neucom.2025.130009}, journal = {Neurocomputing}, month = {7}, pages = {130009}, shortjournal = {Neurocomputing}, title = {Centroid-based contrastive consistency learning for transferable deepfake detection}, volume = {637}, year = {2025}, }
• (2025). CITAL: Counterfactual intervention for temporal action localization with point-level annotation. NEUCOM, 637, 130006. (www)

  @article{ , abstract = {Point-supervised temporal action localization (PTAL) requires only a timestamp annotated on each action instance for training. Most existing PTAL methods use multiple instances learning (MIL) paradigm that localize actions according to contributions of the snippets to the classification results. The gap between classification and localization tasks causes the models to focus more on clues than pure actions. And the models are prone to localize fake actions when there are biased clues between training and test datasets. Inspired by earlier efforts on causal inference, we propose a counterfactual intervention framework for PTAL, CITAL for short. Counterfactual intervention simulates how models respond to counterfactual inputs that contains the same clues without action instances. Intuitively, we can obtain the real response to the pure actions by comparing responses to the inputs before and after counterfactual intervention. Specifically, we propose a background suppression (BS) block to suppresses the background response by guiding the model pay more attention to action instances rather than clues. To fuse the output scores of the various inputs, we propose a fusing by imitation (FI) strategy that further modifies the scores to have a high response to actions and low response to the background segments, generating more accurate proposals. Besides, we propose a counterfactual example generation (CEG) block to generate counterfactual examples with only clues and background contents based on the point labels and snippet-level action scores. Our approach achieves significant mAP gains on THUMOS14, BEOID and GTEA benchmarks comparing to various CAS-based methods without introducing additional parameters.}, archive = {J_NEUCOM}, author = {Yongxiang Hu and Ziying Xia and Zichong Chen and Thupten Tsering and Jian Cheng and Tashi Nyima}, doi = {10.1016/j.neucom.2025.130006}, journal = {Neurocomputing}, month = {7}, pages = {130006}, shortjournal = {Neurocomputing}, title = {CITAL: Counterfactual intervention for temporal action localization with point-level annotation}, volume = {637}, year = {2025}, }
• (2025). Boosting accuracy of student models via masked adaptive self-distillation. NEUCOM, 637, 129988. (www)

  @article{ , abstract = {Knowledge distillation (KD) has achieved impressive success, yet conventional KD approaches are time-consuming and computationally costly. In contrast, self-distillation methods provide an efficient alternative. However, existing self-distillation methods mostly suffer from information redundancy due to the same network architecture from the teacher and student models. Additionally, they simultaneously face the inherent limitation of lacking a high-capacity teacher model. To cope with the above challenges, we propose a novel and efficient method named Masked Adaptive Self-Distillation (MASD). Specifically, we first introduce the Mask Generation Module, which masks random pixels of the feature maps and force it to reconstruct and refine more valuable features on different layers. Moreover, the Adaptive Weighting Mechanism is designed to dynamically adjust and optimize the weights of supervisory signals utilizing the probabilities from the mutual masked supervisory signals, thereby compensating the absence of high-capacity teacher model. We demonstrate the effectiveness of our MASD method on conventional image classification datasets and fine-grained datasets using state-of-the-art CNN architectures, and show that MASD significantly enhances the generalization of various backbone networks. For instance, on the CIFAR-100 classification benchmark, the proposed MASD method achieves an accuracy of 80.40% with the ResNet-18 architecture, surpassing the baseline with a 4.16% margin in Top-1 accuracy.}, archive = {J_NEUCOM}, author = {Haoran Zhao and Shuwen Tian and Jinlong Wang and Zhaopeng Deng and Xin Sun and Junyu Dong}, doi = {10.1016/j.neucom.2025.129988}, journal = {Neurocomputing}, month = {7}, pages = {129988}, shortjournal = {Neurocomputing}, title = {Boosting accuracy of student models via masked adaptive self-distillation}, volume = {637}, year = {2025}, }
• (2025). Feature calibration and feature separation for long-tailed visual recognition. NEUCOM, 637, 129983. (www)

  @article{ , abstract = {Long-tailed data classification is prevalent in real-world scenarios, but training on such datasets can lead to biased classifications and poor performance. We address this challenge by focusing on improving feature representation for tail classes, which is often lower in quality due to their closer proximity to other distinct classes. Inspired by the similarity between head and tail classes, we propose Class-wise Knowledge Distillation (CKD) to help tail classes learn prediction distributions from head classes, thus calibrating their features. Additionally, we introduce Hard Negative Samples Sampling (HNSS) to enhance feature separation by selecting challenging negative examples for contrastive learning. Our Feature Calibration and Feature Separation (FCFS) method achieves competitive results on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT benchmarks, demonstrating effective feature learning for long-tailed classification problems. This approach leverages both knowledge distillation and hard negative sampling to improve model performance.}, archive = {J_NEUCOM}, author = {Qianqian Wang and Fangyu Zhou and Xiangge Zhao and Yangtao Lin and Haibo Ye}, doi = {10.1016/j.neucom.2025.129983}, journal = {Neurocomputing}, month = {7}, pages = {129983}, shortjournal = {Neurocomputing}, title = {Feature calibration and feature separation for long-tailed visual recognition}, volume = {637}, year = {2025}, }
• (2025). DPEC: Dual-path error compensation for low-light image enhancement. NEUCOM, 637, 129980. (www)

  @article{ , abstract = {For the task of low-light image enhancement, deep learning-based algorithms have demonstrated superiority and effectiveness compared to traditional methods. However, these methods, primarily based on Retinex theory, tend to overlook the noise and color distortions in input images, leading to significant noise amplification and local color distortions in enhanced results. To address these issues, we propose the Dual-Path Error Compensation (DPEC) method, designed to improve image quality under low-light conditions by preserving local texture details while restoring global image brightness without amplifying noise. DPEC incorporates precise pixel-level error estimation to capture subtle differences and an independent denoising mechanism to prevent noise amplification. We introduce the HIS-Retinex loss to guide DPEC’s training, ensuring the brightness distribution of enhanced images closely aligns with real-world conditions. To balance computational speed and resource efficiency while training DPEC for a comprehensive understanding of the global context, we integrated the VMamba architecture into its backbone. Comprehensive quantitative and qualitative experimental results demonstrate that our algorithm significantly outperforms state-of-the-art methods in low-light image enhancement. The code is publicly available online at https://github.com/wangshuang233/DPEC .}, archive = {J_NEUCOM}, author = {Shuang Wang and Qianwen Lu and Boxing Peng and Yihe Nie and Qingchuan Tao}, doi = {10.1016/j.neucom.2025.129980}, journal = {Neurocomputing}, month = {7}, pages = {129980}, shortjournal = {Neurocomputing}, title = {DPEC: Dual-path error compensation for low-light image enhancement}, volume = {637}, year = {2025}, }
• (2025). RECoT: Relation-enhanced chains-of-thoughts for knowledge-intensive multi-hop questions answering. NEUCOM, 637, 129903. (www)

  @article{ , abstract = {Open Domain question answering is designed to enable a computer to understand and answer any question on a wide range of topics. The prevalent retrieval-reading paradigm helps large language models (LLMs) when retrieving relevant text from external knowledge sources using questions, however the multi-hop question answering approach based on Chains-of-Thoughts (CoT) may perform poorly when it comes to complex questions. This is because there can be errors in generating sentences at each hop, and these errors accumulate, leading to significant deviations in the final result. In order to solve the above problems, this paper first extracted the relational triples of complex problems. Next, triples are used to select the most representative sentence at each step during CoT generation as the query for the next-hop retrieval. The RECoT with GPT-3 results in significant improvements with F1 score up 5.1 points in downstream QA on 2WikiMultihopQA datasets and up 2.9 points on HotpotQA datasets. In addition, improvements in results can be obtained even with smaller models such as Flan-T5-large without additional training. In conclusion, RECoT reduced model hallucination and accelerated more accurate CoT reasoning to guide retrieval to get improved results. Code is publicly available at https://github.com/XD-BDIV-NLP/RECoT .}, archive = {J_NEUCOM}, author = {Ronghan Li and Dongdong Li and Haowen Yang and Xiaoxi Liu and Haoxiang Jin and RongCheng Pu and Qiguang Miao}, doi = {10.1016/j.neucom.2025.129903}, journal = {Neurocomputing}, month = {7}, pages = {129903}, shortjournal = {Neurocomputing}, title = {RECoT: Relation-enhanced chains-of-thoughts for knowledge-intensive multi-hop questions answering}, volume = {637}, year = {2025}, }
• (2025). A strategy for improving GAN generation: Contrastive self-adversarial training. NEUCOM, 637, 129864. (www)

  @article{ , abstract = {Regularization techniques are essential for addressing issues such as instability and mode collapse in GAN training. However, existing methods have shown limited success in improving GAN training and the quality of generated images. Specifically, most regularization methods emphasize model constraints or noise addition rather than optimizing GAN training strategies, which has prevented further improvement in the quality of generated images. Here, we introduce the C ontrastive S elf- A dversarial ( CSA ) training mode into traditional GAN strategies, leveraging self-competition among generated images to constrain the GAN generation process, the contrastive differences between these images enable the discriminator to capture fine-grained variations in the generation process. Notably, we observed during our experiments that the feature entropy values of generated images fluctuate significantly compared to real images. In light of this phenomenon, we propose a F eature E ntropy R egularization ( FeR ) method that stabilizes the generation process by aligning the feature entropy of generated images with that of the real dataset. Extensive experiments demonstrate the effectiveness of our approach in enhancing GAN training and image quality, advancing the state-of-the-art in regularization benchmarks. Code is available at https://github.com/JOU-UIP/CSA-GAN .}, archive = {J_NEUCOM}, author = {Hansen Zhang and Miao Yang and Haiwen Wang and Yuquan Qiu}, doi = {10.1016/j.neucom.2025.129864}, journal = {Neurocomputing}, month = {7}, pages = {129864}, shortjournal = {Neurocomputing}, title = {A strategy for improving GAN generation: Contrastive self-adversarial training}, volume = {637}, year = {2025}, }
• (2025). SurANet: Surrounding-aware network for concealed object detection via highly-efficient interactive contrastive learning strategy. NEUCOM, 637, 129863. (www)

  @article{ , abstract = {Concealed object detection (COD) in cluttered scenes is significant for various image processing applications. However, due to that concealed objects are always similar to their background, it is extremely hard to distinguish them. Here, the major obstacle is the tiny feature differences between the inside and outside object boundary region, which makes it trouble for existing COD methods to achieve accurate results. In this paper, considering that the surrounding environment information can be well utilized to identify the concealed objects, and thus, we propose a novel deep Surrounding-Aware Network, namely SurANet, for COD tasks, which introduces surrounding information into feature extraction and loss function to improve the discrimination. First, we enhance the semantics of feature maps using differential fusion of surrounding features to highlight concealed objects. Next, a Surrounding-Aware Contrastive Loss is applied to identify the concealed object via learning surrounding feature maps contrastively. Then, SurANet can be trained end-to-end with high efficiency via our proposed Spatial-Compressed Correlation Transmission strategy after our investigation of feature dynamics, and extensive experiments improve that such features can be well reserved respectively. Finally, experimental results demonstrate that the proposed SurANet outperforms state-of-the-art COD methods on multiple real datasets. Our source code will be available at https://github.com/kyh433/SurANet .}, archive = {J_NEUCOM}, author = {Yuhan Kang and Qingpeng Li and Leyuan Fang and Jian Zhao and Xuelong Li}, doi = {10.1016/j.neucom.2025.129863}, journal = {Neurocomputing}, month = {7}, pages = {129863}, shortjournal = {Neurocomputing}, title = {SurANet: Surrounding-aware network for concealed object detection via highly-efficient interactive contrastive learning strategy}, volume = {637}, year = {2025}, }
• (2025). Real-time stereo matching with enhanced geometric comprehension through cross-attention integration. NEUCOM, 636, 130069. (www)

  @article{ , abstract = {Accurate disparity estimation through stereo matching remains a critical challenge, especially for real-time applications. This work introduces a novel and computationally efficient framework that achieves high accuracy and real-time performance in stereo-based disparity estimation. The proposed approach introduces three key innovations. This work proposes a context cross-attention (CCA) module, which enhances the cost volume aggregation process by leveraging localized cross-attention for improved geometric understanding. Guided concatenation volume (GCV) is also implemented, which optimizes feature matching by effectively combining correlation clues with contextual information, reducing computational redundancy while maintaining crucial spatial details. Also, this paper proposes an uncertainty-based refinement (UR) module, which improves accuracy in challenging scenarios by utilizing an uncertainty map, a context feature map, and a geometry feature map to correct errors in challenging areas such as textureless regions and occlusions. Comprehensive experiments on multiple benchmark datasets, including KITTI, Sceneflow, Middlebury, and ETH3D, demonstrate that the proposed model performs better than existing state-of-the-art real-time approaches in accuracy metrics while maintaining comparable computational efficiency. These results establish the framework as a viable solution for demanding real-world applications, particularly in autonomous driving and robotics systems where real-time performance is crucial. The source code is available at https://github.com/kayhan-hashemi/CCAStereo .}, archive = {J_NEUCOM}, author = {Hosein Hashemi and Yasser Baleghi and Mohamad Reza Hassanzadeh}, doi = {10.1016/j.neucom.2025.130069}, journal = {Neurocomputing}, month = {7}, pages = {130069}, shortjournal = {Neurocomputing}, title = {Real-time stereo matching with enhanced geometric comprehension through cross-attention integration}, volume = {636}, year = {2025}, }
• (2025). Drop inherent biases: Multi-level attention calibration for robust cross-domain few-shot classification. NEUCOM, 636, 130056. (www)

  @article{ , abstract = {Few-shot learning (FSL) is a promising approach for addressing the challenge of classifying novel classes with only limited labeled data. Many few-shot studies have elaborated various task-shared inductive biases (meta-knowledge) to solve such tasks and have achieved impressive performance. However, when there is a domain shift between the training and testing tasks, the learned inductive biases fail to generalize across domains. In this paper, we attempt to suppress and correct inherent discriminative inductive biases from the source domain through source domain attention release and target domain attention reaggregation. We propose a few-shot learning framework, which systematically addresses the large domain shift between base and novel classes. Specifically, the framework consists of three parts: prototype-level attention calibration, feature-level attention calibration for attention release and reaggregation, and loss attention calibration. First, the prototype-level attention calibration module highlights key instances via prototype calibration, reducing the influence of noisy instances in few-shot settings. Second, the feature-level attention calibration module suppresses and corrects erroneous discriminative inductive biases from the source domain through base class attention release and novel class attention reaggregation, respectively. Finally, we incorporate the loss attention calibration module into the loss function to balance the discriminability and diversity of the classification matrix, mitigating the decline in generalization ability caused by erroneous discriminative features during domain shift. We conduct experiments on eight classic few-shot cross-domain datasets. The results demonstrate that, under varying domain shifts, our method improves performance, with average accuracy gains of 0.82% and 1.31% in the 5-way 1-shot and 5-way 5-shot settings, respectively, compared to the existing state-of-the-art (SOTA) method.}, archive = {J_NEUCOM}, author = {Minghui Li and Jing Jiang and Hongxun Yao}, doi = {10.1016/j.neucom.2025.130056}, journal = {Neurocomputing}, month = {7}, pages = {130056}, shortjournal = {Neurocomputing}, title = {Drop inherent biases: Multi-level attention calibration for robust cross-domain few-shot classification}, volume = {636}, year = {2025}, }
• (2025). HDCPAA: A few-shot class-incremental learning model for remote sensing image recognition. NEUCOM, 636, 130043. (www)

  @article{ , abstract = {In the scene of remote sensing image (RSI) recognition, it is difficult to obtain a sufficient number of samples for training all categories at once. A more realistic situation is that the recognition task occurs in an open environment, with categories gradually increasing. Additionally, due to the difficulty of collecting certain data, there are only a few samples for each new category. This leads to the problem of few-shot class-incremental learning (FSCIL), where the model learns incrementally and the number of samples for incremental classes is very small, generally only a few, while the number of samples for base classes is relatively large. To address this, this paper proposes a model framework for FSCIL of RSIs, called HDCPAA. The model is mainly divided into three parts. The first part is the feature extraction network, which is pre-trained on the base classes and then its parameters are frozen in subsequent incremental learning to alleviate catastrophic forgetting of the base classes. The second part is a fully connected layer, which transforms the prototypes of each category into quasi-orthogonal prototypes to increase the distance between the prototypes. The third part is the prototype adaptation attention module, which adaptively updates prototypes and query vectors using attention mechanisms. The training process of this module is based on the meta-learning of pseudo-incremental classes. Experiments on two popular benchmark RSI datasets, MSTAR and NWPU-RESISC45, show that our model significantly outperforms the baseline models and sets new state-of-the-art results with remarkable advantages. Our code will be uploaded at: https://github.com/lipeng144/HDCPAA .}, archive = {J_NEUCOM}, author = {Peng Li and Cunqian Feng and Xiaowei Hu and Weike Feng}, doi = {10.1016/j.neucom.2025.130043}, journal = {Neurocomputing}, month = {7}, pages = {130043}, shortjournal = {Neurocomputing}, title = {HDCPAA: A few-shot class-incremental learning model for remote sensing image recognition}, volume = {636}, year = {2025}, }
• (2025). Unifying the syntax and semantics for math word problem solving. NEUCOM, 636, 130042. (www)

  @article{ , abstract = {Math word problem solving is a complex task for natural language processing systems, requiring both comprehension of problem descriptions and deduction of accurate solutions. Existing studies have shown that graph-based approaches can achieve competitive results by applying multilayer graph neural networks to syntactic structure graphs. However, challenges such as incorrect parsing of syntactic dependency trees and insensitivity to numerical information may lead to misinterpretations in the representation. In this paper, we introduce a novel synthetic graph, the N umber- C entered S ynthetic S emantic G raph (NC-SSG), to address these challenges by reorganizing the dependency tree layout around numerical elements. We propose a double-channel graph transformer to enhance the connections between numbers and their contextual elements, thereby improving the understanding of problem descriptions. Additionally, we present a question-driven tree decoder to generate more accurate solutions, aiming to overcome shallow heuristics. Our approach mitigates the impact of parsing errors in syntactic dependency trees, yielding more precise representations and solutions. Experimental evaluations on two benchmark datasets demonstrate that our solver outperforms previous methods and achieves competitive performance compared to large language models.}, archive = {J_NEUCOM}, author = {Xingyu Tao and Yi Zhang and Zhiwen Xie and Zhuo Zhao and Guangyou Zhou and Yongchun Lu}, doi = {10.1016/j.neucom.2025.130042}, journal = {Neurocomputing}, month = {7}, pages = {130042}, shortjournal = {Neurocomputing}, title = {Unifying the syntax and semantics for math word problem solving}, volume = {636}, year = {2025}, }
• (2025). CS4TE: A novel coded self-attention and semantic synergy network for triple extraction. NEUCOM, 636, 130034. (www)

  @article{ , abstract = {The joint entity relation extraction approach holds great potential for extracting triples from unstructured text. However, in current research, two prevalent shortcomings significantly impact the efficacy of triple extraction task. Firstly, since entities constitute only a small proportion of sentences and token embedding contain a substantial amount of irrelevant information, these factors present significant challenges to the performance of classification models. Secondly, the typical process of predicting triples begins with identifying entities and then predicting triples solely based on the obtained entity representation, this process often overlooks the contextual semantic information associated with the entities. In this work, we propose CS4TE: A Novel Coded Self-Attention and Semantic Synergy Network for Triple Extraction. Specifically, we propose a novel Coded Self-Attention Mechanism designed to refine text representation by effectively masking irrelevant information and enhancing entity representation. Additionally, we propose a Semantic Synergy Network, which innovatively integrates semantic information with token pairs to predict triples, addressing the limitations of previous research that often overlooked semantic information. Finally, our model outperforms state-of-the-art baseline models on two public datasets in the joint entity-relation extraction task, and extensive experiments have been conducted to demonstrate the effectiveness of our method from multiple perspectives.}, archive = {J_NEUCOM}, author = {Huiyong Lv and Yurong Qian and Jiaying Chen and Shuxiang Hou and Hongyong Leng and Mengnan Ma}, doi = {10.1016/j.neucom.2025.130034}, journal = {Neurocomputing}, month = {7}, pages = {130034}, shortjournal = {Neurocomputing}, title = {CS4TE: A novel coded self-attention and semantic synergy network for triple extraction}, volume = {636}, year = {2025}, }
• (2025). Single-morphing attack detection using few-shot learning and triplet-loss. NEUCOM, 636, 130033. (www)

  @article{ , abstract = {Face morphing attack detection is challenging and presents a concrete and severe threat to face verification systems. A reliable detection mechanism for such attacks, tested with a robust cross-dataset protocol and unknown morphing tools, is still a research challenge. This paper proposes a framework based on the Few-Shot-Learning approach that shares image information based on the Siamese network using triplet-semi-hard-loss to tackle the morphing attack detection and boost the learning classification process. This network compares a bona fide or potentially morphed image with triplets of morphing face images. Our results show that this new network clusters the morphed images and assigns them to the right classes to obtain a lower equal error rate in a cross-dataset scenario. Few-shot learning helps to boost the learning process by sharing only small image numbers from an unknown dataset. Experimental results using cross-datasets trained with FRGCv2 and tested with FERET datasets reduced the BPCER 10 from 43% to 4.91% using ResNet50. For the AMSL open-access dataset is reduced for MobileNetV2 from BPCER 10 of 31.50% to 2.02%. For the SDD open-access synthetic dataset, the BPCER 10 is reduced for MobileNetV2 from 21.37% to 1.96%.}, archive = {J_NEUCOM}, author = {Juan E. Tapia and Daniel Schulz and Christoph Busch}, doi = {10.1016/j.neucom.2025.130033}, journal = {Neurocomputing}, month = {7}, pages = {130033}, shortjournal = {Neurocomputing}, title = {Single-morphing attack detection using few-shot learning and triplet-loss}, volume = {636}, year = {2025}, }
• (2025). Broad hashing for image retrieval. NEUCOM, 636, 130031. (www)

  @article{ , abstract = {With the popularity of deep learning, deep hashing methods have become mainstream of hashing methods which adopt deep networks to learn better feature representation of images and simultaneously generate compact binary hash codes. Deep hashing methods have a bottleneck in training efficiency due to the complex structure of deep networks. In this work, we propose a broad hashing (BH) method with high retrieval performance and very short learning time. In BH, uncorrelated and balanced binary codes are assigned to each category through a Hadamard matrix. Then, a broad hashing network is constructed to learn hash functions which maps images to binary hash codes with high efficiency. Our method yields higher retrieval precision while its training time is 200 to 700 times faster than that of deep hashing methods. At the same time, more compact hash codes are obtained compared with conventional supervised learning methods. In addition, three incremental algorithms for BH are developed for dynamic environments, which enable the hash network to be remodeled without retraining. Experiments on three benchmark datasets validate the effectiveness and efficiency of BH.}, archive = {J_NEUCOM}, author = {Wing W.Y. Ng and Xuyu Liu and Xing Tian and Ting Wang and Jianjun Zhang and C.L. Philip Chen}, doi = {10.1016/j.neucom.2025.130031}, journal = {Neurocomputing}, month = {7}, pages = {130031}, shortjournal = {Neurocomputing}, title = {Broad hashing for image retrieval}, volume = {636}, year = {2025}, }
• (2025). Adversarial erasure network based on multi-instance learning for weakly supervised video anomaly detection. NEUCOM, 636, 130030. (www)

  @article{ , abstract = {Weakly supervised video anomaly detection (WSVAD) aims to precisely locate temporal windows of abnormal events in untrimmed videos using only video-level labels. By accurately locating anomalies, WSVAD has great application potential in the security domain and contributes to the progress of smart city development. However, the lack of frame-level annotations during training makes it highly challenging to infer the status of each frame. Multiple-Instance Learning (MIL) is the dominant method in WSVAD. Due to the limitation of video-level annotations, most MIL-based methods detect obvious abnormal segments to represent the overall anomaly level of the video while overlooking weak abnormal segments. To focus on the discrimination of weak anomalies, we propose a novel WSVAD framework named Adversarial Erasure Network (AE-Net). AE-Net consists of two key components: (1) a dual-branch architecture that highlights weak anomalies by erasing the most obvious abnormal features and combining the erased features with the original ones. (2) a novel triplet loss function that improves weak anomaly representation by separating abnormal and normal features in the erased feature space. Through the above design, AE-Net can reduce false negatives in real-world anomaly detection. Extensive experiments on three WSVAD benchmarks demonstrate that our method outperforms most existing state-of-the-art methods. Specifically, AE-Net achieves an AUC of 88.40% on the UCF-Crime dataset and 98.27% on the ShanghaiTech dataset, which demonstrates that AE-Net can effectively distinguish between normal and abnormal events. Moreover, AE-Net achieves an AP of 85.13% on the XD-Violence dataset, which highlights that AE-Net can accurately detect abnormal events.}, archive = {J_NEUCOM}, author = {Xin Song and Penghui Liu and Suyuan Li and Siyang Xu and Ke Wang}, doi = {10.1016/j.neucom.2025.130030}, journal = {Neurocomputing}, month = {7}, pages = {130030}, shortjournal = {Neurocomputing}, title = {Adversarial erasure network based on multi-instance learning for weakly supervised video anomaly detection}, volume = {636}, year = {2025}, }
• (2025). Attributed network community detection based on graph contrastive learning and multi-objective evolutionary algorithm. NEUCOM, 636, 130029. (www)

  @article{ , abstract = {Attributed network community detection holds significant research value for network structure analysis and practical applications. However, existing methods still face significant challenges in addressing the conflicts between topological structure and attribute features, as well as balancing structural tightness and attribute similarity in community detection. In light of this, we propose a community detection method based on graph contrastive learning and multi-objective evolutionary algorithm (GCL-MOEA) for attributed networks. Specifically, GCL-MOEA contains two core parts: node embedding and community detection. Considering the conflict between topological structure and attribute features, the node embedding part constructs topology-augmented and attribute-augmented views, which are utilized in a cross-view graph contrastive learning model. This model comprehensively extracts node features to obtain node embedding vectors, effectively preserving the consistency and complementarity between the structure and attributes. The community detection part utilizes clustering results of node embeddings to construct high-quality initial populations. A multi-objective evolutionary algorithm is subsequently employed to obtain community structures where nodes are tightly connected and have similar attributes. The effectiveness of the proposed method is validated on five real-world networks. Experimental results demonstrate that GCL-MOEA outperforms baselines in terms of ACC, NMI, ARI, and F1, obtaining better community detection results.}, archive = {J_NEUCOM}, author = {Yao Liang and Jian Shu and Linlan Liu}, doi = {10.1016/j.neucom.2025.130029}, journal = {Neurocomputing}, month = {7}, pages = {130029}, shortjournal = {Neurocomputing}, title = {Attributed network community detection based on graph contrastive learning and multi-objective evolutionary algorithm}, volume = {636}, year = {2025}, }
• (2025). Fine-grained multi-modal prompt learning for vision–language models. NEUCOM, 636, 130028. (www)

  @article{ , abstract = {Recently advanced pre-trained vision language models have demonstrated outstanding performance in many downstream tasks via prompt learning. Prompt learning provides task-specific prompt information to exploit beneficial knowledge stored in pre-trained models to promote generalization ability for downstream tasks. However, previous work mainly focused on single modal prompt tuning (with only one prompt per modality) and salient distinguished features, which unable to flexibly adjust the two representation spaces on downstream tasks dynamically, yet makes it hard to capture subtle discriminative knowledge, which resulting in suboptimal solutions. In this work, we propose a novel F ine- G rained M ulti-modal P rompt L earning framework, denoted as FGMPL , based on the contrastive language–image pre-trained model (CLIP). To facilitate the pre-trained CLIP model to learn and represent more effective features, we design a dual-grained visual prompt scheme to learn global discrepancies as well as specify the subtle discriminative details among visual classes, and transform random vectors with class names in class-aware text prompt into class-specific discrepancy representation. Moreover, in contrast to the previous prompt approaches, we use shared latent semantic space to generate visual and text prompts to encourage cross-modal interaction. Furthermore, a multimodal prompt tuning evaluator is proposed, which can make the vision and text prompts semantically aligned and enhance each other to promote cross-modal collaborative reasoning to further improve FGMPL. Comprehensive experiments on popular image recognition benchmarks show that our approach has superior generalization and few-shot capabilities.}, archive = {J_NEUCOM}, author = {Yunfei Liu and Yunziwei Deng and Anqi Liu and Yanan Liu and Shengyang Li}, doi = {10.1016/j.neucom.2025.130028}, journal = {Neurocomputing}, month = {7}, pages = {130028}, shortjournal = {Neurocomputing}, title = {Fine-grained multi-modal prompt learning for vision–language models}, volume = {636}, year = {2025}, }
• (2025). Event denoising for dynamic vision sensor using residual graph neural network with density-based spatial clustering. NEUCOM, 636, 130026. (www)

  @article{ , abstract = {The bio-inspired emerging dynamic vision sensor (DVS), characterized by its exceptional high temporal resolution and immediate response, possesses an innate advantage in capturing rapidly changing scenes. Nevertheless, it is also susceptible to severe noise interference, especially in challenging conditions like low illumination and high exposure. Notably, the existing noise processing approaches tend to oversimplify data into 2-dimensional (2D) patterns, disregarding the sparse and irregular crucial event structure information that the DVS intrinsically provides via its asynchronous output. Aiming at these problems, we propose a residual graph neural network (RGNN) framework based on density spatial clustering for event denoising, called DBRGNN. Leveraging the temporal window rule, we extract non-overlapping event segments from the DVS event stream and adopt a density-based spatial clustering algorithm to obtain event groups with spatial correlations. To fully exploit the inherent sparsity and plentiful spatiotemporal information of the raw event stream, we transform each event group as compact graph representations via directed edges and feed them into a graph coding module composed of a series of graph convolutional and pooling layers to learn robust geometric features from event sequences. Importantly, our approach effectively reduces noise levels without compromising the spatial structure and temporal coherence of spike events. Compared with other baseline methods, our DBRGNN achieves competitive performance by quantitative and qualitative evaluations on publicly available datasets under varying lighting conditions and noise ratios.}, archive = {J_NEUCOM}, author = {Weibin Feng and Xiaoping Wang and Xin Zhan and Hongzhi Huang}, doi = {10.1016/j.neucom.2025.130026}, journal = {Neurocomputing}, month = {7}, pages = {130026}, shortjournal = {Neurocomputing}, title = {Event denoising for dynamic vision sensor using residual graph neural network with density-based spatial clustering}, volume = {636}, year = {2025}, }
• (2025). SGCDiff: Sketch-guided cross-modal diffusion model for 3D shape completion. NEUCOM, 636, 130025. (www)

  @article{ , abstract = {Shape completion aims to generate complete shapes based on partial observations. Most recent methods utilize existing information on 3D shapes for shape completion tasks, such as inputting a partial 3D shape into an encoder–decoder structure to obtain a complete 3D shape. Despite the recent rapid evolution of neural networks greatly improving the completion performance of 3D shapes, they usually generate deterministic results. However, the completed shape is inherently diverse, leading to the concept of multimodal shape completion, in which a single partial shape can correspond to multiple plausible complete shapes. Existing multimodal shape completion methods are typically unpredictable, which results in the generated complete shapes exhibiting randomness. To address the challenge of achieving a guided generation process for multimodal shape completion, we propose a novel sketch-based diffusion model. Our key designs encompass the following. We propose a novel diffusion-based framework that employs sketches as guidance to generate complete 3D shapes. Within the framework, we introduce a dual cross-modal attention module that ensures the generated results retain sufficient geometric detail. Experimental results indicate that our approach not only facilitates multimodal shape completion based on sketches but also achieves competitive performance in deterministic shape completion.}, archive = {J_NEUCOM}, author = {Zhenjiang Du and Yan Zhang and Zhitao Liu and Guan Wang and Zeyu Ma and Ning Xie and Yang Yang}, doi = {10.1016/j.neucom.2025.130025}, journal = {Neurocomputing}, month = {7}, pages = {130025}, shortjournal = {Neurocomputing}, title = {SGCDiff: Sketch-guided cross-modal diffusion model for 3D shape completion}, volume = {636}, year = {2025}, }
• (2025). Leveraging diffusion and flow matching models for demographic bias mitigation of facial attribute classifiers. NEUCOM, 636, 130024. (www)

  @article{ , abstract = {Published research highlights the presence of demographic bias in automated facial attribute classification algorithms, notably impacting women and individuals with darker skin tones. Proposed bias mitigation techniques are not generalizable, need demographic annotations, are application-specific, and often obtain fairness by reducing overall classification accuracy. In response to these challenges, this paper proposes a novel bias mitigation technique that systematically integrates diffusion and flow-matching models with a base classifier with minimal additional computational overhead. These generative models are chosen for their extreme success in capturing diverse data distributions and their inherent stochasticity. Our proposed approach augments the base classifier’s accuracy across all demographic sub-groups with enhanced fairness. Further, the stochastic nature of these generative models is harnessed to quantify prediction uncertainty, allowing for test-time rejection, which further enhances fairness. Additionally, novel solvers are proposed to significantly reduce the computational overhead of generative model inference. An exhaustive evaluation carried out on facial attribute annotated datasets substantiates the efficacy of our approach in enhancing the accuracy and fairness of facial attribute classifiers by 0 . 5 % − 3 % and 0 . 5 % − 5 % across datasets over SOTA mitigation techniques. Thus, obtaining state-of-the-art performance. Further, our proposal does not need a demographically annotated training set and is generalizable to any downstream classification task.}, archive = {J_NEUCOM}, author = {Sreeraj Ramachandran and Ajita Rattani}, doi = {10.1016/j.neucom.2025.130024}, journal = {Neurocomputing}, month = {7}, pages = {130024}, shortjournal = {Neurocomputing}, title = {Leveraging diffusion and flow matching models for demographic bias mitigation of facial attribute classifiers}, volume = {636}, year = {2025}, }
• (2025). Multi-level semantic-assisted prototype learning for few-shot action recognition. NEUCOM, 636, 130022. (www)

  @article{ , abstract = {The Few-Shot Action Recognition (FSAR) task involves recognizing new categories with limited labeled data. The conventional fine-tuning-based adaptation approach is often prone to overfitting and lacks temporal modeling for video data. Moreover, the discrepancy in distribution between meta-training and meta-test sets can also lead to suboptimal performance in few-shot scenarios. This paper introduces a simple yet effective multi-level semantic-assisted prototype learning framework to tackle these challenges. Initially, we leverage CLIP to achieve multimodal adaptation learning and present a multi-level semantic-assisted learning module to enhance the prototypes of different action classes based on semantic information. Additionally, we integrate the lightweight adapters into the CLIP visual encoder to support parameter-efficient transfer learning and improve temporal modeling in videos. Especially, a bias compensation block is employed for feature rectification to mitigate the distribution bias in FSAR stemming from data scarcity. Extensive experiments conducted on five standard benchmark datasets demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Dan Liu and Qing Xia and Fanrong Meng and Mao Ye and Jianwei Zhang}, doi = {10.1016/j.neucom.2025.130022}, journal = {Neurocomputing}, month = {7}, pages = {130022}, shortjournal = {Neurocomputing}, title = {Multi-level semantic-assisted prototype learning for few-shot action recognition}, volume = {636}, year = {2025}, }
• (2025). Fault-tolerant and attack-tolerant cooperative event-triggered sampled-data security control for synchronization of RDNNs with stochastic actuator failures and random deception attacks. NEUCOM, 636, 130021. (www)

  @article{ , abstract = {In this article, the fault-tolerant and attack-tolerant cooperative event-triggered sampled-data security (FACETSDS) synchronization problem of space-varying reaction–diffusion neural networks (SVRDNNs) under spatially point measurements (SPMs) with stochastic actuator failures and random deception attacks is investigated. First, to save more communication resources and adapt to the variation of system dynamics subject to stochastic actuator failures and random deception attacks, a FACETSDS control scheme is proposed under SPMs. Second, by constructing a Lyapunov functional and utilizing inequality techniques, some synchronization criteria based on spatial linear matrix inequalities (SLMIs) are derived for SVRDNNs. Then, to solve SLMIs, the FETSDS control for synchronization problem of SVRDNNs under SPMs with stochastic actuator failures and random deception attacks is formulated as an linear matrix inequality feasibility problem. Lastly, the designed FACETSDS synchronization strategy is verified by one numerical example.}, archive = {J_NEUCOM}, author = {Feng-Liang Zhao and Zi-Peng Wang and Junfei Qiao and Huai-Ning Wu and Tingwen Huang}, doi = {10.1016/j.neucom.2025.130021}, journal = {Neurocomputing}, month = {7}, pages = {130021}, shortjournal = {Neurocomputing}, title = {Fault-tolerant and attack-tolerant cooperative event-triggered sampled-data security control for synchronization of RDNNs with stochastic actuator failures and random deception attacks}, volume = {636}, year = {2025}, }
• (2025). Dynamic facial expression recognition in the wild via multi-snippet spatiotemporal learning. NEUCOM, 636, 130020. (www)

  @article{ , abstract = {Dynamic Facial Expression Recognition (DFER) in-the-wild poses a significant challenge in emotion recognition research. Many studies have focused on extracting finer facial features while overlooking the effect of noisy frames on the entire sequence. In addition, the imbalance between short- and long-term temporal relationships remains inadequately addressed. To tackle these issues, we propose the Multi-Snippet Spatiotemporal Learning (MSSL) framework that uses distinct temporal and spatial modeling for snippet feature extraction, enabling more accurate simulation of subtle facial expression changes while capturing finer details. We also introduced a dual-branch hierarchical module, BiTemporal Multi-Snippet Enhancement (BTMSE), which is designed to capture spatiotemporal dependencies and model subtle visual changes across snippets effectively. The Temporal-Transformer further enhances the learning of long-term dependencies, whereas learnable temporal position embeddings ensure consistency between snippet and fused features over time. By leveraging (2+1)D multi-snippet spatiotemporal modeling, BTMSE, and the Temporal-Transformer, MSSL hierarchically explores the complex interrelationships between temporal dynamics and facial expressions. Comparative experiments and ablation studies confirmed the effectiveness of our method on three large-scale in-the-wild datasets: DFEW, FERV39K, and MAFW.}, archive = {J_NEUCOM}, author = {Yang Lü and Fuchun Zhang and Zongnan Ma and Bo Zheng and Zhixiong Nan}, doi = {10.1016/j.neucom.2025.130020}, journal = {Neurocomputing}, month = {7}, pages = {130020}, shortjournal = {Neurocomputing}, title = {Dynamic facial expression recognition in the wild via multi-snippet spatiotemporal learning}, volume = {636}, year = {2025}, }
• (2025). Knowledge distillation for object detection with diffusion model. NEUCOM, 636, 130019. (www)

  @article{ , abstract = {Knowledge distillation is a method that transfers information from a larger network (i.e. the teacher) to a smaller network (i.e. the student), so that the student network can inherit the strong performance of the teacher network while maintaining its computational complexity within a relatively lower range. Currently, knowledge distillation has been widely applied to object detection field to mitigate the rapid expansion of the model size. In this paper, we propose an object detector based on knowledge distillation method. Meanwhile, directly mimicking the features of the teacher often fails to achieve the desired results due to the extra noise in the feature extracted by the student, which causes significant inconsistency and may even weaken the capability of the student. To address this issue, we utilize diffusion model to remove the noise so as to narrow the gap between the features extracted by the teacher and the student, improving the performance of the student. Furthermore, we develop a noise matching module that matches noise level in the student feature during the denoising process. Extensive experiments have been conducted on COCO and Pascal VOC to validate the effectiveness of the proposed method, in which our method achieves 40.0% mAP and 81.63% mAP respectively, while maintaining a frame rate of 27.3FPS, exhibiting the superiority of our model in both accuracy and speed.}, archive = {J_NEUCOM}, author = {Yi Zhang and Junzong Long and Chunrui Li}, doi = {10.1016/j.neucom.2025.130019}, journal = {Neurocomputing}, month = {7}, pages = {130019}, shortjournal = {Neurocomputing}, title = {Knowledge distillation for object detection with diffusion model}, volume = {636}, year = {2025}, }
• (2025). Multimodal multitask similarity learning for vision language model on radiological images and reports. NEUCOM, 636, 130018. (www)

  @article{ , abstract = {In recent years, large-scale Vision-Language Models (VLM) have shown promise in learning general representations for various medical image analysis tasks. However, current medical VLM methods typically employ contrastive learning approaches that have limited ability to capture nuanced yet crucial medical knowledge, particularly within similar medical images, and do not explicitly consider the uneven and complementary semantic information contained in different modalities. To address these challenges, we propose a novel Multimodal Multitask Similarity Learning (M2SL) method that learns joint representations of image–text pairs and captures the relational similarity between different modalities via a coupling network. Our method also notably leverages the rich information in the text inputs to construct a knowledge-driven semantic similarity matrix as the supervision signal. We conduct extensive experiments for cross-modal retrieval and zero-shot classification tasks on radiological images and reports and demonstrate substantial performance gains over existing methods. Our method also accommodates low-resource settings with limited training data availability and has significant implications for enhancing VLM development.}, archive = {J_NEUCOM}, author = {Yang Yu and Jiahao Wang and Weide Liu and Ivan Ho Mien and Pavitra Krishnaswamy and Xulei Yang and Jun Cheng}, doi = {10.1016/j.neucom.2025.130018}, journal = {Neurocomputing}, month = {7}, pages = {130018}, shortjournal = {Neurocomputing}, title = {Multimodal multitask similarity learning for vision language model on radiological images and reports}, volume = {636}, year = {2025}, }
• (2025). Pinning synchronization of higher-order nonlinear networks with time delays. NEUCOM, 636, 130010. (www)

  @article{ , abstract = {This paper investigates the pinning synchronization problem for a class of nonlinear higher-order time-delay networks. In contrast to previous works, the networks studied possess two distinguishing features: (i) the coupling functions governing the higher-order interactions are nonlinear, and (ii) the networks account for time delays, which include intra-node delays and higher-order interaction delays. The above two features are the key factors influencing the synchronization of the networks. By employing Lyapunov stability theory and algebraic graph theory, we derive sufficient conditions for achieving pinning synchronization in the nonlinear higher-order time-delay networks. Two key challenges in deriving the sufficient conditions arise from the two aforementioned features, which introduce: (i) complex tensor computations and (ii) difficulties in decoupling multi-node interactions. In order to address the challenges, we present a pivotal lemma. This lemma serves as a bridge to transform the complex higher-order problem into a first-order one, reducing the complexity of the derivations. Finally, the validity of the theoretical results is demonstrated through two application examples.}, archive = {J_NEUCOM}, author = {Weibin Li and Kaixin Lu and Zhichao Liang and Zhongye Xia and Bo Liu and Yanshan Xiao and Quanying Liu}, doi = {10.1016/j.neucom.2025.130010}, journal = {Neurocomputing}, month = {7}, pages = {130010}, shortjournal = {Neurocomputing}, title = {Pinning synchronization of higher-order nonlinear networks with time delays}, volume = {636}, year = {2025}, }
• (2025). M-MDD: A multi-task deep learning framework for major depressive disorder diagnosis using EEG. NEUCOM, 636, 130008. (www)

  @article{ , abstract = {Major depressive disorder (MDD) is a common and destructive psychiatric disorder worldwide. Traditional MDD diagnosis relies heavily on subjective observation and questionnaires. Recently, a non-invasive method of recording the brain’s spontaneous activity called Electroencephalogram (EEG) has been a useful tool of MDD diagnosis. However, there are still some challenges to be addressed: (1) The model’s robustness to common EEG noise has to be improved, (2) The temporal, spectral and spatial features of EEG need to be extracted and fused appropriately. Learning both robust and powerful features for MDD diagnosis can improve the overall performance, and multi-task learning is a powerful solution. In this paper, we propose M-MDD, a multi-task deep learning framework for MDD diagnosis using EEG. First, we design the Contrastive Noise Robustness Task to learn noise-independent features. Then, we design the Supervised Feature Extraction Task to extract temporal, spectral and spatial features of EEG respectively, and then effectively combine them together. Finally, the above two modules share the same feature space and are trained jointly with the Multi-task Learning Module, improving the overall performance. Validated on two public MDD diagnosis datasets with subject-independent cross-validation, our model achieves the state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Yilin Wang and Sha Zhao and Haiteng Jiang and Shijian Li and Tao Li and Gang Pan}, doi = {10.1016/j.neucom.2025.130008}, journal = {Neurocomputing}, month = {7}, pages = {130008}, shortjournal = {Neurocomputing}, title = {M-MDD: A multi-task deep learning framework for major depressive disorder diagnosis using EEG}, volume = {636}, year = {2025}, }
• (2025). Exploring interaction: Inner-outer spatial–temporal transformer for skeleton-based mutual action recognition. NEUCOM, 636, 130007. (www)

  @article{ , abstract = {Transformer-based methods have achieved significant results in the field of skeleton-based action recognition. However, when dealing with two-person interaction, existing approaches normally embed the skeleton of each person separately and then introduce an additional module to learn their interactions. This risks losing the spatial and semantic connection information between the two entities, which is crucial for interaction identification. To address this issue, a unified interactive spatial–temporal transformer is proposed in this paper. First, a Two-Person Embedding (TPE) is performed to provide a holistic interactive relationship representation, which can effectively avoid the information gap caused by the division of interacting entities. Second, an innovative Inner-Outer Transformer (IOformer) combining with a new spatio-temporal partition strategy is proposed to simultaneously learn the interactions between intra-partition joints and inter-partition skeletal parts. By comprehensively capturing the key spatio-temporal interactive feature, the accuracy and robustness of interaction recognition can be significantly improved. Extensive experiments on three challenging benchmark datasets validate that our method achieves better performance in comprehensive evaluation methods.}, archive = {J_NEUCOM}, author = {Xiaotian Wang and Xiang Jiang and Zhifu Zhao and Kexin Wang and Yifan Yang}, doi = {10.1016/j.neucom.2025.130007}, journal = {Neurocomputing}, month = {7}, pages = {130007}, shortjournal = {Neurocomputing}, title = {Exploring interaction: Inner-outer spatial–temporal transformer for skeleton-based mutual action recognition}, volume = {636}, year = {2025}, }
• (2025). Texture dominated no-reference quality assessment for high resolution image by multi-scale mechanism. NEUCOM, 636, 130003. (www)

  @article{ , abstract = {With the rapid development of new media formats, various high-definition display devices are ubiquitous, and high-resolution (HR) images are essential for high-quality visual experiences. Quality assessment of HR images has become an urgent challenge. However, conventional image quality assessment (IQA) methods with good performance are designed for low-resolution (LR) images, which lacks the perceptual characteristics of HR images, resulting in difficult to achieve satisfactory subjective consistency. Moreover, huge computational costs would have to be consumed when applying those deep neural networks in LR-IQA directly to HR images. Inspired by the fact that regions with rich textures are more sensitive to distortion than others, texture dominated no-reference image quality assessment for HR images are proposed in this paper. Specifically, a dual branch network based on multi-scale technology was designed to extract texture and semantic features separately, and cross scale and dual dimensional attention were introduced to ensure the dominance of texture features. Then, multi-layer perception network is used to map the extracted quality perception feature vectors to the predicted quality score. Worthy of note is that local entropy has been calculated and representative blocks are cropped as inputs to the feature extraction network, greatly reducing computational complexity. Overall, the texture dominated high-resolution IQA network (TD-HRNet) proposed utilizes a reference free method, while could perform excellently on HR datasets of different sizes, image types, and distortion types, accurately predicting the quality of different types of HR images.}, archive = {J_NEUCOM}, author = {Ziqing Huang and Hao Liu and Zhihao Jia and Shuo Zhang and Yonghua Zhang and Shiguang Liu}, doi = {10.1016/j.neucom.2025.130003}, journal = {Neurocomputing}, month = {7}, pages = {130003}, shortjournal = {Neurocomputing}, title = {Texture dominated no-reference quality assessment for high resolution image by multi-scale mechanism}, volume = {636}, year = {2025}, }
• (2025). Fixed-time synchronization in p-th moment for stochastic multi-layer neural networks: An adaptive graph-theoretic lyapunov functional approach. NEUCOM, 636, 130002. (www)

  @article{ , abstract = {In this paper, the p-th moment synchronization problem for a class of stochastic multi-layer neural networks with intra-layer and inter-layer connections is investigated. Due to the multiple connections with delays and stochastic noise, the typical methodologies that build a canonical linear or expanded matrix model to analyze its stability by constraining eigenvalues in the left-half plane, such as the Kronecker product method, linear matrix inequality and M -matrix approach are tough to tackle the problem. Consequently, a graph-theory-based Lyapunov functional is constructed by combining multiplicative principles and a graph-theoretic approach to help examine the effect of inter- and intra-layer connectivity on a unified framework. With the proposed adaptive fixed-time controller, sufficient conditions for the p-th moment synchronization in a fixed time are derived in terms of algebraic inequality. A corollary, together with a constant-gain fixed-time controller, is presented in case there is no delay. Finally, a confirmatory and two comparative simulations show the effectiveness and convenient implementation of the proposed control strategy.}, archive = {J_NEUCOM}, author = {Guan-Nan Yu and Xiao-Kang Liu and Yan Lei and Yan-Wu Wang}, doi = {10.1016/j.neucom.2025.130002}, journal = {Neurocomputing}, month = {7}, pages = {130002}, shortjournal = {Neurocomputing}, title = {Fixed-time synchronization in p-th moment for stochastic multi-layer neural networks: An adaptive graph-theoretic lyapunov functional approach}, volume = {636}, year = {2025}, }
• (2025). Lag-bipartite consensus control of nonlinear multi-agent systems with exogenous disturbances via dynamic event-triggered strategy. NEUCOM, 636, 130001. (www)

  @article{ , abstract = {This paper explores the lag-bipartite consensus issue for nonlinear multi-agent systems with external disturbances via event-triggered mechanisms. Firstly, a disturbance observer is devised to offset disturbances induced from ambient noise or parameter uncertainties. To save needless communication among neighbor agents and enhance the system’s anti-disturbance abilities, the centralized event-based approach and a distributed dynamic event-triggered control scheme with internal dynamic parameters are raised via combining the disturbance compensation strategy, respectively. Unlike existent static triggering approaches, this dynamic triggering scheme widens interval duration between two successive triggering instants. On the basis of both control schemes, a few sufficient conditions are provided to reach lag-bipartite consensus for nonlinear multi-agent systems, while Zeno behavior cannot arise via developed triggering rules. Finally, the validity of presented schemes is illustrated under numerical examples.}, archive = {J_NEUCOM}, author = {Junsheng Yu and Huizhi Xu and Zhongjun Ma}, doi = {10.1016/j.neucom.2025.130001}, journal = {Neurocomputing}, month = {7}, pages = {130001}, shortjournal = {Neurocomputing}, title = {Lag-bipartite consensus control of nonlinear multi-agent systems with exogenous disturbances via dynamic event-triggered strategy}, volume = {636}, year = {2025}, }
• (2025). Distributed approximate aggregative optimization of multiple Euler–Lagrange systems using only sampling measurements. NEUCOM, 636, 130000. (www)

  @article{ , abstract = {This article studies the distributed aggregative optimization for multiple Euler–Lagrange systems over directed networks. First, a new class of auxiliary aggregative variables is proposed that only utilize sampling measurements of adjacent outputs. Then, by selecting a smoothing function, we can gradually integrate the sampling information into new variables within the sampling period. Given the proposed variables, a key theorem is derived to transform the approximate aggregative optimization problem into a regulation problem, such that classical control methods can be utilized to regulate the aggregative variables for more complex dynamics. In addition, an adaptive fuzzy distributed control law is constructed based on aggregative variables, deadzone function and fuzzy system to solve the aggregative optimization for fully actuated Lagrangian agents with bounded disturbance. Finally, a numerical experiment is conducted to demonstrate the validity and effectiveness of the theoretical results.}, archive = {J_NEUCOM}, author = {Cong Li and Qingling Wang}, doi = {10.1016/j.neucom.2025.130000}, journal = {Neurocomputing}, month = {7}, pages = {130000}, shortjournal = {Neurocomputing}, title = {Distributed approximate aggregative optimization of multiple Euler–Lagrange systems using only sampling measurements}, volume = {636}, year = {2025}, }
• (2025). Spatial–spectral morphological mamba for hyperspectral image classification. NEUCOM, 636, 129995. (www)

  @article{ , abstract = {Recent advancements in transformers, specifically self-attention mechanisms, have significantly improved hyperspectral image (HSI) classification. However, these models often have inefficiencies, as their computational complexity scales quadratically with sequence length. To address these challenges, we propose the morphological spatial mamba (SMM) and morphological spatial–spectral Mamba (SSMM) model (MorpMamba), which combines the strengths of morphological operations and the state space model framework, offering a more computationally efficient alternative to transformers. In MorpMamba, a novel token generation module first converts HSI patches into spatial–spectral tokens. These tokens are then processed through morphological operations such as erosion and dilation, utilizing depthwise separable convolutions to capture structural and shape information. A token enhancement module refines these features by dynamically adjusting the spatial and spectral tokens based on central HSI regions, ensuring effective feature fusion within each block. Subsequently, multi-head self-attention is applied to enrich the feature representations further, allowing the model to capture complex relationships and dependencies within the data. Finally, the enhanced tokens are fed into a state space module, which efficiently models the temporal evolution of the features for classification. Experimental results on widely used HSI datasets demonstrate that MorpMamba achieves superior parametric efficiency compared to traditional CNN and transformer models while maintaining high accuracy. The source code is available at https://github.com/mahmad000/MorpMamba .}, archive = {J_NEUCOM}, author = {Muhammad Ahmad and Muhammad Hassaan Farooq Butt and Adil Mehmood Khan and Manuel Mazzara and Salvatore Distefano and Muhammad Usama and Swalpa Kumar Roy and Jocelyn Chanussot and Danfeng Hong}, doi = {10.1016/j.neucom.2025.129995}, journal = {Neurocomputing}, month = {7}, pages = {129995}, shortjournal = {Neurocomputing}, title = {Spatial–spectral morphological mamba for hyperspectral image classification}, volume = {636}, year = {2025}, }
• (2025). Decoding motor imagery hand direction in brain computer interface from direction-dependent modulation of parietal connectivity using a new brain functional connectivity measure. NEUCOM, 636, 129994. (www)

  @article{ , abstract = {The posterior Parietal Cortex (PPC) of human and nonhuman primates plays a vital role in motor planning. However, EEG functional connectivity correlates within PPC associated with motor intentions are less investigated in the literature. In this study, we investigate whether parietal EEG exhibits direction-dependent modulation of functional connectivity, during bidirectional hand movement imagination in right and left directions. Further, the utility of parietal connectivity modulation patterns, in decoding the directions of imagined hand movement is also evaluated. Imagined movement directions of the dominant hand are decoded using connectivity features derived from parietal EEG. A new brain functional connectivity measure called Cumulative Phase Lag is proposed to evaluate the functional connectivity within the right and left hemispheres of the posterior parietal cortex. Parietal connectivity features are derived from twenty-three EEG subbands from both hemispheres. Further, hemispherical asymmetry is exploited to identify the hemisphere with dominant directive discriminability. Connectivity features of the selected hemisphere are used to identify the most discriminative subband and selected features of the discriminative subband are used to classify the directions of imagined hand movement. The proposed algorithm employing subject-specific connectivity features yielded an average right vs left-hand motor imagery direction decoding accuracy of 79.67 % among 15 healthy subjects. The study results revealed that connectivity patterns in the posterior parietal cortex exhibited direction-dependent variability, suggesting a direction-dependent modulation of connectivity within the posterior parietal cortex. The results suggest the use of the posterior parietal cortex as a potential source of control signals for neuro-prosthetic applications.}, archive = {J_NEUCOM}, author = {K. Sagila Gangadharan and A.P. Vinod}, doi = {10.1016/j.neucom.2025.129994}, journal = {Neurocomputing}, month = {7}, pages = {129994}, shortjournal = {Neurocomputing}, title = {Decoding motor imagery hand direction in brain computer interface from direction-dependent modulation of parietal connectivity using a new brain functional connectivity measure}, volume = {636}, year = {2025}, }
• (2025). Zero-reference generative exposure correction and adaptive fusion for low-light image enhancement. NEUCOM, 636, 129992. (www)

  @article{ , abstract = {Existing low-light image enhancement methods have the problem of difficulty in enhancing dark areas while controlling overexposed areas in natural images. To address this issue, a Generative Exposure Correction method based on Retinex theory is proposed in this paper, in which the Pseudo-Exposure Residual map and illumination map are deeply coupled based on the proposed intensity compensation prior to constrain the generative network’s output in order to simultaneously deal with overexposure and underexposure. Furthermore, to enhance the effect and prevent over-correction, an exposure fusion technique is proposed, which adaptively selects the best exposure area from the two corrected images and achieves a globally balanced exposure by using an intensity correction compensation operator. More importantly, our proposed method does not require the collection of additional external datasets, which also overcomes the difficulty of data acquisition. Experimental comparisons of our method with the other seven state-of-the-art methods on five public datasets demonstrate that our method achieves the best performance in terms of detail enhancement and natural color preservation.}, archive = {J_NEUCOM}, author = {Qing Pan and Zirong Zhang and Nili Tian}, doi = {10.1016/j.neucom.2025.129992}, journal = {Neurocomputing}, month = {7}, pages = {129992}, shortjournal = {Neurocomputing}, title = {Zero-reference generative exposure correction and adaptive fusion for low-light image enhancement}, volume = {636}, year = {2025}, }
• (2025). Extending evolution-guided policy gradient learning into the multi-objective domain. NEUCOM, 636, 129991. (www)

  @article{ , abstract = {Multi-Objective Reinforcement Learning (MORL) poses significant challenges, primarily due to the necessity of balancing conflicting objectives—a limitation that traditional single-objective approaches fail to address. This paper introduces Multi-Objective Evolutionary Reinforcement Learning (MO-ERL), the first adaptation of Evolutionary Reinforcement Learning (ERL) specifically designed to address the complexities of the multi-objective domain effectively. MO-ERL integrates policy gradient-based reinforcement learning (RL), which optimizes expected utility, with evolutionary algorithms (EAs) that maintain diversity across the Pareto front. This combination leverages RL’s strength in exploitation and EAs’ proficiency in exploration, enabling MO-ERL to effectively navigate the trade-offs inherent in multi-objective optimization problems. Evaluation on multi-objective continuous control tasks using the MuJoCo physics engine demonstrates that MO-ERL outperforms state-of-the-art baselines, achieving up to 62.71% higher hypervolume and 196.28% greater expected utility. These results validate MO-ERL’s ability to balance solution diversity and optimality, setting a new benchmark for solving MORL tasks.}, archive = {J_NEUCOM}, author = {Adam Callaghan and Karl Mason and Patrick Mannion}, doi = {10.1016/j.neucom.2025.129991}, journal = {Neurocomputing}, month = {7}, pages = {129991}, shortjournal = {Neurocomputing}, title = {Extending evolution-guided policy gradient learning into the multi-objective domain}, volume = {636}, year = {2025}, }
• (2025). HSACT: A hierarchical semantic-aware CNN-transformer for remote sensing image spectral super-resolution. NEUCOM, 636, 129990. (www)

  @article{ , abstract = {Hyperspectral remote sensing technology has demonstrated its spectral diagnosis advantages in numerous remote sensing observation fields. However, hyperspectral imaging is expensive and less portable compared to RGB imaging. To recover the corresponding hyperspectral image (HSI) from a remote sensing RGB image, this paper proposes a new hierarchical semantic-aware convolutional neural network (CNN)-Transformer (HSACT) for remote sensing image spectral super-resolution (SSR). Particularly, this work aims to reconstruct HSIs from RGB images within the same field of view using a lightweight semantic embedding architecture. Our HSACT consists of the following steps. First, an initial spectrum estimation module (from the RGB image to the HSI) is designed to progressively consider spectral estimation between RGB wavelength-inner and wavelength-outer information. Then, an attention-driven semantic-aware CNN-Transformer is developed to reconstruct the spatial and spectral details of HSI. Specifically, a trainable polymorphic superpixel convolution (PSConv) is proposed to capture features efficiently in the above module. Next, we introduce an information-lossless hierarchical network architecture to link the above modules and achieve end-to-end RGB image SSR through weight sharing. Experimental results on several datasets demonstrated that our HSACT outperforms traditional and advanced SSR methods. The codes of this paper are available from https://github.com/chengle-zhou/HSACT .}, archive = {J_NEUCOM}, author = {Chengle Zhou and Zhi He and Liwei Zou and Yunfei Li and Antonio Plaza}, doi = {10.1016/j.neucom.2025.129990}, journal = {Neurocomputing}, month = {7}, pages = {129990}, shortjournal = {Neurocomputing}, title = {HSACT: A hierarchical semantic-aware CNN-transformer for remote sensing image spectral super-resolution}, volume = {636}, year = {2025}, }
• (2025). PCSViT: Efficient and hardware friendly pyramid vision transformer with channel and spatial self-attentions. NEUCOM, 636, 129987. (www)

  @article{ , abstract = {Vision Transformers (ViT) have been widely used in various visual tasks and have achieved great success due to their advantageous self-attention mechanism. However, most ViT models focus primarily on spatial self-attention, often overlooking the importance of channel attention. In this paper, we propose a channel self-attention module as a complementary addition to the standard self-attention module in ViTs. Then, we introduce an adaptive feed-forward network designed for different attention modules. Based on the proposed self-attention module and adaptive feed-forward network, we propose a flexible Vision Transformer with channel and spatial attentions (CSViT) and conduct a series of experiments to explore the optimal position of different attention modules. Additionally, we introduce PCSViT, which combines the strengths of CSViT and convolutional neural networks (CNNs). PCSViT features a pyramid architecture and incorporates local spatial attention, global spatial attention, and channel attention. We further explore hardware-friendly designs to efficiently implement and accelerate PCSViT on embedded devices. The performance of the proposed methods is evaluated on small datasets CIFAR and Fashion-MNIST, as well as the larger dataset ImageNet. Experimental results show that the proposed model reduces ViT’s reliance on large datasets and outperforms several lightweight state-of-the-art CNN and ViT models across a range of model sizes. The hardware-friendly designs achieve about 10% acceleration on a RISC-V CPU.}, archive = {J_NEUCOM}, author = {Xiaofeng Zou and Yuanxi Peng and Guoqing Li and Xinye Cao}, doi = {10.1016/j.neucom.2025.129987}, journal = {Neurocomputing}, month = {7}, pages = {129987}, shortjournal = {Neurocomputing}, title = {PCSViT: Efficient and hardware friendly pyramid vision transformer with channel and spatial self-attentions}, volume = {636}, year = {2025}, }
• (2025). Adjustable behavior-guided adaptive dynamic programming for neural learning control. NEUCOM, 636, 129986. (www)

  @article{ , abstract = {In this article, an adjustable behavior-guided adaptive dynamic programming (BGADP) algorithm is designed to solve the optimal regulation problem for discrete-time systems. In conventional adaptive dynamic programming methods, gradient information of system dynamics is necessary for conducting policy improvement. However, these methods face challenges when gradient information cannot be computed or when the system dynamics is non-differentiable. To overcome these limitations, a human-behavior-inspired swarm intelligence approach is used to search for superior policies during the iterative process, eliminating the need for gradient information. Additionally, a relaxation factor is introduced into the value function update to accelerate the convergence speed of the algorithm. The monotonicity and convergence properties of the iterative value function are rigorously analyzed. Finally, the effectiveness and practicality of the adjustable BGADP algorithm are validated through two simulation studies, which are implemented using the actor–critic framework with neural networks.}, archive = {J_NEUCOM}, author = {Guohan Tang and Ding Wang and Ao Liu and Junfei Qiao}, doi = {10.1016/j.neucom.2025.129986}, journal = {Neurocomputing}, month = {7}, pages = {129986}, shortjournal = {Neurocomputing}, title = {Adjustable behavior-guided adaptive dynamic programming for neural learning control}, volume = {636}, year = {2025}, }
• (2025). Grid mamba: Grid state space model for large-scale point cloud analysis. NEUCOM, 636, 129985. (www)

  @article{ , abstract = {Large-scale point cloud semantic segmentation aims to classify scene point clouds at the pixel level, which is crucial for understanding 3D real-world scenes. Existing Transformer-based models for point cloud segmentation face the challenge of quadratic computational complexity, limiting their ability to handle high-resolution and large-scale point clouds. Inspired by the recently proposed Mamba model, known for its efficient long-sequence modeling capabilities, we propose Grid Mamba in this work, which is a specialized network tailored for large-scale point cloud learning, achieving global linear computational complexity. Grid Mamba’s highlights consist of three parts: Grid Multi-view Scanning, Grid Sparsity Pooling, and Grid Mamba Block. Grid Multi-view Scanning can reduce the loss of spatial proximity caused by serialization. Grid Sparsity Pooling addresses the issue of local information loss during the pooling stage of large-scale point clouds. Additionally, Grid Mamba Block overcomes the limitations of Mamba in scene point cloud feature interactions. Extensive experimental results demonstrate that Grid Mamba achieves outstanding performance across multiple indoor and outdoor scene datasets.}, archive = {J_NEUCOM}, author = {Yulong Yang and Tianzhou Xun and Kuangrong Hao and Bing Wei and Xue-song Tang}, doi = {10.1016/j.neucom.2025.129985}, journal = {Neurocomputing}, month = {7}, pages = {129985}, shortjournal = {Neurocomputing}, title = {Grid mamba: Grid state space model for large-scale point cloud analysis}, volume = {636}, year = {2025}, }
• (2025). GCD-net: Global consciousness-driven open-vocabulary semantic segmentation network. NEUCOM, 636, 129982. (www)

  @article{ , abstract = {Open-vocabulary semantic segmentation aims to achieve accurate classification of different categories of pixels, even if these categories are not explicitly labeled during training. The current research trend in this field emphasizes the utilization of pre-trained visual–language models to augment exploration capabilities. The core of these methods is to use image-level models to guide the segmentation process at the pixel level, thereby enhancing the model’s ability to recognize and segment unseen categories during training. However, many approaches overlook global information, which may lead to a lack of comprehensive scene understanding when processing images. Thereby, GCD-Net is introduced as an innovative open-vocabulary semantic segmentation framework, which integrates a novel decoder with a hierarchical encoder to form an encoder–decoder architecture. The hierarchical encoder leverages a hierarchical backbone network to generate a pixel-level image–text cost map, which preserves spatial information effectively at different levels. The proposed decoder, known as the Feature Fusion Decoder, comprises three pivotal modules: the Global Feature Extraction Module, the Visual Enhancement Module, and the Feature Aggregation Module. These modules cooperate to process hierarchical feature maps from different levels to capture global context information and effectively aggregate pixel blocks into semantic regions for high-quality open-vocabulary semantic segmentation. Experiments on multiple open-vocabulary semantic segmentation datasets demonstrate that GCD-Net achieves an mIoU score of 17.5% on PC-459 and 94.3% on PAS-20, verifying the effectiveness and superiority of the method.}, archive = {J_NEUCOM}, author = {Xing Wu and Zhenyao Xu and Quan Qian and Bin Huang}, doi = {10.1016/j.neucom.2025.129982}, journal = {Neurocomputing}, month = {7}, pages = {129982}, shortjournal = {Neurocomputing}, title = {GCD-net: Global consciousness-driven open-vocabulary semantic segmentation network}, volume = {636}, year = {2025}, }
• (2025). Optimized inverse dead-zone formation control using reinforcement learning for the nonlinear single-integrator dynamic multi-agent system. NEUCOM, 636, 129981. (www)

  @article{ , abstract = {In this article, an optimized inverse dead-zone formation control using identifier–critic–actor reinforcement learning (RL) is studied for the nonlinear single-integral dynamic multi-agent system (MAS). Since MAS formation is often accompanied with a high energy expenditure, it is very necessary and essential to take optimization as a control design principle. In order to smoothly achieve the optimized MAS formation control, a simplified RL is developed by performing the gradient descent method to a simple positive function, which is equivalent to Hamilton–Jacobi–Bellman (HJB) equation. Furthermore, since the MAS formation cooperation is depended on the information exchange among agents, it is very possible to happen the control dead-zone phenomenon, which makes the actuator without control signal. For eliminating the effect of dead-zone, an adaptive inverse dead-zone method is developed and then is combined with RL for this optimized formation control. In comparison to the conventional inverse dead-zone approach, this design has the less adaptive parameters. Finally, the results of theoretical and simulation demonstrate viability of the proposed method.}, archive = {J_NEUCOM}, author = {Guoxing Wen and Wenxia Sun and Shuaihua Ma}, doi = {10.1016/j.neucom.2025.129981}, journal = {Neurocomputing}, month = {7}, pages = {129981}, shortjournal = {Neurocomputing}, title = {Optimized inverse dead-zone formation control using reinforcement learning for the nonlinear single-integrator dynamic multi-agent system}, volume = {636}, year = {2025}, }
• (2025). Shading- and geometry-aware lighting calibration network for uncalibrated photometric stereo. NEUCOM, 636, 129979. (www)

  @article{ , abstract = {Three-dimensional measurement provides essential geometric information for fault diagnosis and product optimization in intelligent manufacturing applications. Photometric stereo is a non-destructive 3D measurement technique that estimates the surface normals of objects using shading cues from images under different lighting conditions. However, the generalized bas-relief (GBR) ambiguity caused by unknown or varying lighting will significantly decrease measurement accuracy. To address this issue, we propose a shading- and geometry-aware lighting calibration network (SGLC-Net) to mitigate the inherent ambiguity and enhance surface normal estimation in uncalibrated photometric stereo by generating accurate lighting information. The proposed method iteratively optimizes lighting direction and intensity by leveraging self-generated shading and normal prior features. To further improve the accuracy of the lighting estimation, we introduce collocated light into SGLC-Net to implicitly extract shading features of images to generate accurate rough lighting. Accurate rough lighting can generate accurate shading and normal prior features, which can be used to optimize rough lighting to generate fine lighting. Experimental results indicate that the proposed method significantly outperforms most uncalibrated photometric stereo methods in lighting estimation on multiple real-world datasets. Furthermore, our method can seamlessly integrate with most uncalibrated photometric stereo methods to effectively enhance the accuracy of the surface normal estimation under unknown illumination.}, archive = {J_NEUCOM}, author = {Yuze Yang and Jiahang Liu and Yangyu Fu and Yue Ni and Yan Xu}, doi = {10.1016/j.neucom.2025.129979}, journal = {Neurocomputing}, month = {7}, pages = {129979}, shortjournal = {Neurocomputing}, title = {Shading- and geometry-aware lighting calibration network for uncalibrated photometric stereo}, volume = {636}, year = {2025}, }
• (2025). Enhancing offline reinforcement learning for wastewater treatment via transition filter and prioritized approximation loss. NEUCOM, 636, 129977. (www)

  @article{ , abstract = {Wastewater treatment plays a crucial role in urban society, requiring efficient control strategies to optimize its performance. In this paper, we propose an enhanced offline reinforcement learning (RL) approach for wastewater treatment. Our algorithm improves the learning process. It uses a transition filter to sort out low-performance transitions and employs prioritized approximation loss to achieve prioritized experience replay with uniformly sampled loss. Additionally, the variational autoencoder is introduced to address the problem of distribution shift in offline RL. The proposed approach is evaluated on a nonlinear system and wastewater treatment simulation platform, demonstrating its effectiveness in achieving optimal control. The contributions of this paper include the development of an improved offline RL algorithm for wastewater treatment and the integration of transition filtering and prioritized approximation loss. Evaluation results demonstrate that the proposed algorithm achieves lower tracking error and cost.}, archive = {J_NEUCOM}, author = {Ruyue Yang and Ding Wang and Menghua Li and Chengyu Cui and Junfei Qiao}, doi = {10.1016/j.neucom.2025.129977}, journal = {Neurocomputing}, month = {7}, pages = {129977}, shortjournal = {Neurocomputing}, title = {Enhancing offline reinforcement learning for wastewater treatment via transition filter and prioritized approximation loss}, volume = {636}, year = {2025}, }
• (2025). VLSG-net: Vision-language scene graphs network for paragraph video captioning. NEUCOM, 636, 129976. (www)

  @article{ , abstract = {Paragraph Video captioning seeks to automatically describe multiple events in a video. Despite significant progress, most current approaches fail to fully leverage scene graph knowledge when performing cross-modal alignment between video and text representations. Consequently, such methods may not learn causal associations between entities, leading to a degradation in captioning performance. In this paper, we propose an end-to-end Vision-Language Scene Graphs Network (VLSG-net) to address this issue. We first introduce an encoder that integrates scene graph knowledge with global features and predicates to understand visual scenes. Specifically, scene graph knowledge detects entities and models their correlations and constraints, enabling the representation of relationships between various entities. We then introduce a Knowledge-Enhanced Encoder paired with a contrastive loss to leverage scene graph knowledge, thereby enhancing multimodal structured representations. Finally, we propose a transformer-in-transformer decoder to model the coherency of intra- and inter-event relationships within the video and generate captions. By incorporating relationship reasoning among entities through scene graphs and video-language alignment learning, VLSG-net generates more logical and detailed captions. Extensive experiments confirm that VLSG-net performs favorably against the state-of-the-art methods on two widely used benchmark datasets, ActivityNet Captions, and YouCookII.}, archive = {J_NEUCOM}, author = {Yufeng Hou and Qingguo Zhou and Hui Lv and Lan Guo and Yan Li and La Duo and Zhenyu He}, doi = {10.1016/j.neucom.2025.129976}, journal = {Neurocomputing}, month = {7}, pages = {129976}, shortjournal = {Neurocomputing}, title = {VLSG-net: Vision-language scene graphs network for paragraph video captioning}, volume = {636}, year = {2025}, }
• (2025). Privacy-preserving and byzantine-robust federated broad learning with chain-loop structure. NEUCOM, 636, 129975. (www)

  @article{ , abstract = {Federated learning (FL) can collaboratively train a model by aggregating local models instead of aggregating raw data, which can protect privacy by ensuring that data remains on the client. However, the traditional FL still faces some challenges such as privacy leakage and the presence of Byzantine clients. We propose a privacy-preserving and Byzantine-robust federated broad learning framework with chain-loop structure i.e., PBFBL-CL, and this algorithm can simultaneously achieve protection of clients’ privacy and robustness against Byzantine attacks. In this paper, we apply Byzantine step-by-step co-validation algorithm to address the existence of Byzantine clients. We pass the aggregated model through the chain, so each client’s privacy is well protected. Moreover, PBFBL-CL can reduce the communication overhead between clients and server. Finally, we evaluate the PBFBL-CL algorithm in MNIST, Fashion-MNIST and NORB datasets, and the results show that our algorithm is better than existing FL algorithms in terms of model accuracy and training speed. Experimental results demonstrate that under the extreme scenario where Byzantine client proportion reaches 90%, the model achieves an accuracy of 89.53%, only 4.17% lower than the 93.7% accuracy observed in the ideal scenario without Byzantine clients.}, archive = {J_NEUCOM}, author = {Nan Li and Chang-E Ren and Siyao Cheng}, doi = {10.1016/j.neucom.2025.129975}, journal = {Neurocomputing}, month = {7}, pages = {129975}, shortjournal = {Neurocomputing}, title = {Privacy-preserving and byzantine-robust federated broad learning with chain-loop structure}, volume = {636}, year = {2025}, }
• (2025). Spectrum-guided spatial feature enhancement network for event-based lip-reading. NEUCOM, 636, 129974. (www)

  @article{ , abstract = {The Automatic Lip-reading task aims to recognize spoken words through visual cues from the speaker’s lip movements. This crucial task complements audio-based speech recognition systems and can substitute them when sound is unavailable. Event-based lip-reading methods have gained increasing attention due to the advantages of event cameras, such as high temporal resolution and low power consumption. However, existing methods often fail to fully utilize the spatial information in event data due to its sparsity and the presence of random activations. To address this, we propose a novel Spectral-guided Spatial Enhancement Network (SSE-Net). SSE-Net introduces two core innovations: the Spectrum-guided Spatial Feature Enhance Module (SSEM) and the Multi-Scale Spatial Interaction Module (MS-SIM). SSEM employs frequency domain enhancement and spatial feature enhancement strategies to augment spatial features crucial for event-based lipreading tasks. MS-SIM conducts the fusion and interaction of multi-level semantics, enriching the contextual information of lip representations. We conducted experiments on the event-based lip-reading dataset DVS-Lip with our proposed method and demonstrated its superiority over other state-of-the-art event-based lip-reading methods.}, archive = {J_NEUCOM}, author = {Yi Zhang and Xiuping Liu and Hongchen Tan and Xin Li}, doi = {10.1016/j.neucom.2025.129974}, journal = {Neurocomputing}, month = {7}, pages = {129974}, shortjournal = {Neurocomputing}, title = {Spectrum-guided spatial feature enhancement network for event-based lip-reading}, volume = {636}, year = {2025}, }
• (2025). HRL-painter: Optimal planning painter based on hierarchical reinforcement learning. NEUCOM, 636, 129972. (www)

  @article{ , abstract = {Stroke-based rendering method has shown its superiority in generating stylized paintings from realistic photographs. However, the existing methods often divide the image into regular blocks for parallel painting or start painting by progressively narrowing down the painting region from the entire canvas. Not only does this lead to an irrational allocation of stroke resources, but also deviates from the painting approach employed by human artists. To address this, we propose a novel painting method based on hierarchical reinforcement learning, namely HRL-Painter, which consists of a high-level agent that strategically plans the sequence of painting regions and a low-level agent that carries out specific painting tasks in the corresponding regions. In the initial stage, we consider the entire canvas as the painting region and then use a small number of strokes for a rough depiction. Next, our high-level agent plans the optimal sequence of painting regions based on the content of the target image, taking into account the error between the current canvas and the target image. Finally, the low-level agent is dedicated to executing detailed painting tasks within the painting regions proposed by the high-level agent. Extensive experiments on standard datasets including CelebA , ImageNet , CUB-200 Birds and Stanford Cars-196 demonstrate that our proposed hierarchical painting agent not only produce high-quality canvases but also exhibit a painting process that closely resembles the human painting style, showcasing excellent interpretability.}, archive = {J_NEUCOM}, author = {Jiong Zhang and Guangxin Xu and Xiaoyan Zhang}, doi = {10.1016/j.neucom.2025.129972}, journal = {Neurocomputing}, month = {7}, pages = {129972}, shortjournal = {Neurocomputing}, title = {HRL-painter: Optimal planning painter based on hierarchical reinforcement learning}, volume = {636}, year = {2025}, }
• (2025). Distributed neural predictor enhanced coordinated control of AUVs. NEUCOM, 636, 129971. (www)

  @article{ , abstract = {This article investigates an enhanced tunnel prescribed performance coordinated control problem of multiple autonomous underwater vehicles (AUVs) under initial constraints. To meet high performance requirements in complex underwater conditions, AUV control faces challenges. In order to address these, an enhanced tunnel prescribed performance (ETPP) method is proposed, which is composed of composite error scaling function (CESF) and tunnel prescribed performance (TPP). In particular, a CESF-based error transformation is performed to scale the tracking error within the TPP limits. In the guidance loop, an ETPP-based guidance law is devised to guarantee the transient and steady-state behavior of the tracking error. In the control loop, based on the distributed learning strategy with weighted average, a quantized input-based distributed neural predictor (QDNP) is proposed to estimate the unknown external disturbances. Using the antidisturbance technique, a QDNP-based quantized control law is designed to stabilize multi-AUV formations. The uniformly ultimately bounded (UUB) stability of the overall closed-loop system is established in the Lyapunov sense. Finally, simulation examples with four AUVs are provided to demonstrate the effectiveness of the proposed distributed tunnel performance-guaranteed coordinated control method.}, archive = {J_NEUCOM}, author = {Minjing Wang and Di Wu and Lei Qiao and Rui Gao and Wenlong Feng}, doi = {10.1016/j.neucom.2025.129971}, journal = {Neurocomputing}, month = {7}, pages = {129971}, shortjournal = {Neurocomputing}, title = {Distributed neural predictor enhanced coordinated control of AUVs}, volume = {636}, year = {2025}, }
• (2025). Rad-mark: Reliable adversarial zero-watermarking. NEUCOM, 636, 129970. (www)

  @article{ , abstract = {Zero-watermarking is a lossless protection technique, and thus it is widely used in medical images, artworks and other carriers that require lossless protection. However, the current zero-watermarking suffers from the problem of high similarity between the feature images of different host images, which results in a high false positive rate. To address this challenge, we propose Rad-Mark, a deep learning-based zero-watermarking framework that leverages adversarial feature optimization to enhance the robustness and accuracy of watermark detection significantly for the first time. The adversarial samples are employed to significantly improve the framework’s security, which can achieve the NC value of false positives close to 0.5. Both image perturbation and Gaussian noise are incorporated into the training process. Specifically, our Rad-Mark involves a feature fusion design, a mapping network based on the fusion of locally filtered and global handcrafted features. We conduct an in-depth analysis of key parameters, including Gaussian noise, watermark dimensions, and weighting factors, exploring their impact on the performance of our Rad-Mark. Extensive experimental results demonstrate that Rad-Mark outperforms existing zero-watermarking methods in terms of both security and robustness.}, archive = {J_NEUCOM}, author = {Kun Hu and Dakai Zhai and Heng Gao and Haoyu Xie and Xingjun Wang}, doi = {10.1016/j.neucom.2025.129970}, journal = {Neurocomputing}, month = {7}, pages = {129970}, shortjournal = {Neurocomputing}, title = {Rad-mark: Reliable adversarial zero-watermarking}, volume = {636}, year = {2025}, }
• (2025). An event-triggered reliable cloud control scheme based on ADP and integral sliding mode. NEUCOM, 636, 129968. (www)

  @article{ , abstract = {This paper investigates event-triggered (ET) reliable control problems for cloud control systems under actuator faults and data injection attacks via the adaptive dynamic programming (ADP) and integral sliding mode (ISM). A mist-fog-regional cloud control architecture is first given, which can improve computing efficiency of the cloud platform. In this architecture, a fog-based fault parameter estimation method is proposed with the aid of neural networks. It is driven by the feedback of fault parameter estimation errors, so as to achieve more accurate estimations of fault parameters. A double ET reliable cloud control scheme is further presented. It is composed of an ISM-based and an ADP-based regional cloud controllers. As a result, it not only saves communication resources, but also eliminates the influence of the attacks and matched uncertainties, as well as ensures the stability of the equivalent sliding-mode dynamics with optimal performance. Finally, the effectiveness of the proposed method is verified by the simulation results.}, archive = {J_NEUCOM}, author = {Xin Huang and Sicheng Bi and Xinyu Han and Shuyi Xiao and Qingyu Su}, doi = {10.1016/j.neucom.2025.129968}, journal = {Neurocomputing}, month = {7}, pages = {129968}, shortjournal = {Neurocomputing}, title = {An event-triggered reliable cloud control scheme based on ADP and integral sliding mode}, volume = {636}, year = {2025}, }
• (2025). A memory failure computational model in alzheimer-like disease via continuous delayed hopfield network with lurie control system based healing. NEUCOM, 636, 129967. (www)

  @article{ , abstract = {Alzheimer’s disease (AD) is a degenerative neurological condition that impacts millions of individuals across the globe and remains without a healing. In the search for new possibilities of treatments for this terrible disease, this work presents the improved Alzheimer-like disease (IALD) model for memory failure and connects it to a new control technique that establishes a cure for the memory lost, either in biological or in artificial neural networks. For the IALD model, continuous Hopfield neural networks (HNN) with time delay are used. From the healing side, a robust control technique is used, which is based on new discoveries in Lurie control systems. In addition, this paper reviews the development of Alzheimer-like disease (ALD) model, as well as, the relationship of HNN with Lurie system. Simulations are executed to validate the model and to show the efficacy of applying a new theorem from Lurie problem. With the results presented, this work proposes a new conceptual paradigm that could potentially be applied in memory failure treatments in AD, as well as in hardware implemented HNN under adversarial attacks or adverse environmental conditions.}, archive = {J_NEUCOM}, author = {Rafael Fernandes Pinheiro and Diego Colón and Rui Fonseca-Pinto}, doi = {10.1016/j.neucom.2025.129967}, journal = {Neurocomputing}, month = {7}, pages = {129967}, shortjournal = {Neurocomputing}, title = {A memory failure computational model in alzheimer-like disease via continuous delayed hopfield network with lurie control system based healing}, volume = {636}, year = {2025}, }
• (2025). Fine-grained hierarchical singular value decomposition for convolutional neural networks compression and acceleration. NEUCOM, 636, 129966. (www)

  @article{ , abstract = {Convolutional neural networks (CNNs) still remain crucial in the field of computer vision, especially in industrial-embedded scenarios. Although modern artificial intelligence chips such as embedded graphics processing units (GPUs) and neural process units (NPUs) are equipped with sufficient computability, making CNNs more lightweight always has non-negligible significance. Until now, many researchers have made multiple corresponding achievements, in which a series of tensor decomposition methods have represented their unique advantages such as concision, flexibility, and low-rank approximation theory. However, balancing the compression, acceleration, and precision, is still an open issue, because the traditional tensor decompositions are hard to deal with the trade-off between approximation and compression ability, while the so-called fine-grained tensor decompositions such as Kronecker canonical polyadic (KCP) have not created a way to merge the factors for efficient inference. In this paper, we first review related works on convolutional neural network (CNN) compression and the necessary prior knowledge. We then propose a novel matrix decomposition method, termed hierarchical singular value (HSV) decomposition, and validate its effectiveness. Subsequently, we introduce a fast contraction strategy based on the merged factors of HSV and explain how our method addresses the inefficiencies in inference associated with traditional contraction processes. Additionally, we validate the advantages of HSV by comparing its complexity with that of other classical tensor decomposition methods. Thereafter, we apply HSV to CNN compression and acceleration by transforming convolution operations into matrix multiplication. We also propose a self-adaptive rank selection algorithm tailored to standard CNN architecture and conduct a theoretical analysis of the convergence of our method. Multiple experiments on CIFAR-10, ImageNet, COCO, and Cityscapes benchmark datasets show that the proposed HSV-Conv can simultaneously gain considerable compression ratio and acceleration ratio, while the precision loss is almost non-existent. We also make a comprehensive comparison with the other related works, and the superiority of our method is further validated. Besides, we give a deep discussion about the rank selection issue of HSV in the aspects of practice and theory, which explains the strategy of the proposed self-adaptive rank selection and the reason for choosing fine-tuning rather than training from scratch.}, archive = {J_NEUCOM}, author = {Mengmeng Qi and Dingheng Wang and Wei Yang and Baorong Liu and Fuyong Wang and Zengqiang Chen}, doi = {10.1016/j.neucom.2025.129966}, journal = {Neurocomputing}, month = {7}, pages = {129966}, shortjournal = {Neurocomputing}, title = {Fine-grained hierarchical singular value decomposition for convolutional neural networks compression and acceleration}, volume = {636}, year = {2025}, }
• (2025). Multi-level feature splicing 3D network based on multi-task joint learning for video anomaly detection. NEUCOM, 636, 129964. (www)

  @article{ , abstract = {In video anomaly detection research, deep learning is dedicated to identifying anomalous events accurately and efficiently. However, due to the scarcity and diversity of anomaly samples, previous methods have not adequately taken into account important information about location and timing. In addition, the overpowered generalization ability of the models leads to the fact that anomalies can also be well reconstructed or predicted. To address the above challenges, we propose a 3D network based on multi-level feature splicing with joint multi-task learning. The network is improved by the autoencoder (AE) as a backbone network. Firstly, we design a normal sample training task and a Gaussian noise task from a spatial perspective to enhance the reconstruction of positive samples. The frame-skipping task and the inverse sequence task of the video are designed from the temporal perspective to suppress the reconstruction ability of negative samples. Secondly, we use multi-level feature splicing in the encoding and decoding process to equip the network with the ability to explore sufficient information from the full scale. At the same time, we use an attention gating module to filter redundant features. The results show that our network is competitive with state-of-the-art methods. In terms of AUC, UCSD Ped2 achieves 99.3%, CUHK Avenue achieves 88.4%, and ShanghaiTech Campus achieves 74.2%.}, archive = {J_NEUCOM}, author = {Yang Li and Guoxiang Tong}, doi = {10.1016/j.neucom.2025.129964}, journal = {Neurocomputing}, month = {7}, pages = {129964}, shortjournal = {Neurocomputing}, title = {Multi-level feature splicing 3D network based on multi-task joint learning for video anomaly detection}, volume = {636}, year = {2025}, }
• (2025). L2M-GCN: A new framework for learning robust GCN against structural attacks. NEUCOM, 636, 129962. (www)

  @article{ , abstract = {Graph Convolutional Networks (GCNs) have gained extensive attention due to their strong ability to learn from graphs. However, with the advent of stealthy attacks that cause significant differences in node embeddings, the vulnerability of GCNs to malicious attacks has been exposed. Although there are many studies on defense in the spatial or spectral domains, they neglect the complementary roles of the two. In this paper, we propose a new framework, Low frequency and 2-hop in Multi-channel GCN (L2M-GCN), which combines spatial and spectral defense. L2M-GCN has two GCN-based modules. In module one, a new structure reconstructed from learnable spectrum and low-frequency components replaces the adjacency matrix in GCN. In module two, purified 2-hop is introduced and the attention mechanism is used to learn the importance weights of node embeddings. The two modules are eventually assembled into L2M-GCN for joint learning in a parameter-sharing and end-to-end fashion. Extensive experiments demonstrate that L2M-GCN significantly improves the defense performance against structural attacks and outperforms the baselines and state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Haoran Chen and Xianchen Zhou and Jiwei Zhang and Hongxia Wang}, doi = {10.1016/j.neucom.2025.129962}, journal = {Neurocomputing}, month = {7}, pages = {129962}, shortjournal = {Neurocomputing}, title = {L2M-GCN: A new framework for learning robust GCN against structural attacks}, volume = {636}, year = {2025}, }
• (2025). Channel pruning for convolutional neural networks using l0-norm constraints. NEUCOM, 636, 129925. (www)

  @article{ , abstract = {Channel pruning can effectively reduce the size and inference time of Convolutional Neural Networks (CNNs). However, existing channel pruning methods still face several issues, including high computational costs, extensive manual intervention, difficulty in hyperparameter tuning, and challenges in directly controlling the sparsity. To address these issues, this paper proposes two channel pruning methods based on l 0 -norm sparse optimization: the l 0 -norm Pruner and the Automated l 0 -norm Pruner. The l 0 -norm Pruner formulates the channel pruning problem as a sparse optimization problem involving the l 0 -norm and achieves a fast solution through a series of approximations and transformations. Inspired by this solution process, we devise the Zero-Norm (ZN) module, which can autonomously select output channels for each layer based on a predefined global pruning ratio. This approach incurs low computational cost and allows for precise control over the overall pruning ratio. Furthermore, to further enhance the performance of the pruned model, we have developed the Automated l 0 -norm Pruner. This method utilizes a Bee Colony Optimization algorithm to adjust the pruning ratio, mitigating the negative impact of manually preset pruning ratios on model performance. Our experiments demonstrate that the proposed pruning methods outperform several state-of-the-art techniques. The source code for our proposed methods is available at: https://github.com/TCCofWANG/l0_prune .}, archive = {J_NEUCOM}, author = {Enhao Chen and Hao Wang and Zhanglei Shi and Wei Zhang}, doi = {10.1016/j.neucom.2025.129925}, journal = {Neurocomputing}, month = {7}, pages = {129925}, shortjournal = {Neurocomputing}, title = {Channel pruning for convolutional neural networks using l0-norm constraints}, volume = {636}, year = {2025}, }
• (2025). Cascade-UDA: A cascade paradigm for unsupervised domain adaptation. NEUCOM, 636, 129924. (www)

  @article{ , abstract = {Unsupervised domain adaptation (UDA) aims to enhance the performance of models on unlabeled target domains by utilizing labeled data from a related source domain. However, existing UDA methods often struggle with semantic confusion and distribution shifts. To address these issues, we propose a novel two-stage UDA framework called Cascade-UDA. In the first stage, we fine-tune CLIP-LoRA on the source domain to learn class-related, domain-invariant features while preserving semantic integrity. In the second stage, we freeze the fine-tuned CLIP-LoRA and introduce a textual prompt for target domain adaptation, refining pseudo-labels with knowledge from the source domain. Our proposed method effectively decouples semantic learning from domain-specific adaptation, enhancing performance on the target domain. Extensive experiments on public datasets demonstrate the superiority of our approach over existing methods.}, archive = {J_NEUCOM}, author = {Mengmeng Zhan and Zongqian Wu and Huafu Xu and Xiaofeng Zhu and Rongyao Hu}, doi = {10.1016/j.neucom.2025.129924}, journal = {Neurocomputing}, month = {7}, pages = {129924}, shortjournal = {Neurocomputing}, title = {Cascade-UDA: A cascade paradigm for unsupervised domain adaptation}, volume = {636}, year = {2025}, }
• (2025). YOLO-ELWNet: A lightweight object detection network. NEUCOM, 636, 129904. (www)

  @article{ , abstract = {This paper proposes a YOLO-based efficient lightweight network (YOLO-ELWNet) for onboard object detection based on the YOLOv3. A channel split and shuffle with coordinate attention module is developed in the backbone block, which effectively reduces the size of model parameters and computational cost while maintaining the detection accuracy. A new feature fusion network is proposed in the neck block, where a cross-stage partial with efficient bottleneck module is put forward to improve the feature extraction ability and reduce the computational cost. The Scylla intersection over union-based loss function is utilized in the head block, which accelerates the convergence speed of the YOLO-ELWNet. The effectiveness of the proposed YOLO-ELWNet is validated on the open source KITTI vision benchmark. The performance of YOLO-ELWNet is superior to some mainstream lightweight object detection models in terms of detection accuracy and computational cost, which demonstrates its applicability for resource-constrained onboard object detection.}, archive = {J_NEUCOM}, author = {Baoye Song and Jianyu Chen and Weibo Liu and Jingzhong Fang and Yani Xue and Xiaohui Liu}, doi = {10.1016/j.neucom.2025.129904}, journal = {Neurocomputing}, month = {7}, pages = {129904}, shortjournal = {Neurocomputing}, title = {YOLO-ELWNet: A lightweight object detection network}, volume = {636}, year = {2025}, }
• (2025). DCFT: Dependency-aware continual learning fine-tuning for sparse LLMs. NEUCOM, 636, 129897. (www)

  @article{ , abstract = {As the size of Large Language Models (LLMs) increasing, they exhibit enhanced capabilities in general intelligence but also present greater challenges in deployment. Consequently, compressing LLMs has become critically important. Among the various compression techniques, post-training pruning is highly favored by researchers due to its efficiency. However, this one-shot pruning approach often results in a significant deterioration of model performance. To mitigate this issue, we introduce Dependency-aware Continual learning Fine-Tuning (DCFT) for sparse LLMs. This method facilitates fine-tuning across sequential tasks without compromising the model’s sparsity. Initially, we revisit the inference process in LLMs from a novel perspective, treating two matrices that previously required independent optimization as a unified entity. This strategy involves introduces merely 0.011‰ additional parameters to achieve efficient fine-tuning. Furthermore, we re-evaluate the parameter fine-tuning process through the lens of matrix space mapping. By constraining the similarity of the mapping matrices, our approach enables the model to retain its performance on prior tasks while learning new ones. We tested our method on models from the LLaMA-V1/V2 families, with parameters ranging from 7B to 70B, and under various sparsity ratios and patterns (unstructured and N:M sparsity). The results consistently demonstrate outstanding performance.}, archive = {J_NEUCOM}, author = {Yanzhe Wang and Yizhen Wang and Baoqun Yin}, doi = {10.1016/j.neucom.2025.129897}, journal = {Neurocomputing}, month = {7}, pages = {129897}, shortjournal = {Neurocomputing}, title = {DCFT: Dependency-aware continual learning fine-tuning for sparse LLMs}, volume = {636}, year = {2025}, }
• (2025). Parallel spatiotemporal network to recognize micro-expression. NEUCOM, 636, 129891. (www)

  @article{ , abstract = {Micro-expressions are fleeting spontaneous facial expressions that commonly occur in high-stakes scenarios and reflect humans’ mental states. Thus, it is one of the crucial clues for lie detection. Furthermore, due to the brief duration of micro-expression, temporal information is important for micro-expression recognition. The paper proposes a Parallel Spatiotemporal Network (PSN) to recognize micro-expression. The proposed PSN includes a spatial sub-network and a temporal sub-network. The spatial sub-network is a shallow network with subtle motion information as the input. And the temporal sub-network is a network with a novel temporal feature extraction unit that extracts sparse temporal features of micro-expressions. Finally, we propose an element-wise addition with 1 × 1 convolutional kernel fusion model to fuse the spatial and temporal features. The proposed PSN gets better measurement metrics (such as recognition rate, F1 score, true positive rate, and true negative rate) than the other state-of-the-art methods on the consisted databases consisting of CASME, CASME II, CAS(ME) 2 , and SAMM.}, archive = {J_NEUCOM}, author = {Jingting Li and Su-Jing Wang and Yong Wang and Haoliang Zhou and Xiaolan Fu}, doi = {10.1016/j.neucom.2025.129891}, journal = {Neurocomputing}, month = {7}, pages = {129891}, shortjournal = {Neurocomputing}, title = {Parallel spatiotemporal network to recognize micro-expression}, volume = {636}, year = {2025}, }
• (2025). Farewell to CycleGAN: Single GAN with decoupled constraint for unpaired image dehazing. NEUCOM, 636, 129888. (www)

  @article{ , abstract = {Unpaired image dehazing has attracted more and more attention, since the pair-wise training data which is prerequisite for the supervised dehazing methods leads to high cost if they are really captured or performance degradation on the real-hazy scenes if they are synthesized. The existing methods for unpaired image dehazing are all based on the CycleGAN-like framework with pixel-to-pixel constraint, which leads to burdensome model complexity and unstable training. In this paper, we propose a novel single GAN model for unpaired image dehazing (SinGAN-Dehaze), which gets rid of the cycle-consistency constraint. To be specific, the cycle-consistency is decoupled to content-consistency and style-consistency, where the pixel-to-pixel mapping is replaced by the patch-to-patch semantic mapping. The content-consistency is ensured by capturing local distinctive representations and global contextual dependencies. The style-consistency is achieved by forcing the high-frequency information distribution of dehazing result close to that of the clear image with similar style. Extensive experiments demonstrate that our proposal can achieve superior performance for unpaired image dehazing in terms of the objective index and visual effect on both synthetic and real-hazy scenarios.}, archive = {J_NEUCOM}, author = {Xiaotong Luo and Wenjin Yang and Yuan Xie and Yanyun Qu}, doi = {10.1016/j.neucom.2025.129888}, journal = {Neurocomputing}, month = {7}, pages = {129888}, shortjournal = {Neurocomputing}, title = {Farewell to CycleGAN: Single GAN with decoupled constraint for unpaired image dehazing}, volume = {636}, year = {2025}, }
• (2025). A hybrid approach combining sentiment analysis and deep learning to mitigate data sparsity in recommender systems. NEUCOM, 636, 129886. (www)

  @article{ , abstract = {The optimization of recommendation systems (RS) is crucial for delivering personalized product suggestions. Despite their successes, RS approaches often face challenges, such as data sparsity in the user–item matrix, which can undermine their performance. To address these challenges, integrating additional information sources, such as item/user profiles and textual reviews, is essential. These sources offer valuable insights into user preferences and item characteristics, helping in understanding the contextual details of both. This study focuses on developing an advanced RS architecture that combines Singular Value Decomposition (SVD) with BERT-CB methods and a Hybrid Model-based Sentiment Analysis. By integrating BERT with Multilayer Perceptron (MLP) methods, the system gains a deeper understanding of item profiles, improving the comprehension of user preferences and item characteristics. Additionally, a novel hybrid approach for sentiment analysis is proposed, using GloVe embeddings and CNN-BiGRU, improving the accuracy and robustness of sentiment detection in user reviews. This comprehensive understanding, combined with collaborative filtering models like SVD, enables the system to provide highly accurate recommendations. The proposed approach consists of four main phases: first, embedding review text using GloVe embeddings and developing a hybrid sentiment analysis approach with CNN and BiGRU architectures; second, creating a BERT language model for generating embeddings from item profile texts, followed by dimensionality reduction using Auto-Encoder; third, using these vectors to build a novel MLP model; fourth, developing a Collaborative Filtering method using SVD, and finally, combining these methods into a hybrid approach and conducting a comprehensive evaluation. Empirical results clearly show the effectiveness of our approach, particularly the combination of GloVe-CNN-BiGRU and BERT-CB with SVD methodology, demonstrating significant improvements across various performance metrics. This confirms the practical value of using contextualized data from BERT-CB and the sentiment analysis approach, enhancing the recommendation system’s effectiveness.}, archive = {J_NEUCOM}, author = {Ikram Karabila and Nossayba Darraz and Anas El-Ansari and Nabil Alami and Mostafa El Mallahi}, doi = {10.1016/j.neucom.2025.129886}, journal = {Neurocomputing}, month = {7}, pages = {129886}, shortjournal = {Neurocomputing}, title = {A hybrid approach combining sentiment analysis and deep learning to mitigate data sparsity in recommender systems}, volume = {636}, year = {2025}, }
• (2025). Neural learning rules from associative networks theory. NEUCOM, 636, 129865. (www)

  @article{ , abstract = {Associative networks theory is increasingly providing tools to interpret update rules of artificial neural networks. At the same time, deriving neural learning rules from a solid theory remains a fundamental challenge. We make some steps in this direction by considering general energy-based associative networks of continuous neurons and synapses that evolve in multiple time scales. We use the separation of these timescales to recover a limit in which the activation of the neurons, the energy of the system and the neural dynamics can all be recovered from a generating function. By allowing the generating function to depend on memories, we recover the conventional Hebbian modeling choice for the interaction strength between neurons. Finally, we propose and discuss a dynamics of memories that enables us to include learning in this framework.}, archive = {J_NEUCOM}, author = {Daniele Lotito}, doi = {10.1016/j.neucom.2025.129865}, journal = {Neurocomputing}, month = {7}, pages = {129865}, shortjournal = {Neurocomputing}, title = {Neural learning rules from associative networks theory}, volume = {636}, year = {2025}, }
• (2025). Adaptive temperature distillation method for mining hard samples’ knowledge. NEUCOM, 636, 129745. (www)

  @article{ , abstract = {Knowledge distillation can transfer knowledge from a complex teacher network into a simple student one through a high temperature factor, improving the latter’s performance. However, existing studies usually use fixed temperatures, making them ineffective in mining the rich knowledge contained in hard samples. Specifically, high temperature tends to over-smooth knowledge on hard samples, whereas low temperature makes knowledge almost equivalent to hard labels on easy samples. In this paper, we propose an Adaptive Temperature Distillation (ATD) method to effectively address these challenges. A well-trained teacher network’s information entropy is used to assess a sample’s relative difficulty. Then, low temperature is used in a hard sample, which allows the student network to learn its dark knowledge more effectively. And high temperature is employed in an easy sample to prevent the student network from becoming overconfident and ignoring the dark knowledge of negative classes. Furthermore, we propose a mixup variant to enable the student network to access more hard samples with rich dark knowledge. Instead of focusing on data augmentation as the existing mixup studies, ATD pays attention to increasing the richness of dark knowledge by mixing the output logits of easy and hard samples. The overall performance of ATD is verified in multiple benchmark datasets by comparing it with state-of-the-art knowledge distillation methods.}, archive = {J_NEUCOM}, author = {Shunzhi Yang and Xiong Yang and Jin Ren and Liuchi Xu and Jinfeng Yang and Zhenhua Huang and Zheng Gong and Wenguang Wang}, doi = {10.1016/j.neucom.2025.129745}, journal = {Neurocomputing}, month = {7}, pages = {129745}, shortjournal = {Neurocomputing}, title = {Adaptive temperature distillation method for mining hard samples’ knowledge}, volume = {636}, year = {2025}, }
• (2025). A reformulation neurodynamic algorithm for distributed nonconvex optimization. NEUCOM, 635, 130023. (www)

  @article{ , abstract = {This paper presents a reformulation neurodynamic algorithm for solving distributed nonconvex optimization problems. A class of general Lagrangian functions is introduced to eliminate the dual gap in nonconvex problems. This algorithm extends the application of neurodynamic algorithms based on the p -power reformulation transformation of Lagrangian functions. Under mild conditions, the initial point of the decision vector can be arbitrarily chosen. It is proven that the output trajectories will eventually converge to a strict local minimum point of the distributed nonconvex optimization problem. Finally, numerical experiments demonstrate the effectiveness of the proposed algorithm, which is also applied to solve the oblique throwing problem and the distributed source localization problem.}, archive = {J_NEUCOM}, author = {Xin Yu and Qingzhou Huang and Rixin Lin}, doi = {10.1016/j.neucom.2025.130023}, journal = {Neurocomputing}, month = {6}, pages = {130023}, shortjournal = {Neurocomputing}, title = {A reformulation neurodynamic algorithm for distributed nonconvex optimization}, volume = {635}, year = {2025}, }
• (2025). ASTR: Transformer-based alert-to-stage translator for multi-stage attack detection. NEUCOM, 635, 130016. (www)

  @article{ , abstract = {Multi-stage attacks have complex patterns and intricate inter-stage dependencies, making them difficult to detect using conventional methods. Traditional detection techniques struggle to capture long-term dependencies in alert sequences and often fail to map individual alerts to their specific attack stages. Aiming to develop an advanced detection method enhancing the accuracy and granularity of multi-stage attack detection, this paper proposes ASTR (Alert Stage TRanslator for MSA), a Transformer-based detection architecture. ASTR utilizes a Doc2Vec-based frontend to embed IDS alerts into numerical vectors and applies positional encoding to preserve sequential information. The core Transformer backbone, with its attention mechanism, is employed to extract deep contextual features and long-term dependencies from alert sequences. Finally, a refined backend maps these features to specific attack stages. ASTR is evaluated against state-of-the-art methods using three public datasets: DARPA2000, ISCXIDS2012, and CIC-IDS2017. Experimental results demonstrate that ASTR achieves detection accuracy of over 97% across all datasets. Compared to existing methods, ASTR not only improves precision, recall, and F1-scores but also provides detailed, one-to-one mappings from alerts to attack stages, thereby offering a more fine-grained detection capability.}, archive = {J_NEUCOM}, author = {Wei Ma and Yunyun Hou and Aina Sui and Pengpeng Jian}, doi = {10.1016/j.neucom.2025.130016}, journal = {Neurocomputing}, month = {6}, pages = {130016}, shortjournal = {Neurocomputing}, title = {ASTR: Transformer-based alert-to-stage translator for multi-stage attack detection}, volume = {635}, year = {2025}, }
• (2025). PRG-net: Point relationship-guided network for 3D human action recognition. NEUCOM, 635, 130015. (www)

  @article{ , abstract = {Point clouds contain rich spatial information, providing important supplementary clues for human action recognition. Recent methods for action recognition based on point cloud sequences primarily rely on complex spatiotemporal local encoding. However, these methods often utilize max-pooling operations to select features when extracting local features, restricting feature updates to local neighborhoods and failing to fully exploit the relationships between regions. Moreover, cross-frame encoding can also lead to the loss of spatiotemporal information. In this study, we propose PRG-Net, a Point Relation Guided Network, to further improve the learning of spatiotemporal features in point clouds. First, we designed two core modules: the Spatial Feature Aggregation (SFA) and the Spatial Feature Descriptor (SFD) modules. The SFA module expands the spatial structure between regions using dynamic aggregation techniques, while the SFD module guides the region aggregation process by Attention-Weighted Descriptors. They enhance the modeling of human spatial structure by expanding the relationships between regions. Second, we introduce inter-frame motion encoding techniques that can obtain the final spatiotemporal representation of the human body through the aggregation of cross-frame vectors, without relying on complex spatiotemporal local encoding. We evaluate PRG-Net on publicly available human action recognition datasets, including NTU RGB+D 60, NTU RGB+D 120, UTD-MHAD, and MSR Action 3D. Experimental results demonstrate that our method outperforms state-of-the-art point-based 3D action recognition methods significantly. Furthermore, we conduct extended experiments on the SHREC 2017 dataset for gesture recognition, and the results show that our method maintains competitive performance on that dataset as well.}, archive = {J_NEUCOM}, author = {Yao Du and Zhenjie Hou and En Lin and Xing Li and Jiuzhen Liang and Xinwen Zhou}, doi = {10.1016/j.neucom.2025.130015}, journal = {Neurocomputing}, month = {6}, pages = {130015}, shortjournal = {Neurocomputing}, title = {PRG-net: Point relationship-guided network for 3D human action recognition}, volume = {635}, year = {2025}, }
• (2025). Deep consistent penalizing hashing with noise-robust representation for large-scale image retrieval. NEUCOM, 635, 130014. (www)

  @article{ , abstract = {Benefiting from the powerful representational capacity of deep learning and the attractive computational efficiency of binary codes, deep hashing frameworks have made large progress for large-scale image retrieval applications. By calculating the priori labels, most existing deep supervised hashing usually introduces the effective margin-based objective loss to generate label-level penalizing boundaries for training samples during the model optimization. However, the decision boundaries from label-level penalizing may be inconsistent with semantic relations hidden in raw samples, compromising the performance. Besides, for classes with low intra-class variances or inter-class correlations, the force field of the margin-based methods might be too weak to learn the discriminant embedding space. In this paper, we solve this dilemma with a novel unified deep hashing framework, termed Deep Consistent Penalizing Hashing with noise-robust representation (DCPH), to generate compact yet discriminative binary codes for efficient and accurate image retrieval. Specifically, by learning the unbalanced correlations of training samples, the semantic consistency penalizing loss, consisting of pulling penalizing elements and pushing penalizing elements, is proposed to generate the semantic decision boundaries across classes. For parameter optimization, the dice-like optimization strategy is introduced to balance the pulling and pushing field, facilitating the generation of highly separable Hamming space. Besides, to mitigate the negative influence caused by objective-unrelated information or noise, by introducing patch-wise attention strategy and depth-wise convolution operation, the noise-robust representation module is developed to capture the robust feature descriptor with abundant fine-grained information. Comprehensive evaluations are performed on several benchmark datasets, and the experimental results consistently validate the effectiveness of our proposed DCPH framework, which significantly outperforms the state-of-the-art deep hashing methods.}, archive = {J_NEUCOM}, author = {Qibing Qin and Hong Wang and Wenfeng Zhang and Lei Huang and Jie Nie}, doi = {10.1016/j.neucom.2025.130014}, journal = {Neurocomputing}, month = {6}, pages = {130014}, shortjournal = {Neurocomputing}, title = {Deep consistent penalizing hashing with noise-robust representation for large-scale image retrieval}, volume = {635}, year = {2025}, }
• (2025). Domain-wise knowledge decoupling for personalized federated learning via radon transform. NEUCOM, 635, 130013. (www)

  @article{ , abstract = {Personalized federated learning (pFL) customizes local models to address heterogeneous data across clients. One prominent research direction in pFL is model decoupling, where the knowledge of a global model is selectively utilized to assist local model personalization. Prior studies primarily use decoupled global-model parameters to convey this selected knowledge. However, due to the task-related knowledge-mixing nature of deep learning models, using these parameters may introduce irrelevant knowledge to specific clients, impeding personalization. To address this, we propose a domain-wise knowledge decoupling approach (pFedDKD), which decouples global-model knowledge into diverse projection segments in the representation space, meeting the specific needs of clients on heterogeneous local domains. A Radon transform-based method is provided to facilitate this decoupling, enabling clients to extract relevant knowledge segments for personalization. Besides, we provide a distillation-based back-projection learning method to fuse local-model knowledge into the global model, ensuring the updated global-model knowledge remains decouplable by projection. A theoretical analysis confirms that our approach improves generalization. Extensive experiments on four datasets demonstrate that pFedDKD consistently outperforms eleven state-of-the-art baselines, achieving an average improvement of 1.21% in test accuracy over the best-performing baseline.}, archive = {J_NEUCOM}, author = {Zihao Lu and Junli Wang and Changjun Jiang}, doi = {10.1016/j.neucom.2025.130013}, journal = {Neurocomputing}, month = {6}, pages = {130013}, shortjournal = {Neurocomputing}, title = {Domain-wise knowledge decoupling for personalized federated learning via radon transform}, volume = {635}, year = {2025}, }
• (2025). Multiobjective backbone network architecture search based on transfer learning in steel defect detection. NEUCOM, 635, 130012. (www)

  @article{ , abstract = {In recent years, steel defect detection methods based on deep learning have been widely used. However, due to the shape specificity of steel defects and data scarcity, using existing convolutional neural network architectures for training requires significant expertise and time to fine-tune the hyperparameters. Transfer learning effectively tackles the challenges of data scarcity or limited computing resources by transferring domain knowledge from source tasks to related target tasks, reducing the resource consumption of model training from scratch. In this paper, we propose a transfer learning-based multiobjective backbone network architecture search method (TMBNAS). First, TMBNAS formulates defect detection network search as a multiobjective problem while optimizing its detection accuracy and model complexity. Second, an effective variable-length encoding strategy is designed to represent different building blocks and unpredictable optimal depths in convolutional neural networks, and targeted improvements are made to the crossover and mutation operators. For the specificity of the steel defect detection task, a transfer learning strategy based on similar knowledge is used to transfer the architecture and weight parameters obtained from the search in the source task to the target task, and adjust and optimize them. Finally, a dynamic adjustment mechanism based on actual constraints is designed during the search process to gradually approximate the optimal non-dominated solution set with higher detection accuracy without losing its population diversity. The proposed method is tested on the continuous casting slab and workpiece defect datasets. The experimental results show that the model searched by the proposed method can achieve better detection performance compared with manually designed deep learning algorithms and classical network architecture search methods.}, archive = {J_NEUCOM}, author = {Tianchen Zhao and Xianpeng Wang and Xiangman Song}, doi = {10.1016/j.neucom.2025.130012}, journal = {Neurocomputing}, month = {6}, pages = {130012}, shortjournal = {Neurocomputing}, title = {Multiobjective backbone network architecture search based on transfer learning in steel defect detection}, volume = {635}, year = {2025}, }
• (2025). Pool-mamba: Pooling state space model for low-light image enhancement. NEUCOM, 635, 130005. (www)

  @article{ , abstract = {Mamba, with its advantages in long-distance modeling and computational efficiency, has been promptly applied in low-light image enhancement (LLIE). Nevertheless, Mamba faces two key issues when processing low-light images: (1) overexposure in scenes with uneven illumination due to the lack of multi-scale modeling; (2) insufficient local detail recovery, as its sequential operation weakens the perception of local lighting changes. To alleviate these problems, we propose a Pooling State Space Model (Pool-Mamba) by integrating the state space model with pooling techniques. First, we devise the Pyramid-pooling Mamba (PMamba) module, which leverages pyramid pooling to capture multi-scale information, effectively mitigating uneven exposure under varying lighting conditions. Next, the Axis-pooling Mamba (AMamba) module is proposed to model local correlations along specific spatial dimensions (height and width), generating more refined local representations and enhancing the model’s ability to adapt to local lighting variations and intricate details. Finally, we incorporate a Dual Gated Enhancement Module (DGEM) to strengthen the channel correlations between PMamba and AMamba, facilitating the integration of multi-scale and local features. Benchmark assessments demonstrate that Pool-Mamba surpasses current state-of-the-art (SOTA) methods, achieving superior quantitative evaluations and less distorted visual results.}, archive = {J_NEUCOM}, author = {Qiao Zhang and Mingwen Shao and Xinyuan Chen}, doi = {10.1016/j.neucom.2025.130005}, journal = {Neurocomputing}, month = {6}, pages = {130005}, shortjournal = {Neurocomputing}, title = {Pool-mamba: Pooling state space model for low-light image enhancement}, volume = {635}, year = {2025}, }
• (2025). MKGF: A multi-modal knowledge graph based RAG framework to enhance LVLMs for medical visual question answering. NEUCOM, 635, 129999. (www)

  @article{ , abstract = {Medical visual question answering (MedVQA) is a challenging task that requires models to understand medical images and return accurate responses for the given questions. Most recent methods focus on transferring general-domain large vision–language models (LVLMs) to the medical domain by constructing medical instruction datasets and in-context learning. However, the performance of these methods are limited due to the hallucination issue of LVLMs. In addition, fine-tuning the abundant parameters of LVLMs on medical instruction datasets is high time and economic cost. Hence, we propose a MKGF framework that leverages a multi-modal medical knowledge graph (MMKG) to relieve the hallucination issue without fine-tuning the abundant parameters of LVLMs. Firstly, we employ a pre-trained text retriever to build question–knowledge relations on training set. Secondly, we train a multi-modal retriever with these relations. Finally, we use it to retrieve question-relevant knowledge and enhance the performance of LVLMs on the test set. To evaluate the effectiveness of MKGF, we conduct extensive experiments on two public datasets Slake and VQA-RAD. Our method improves the pre-trained SOTA LVLMs by 10.15% and 9.32%, respectively. The source codes are available at https://github.com/ehnal/MKGF .}, archive = {J_NEUCOM}, author = {Yinan Wu and Yuming Lu and Yan Zhou and Yifan Ding and Jingping Liu and Tong Ruan}, doi = {10.1016/j.neucom.2025.129999}, journal = {Neurocomputing}, month = {6}, pages = {129999}, shortjournal = {Neurocomputing}, title = {MKGF: A multi-modal knowledge graph based RAG framework to enhance LVLMs for medical visual question answering}, volume = {635}, year = {2025}, }
• (2025). Label self-correction intelligent diagnosis method and embedded system for axle box bearings of high-speed trains with noisy labels. NEUCOM, 635, 129998. (www)

  @article{ , abstract = {Due to annotation errors, delayed labeling, and noise interference, data label noise is a common issue in high-speed train datasets, leading to overfitting of existing intelligent diagnostic methods on noisy-label samples and a decline in the accuracy of fault diagnosis, which affects the correct assessment of high-speed train bearing health. To tackle this issue, this article presents an adaptive label self-correction intelligent diagnostic method. The method consists of three main parts: First, it employs dynamic thresholds and multi-network interactive training to separate clean from noisy labels. Second, it corrects noisy labels using classifiers trained on clean data, with two designed correction methods for high-accuracy label correction. Third, it retrains the model by reweighting loss to ensure that it fully captures information from noisy label data. Additionally, based on the proposed method, an AI microprocessor diagnosis system is developed for real-world health monitoring of axle box bearings. Both the method and the system have been validated through diagnostic cases of axle box bearings. Validation through diagnostic cases demonstrates that the method can train high-accuracy diagnostic models under label noise conditions and the system can rapidly diagnose data in real-time.}, archive = {J_NEUCOM}, author = {Yaning Li and Yang Gao and Bin Yang and Yaguo Lei and Xiang Li and Yue Shu and Ke Feng}, doi = {10.1016/j.neucom.2025.129998}, journal = {Neurocomputing}, month = {6}, pages = {129998}, shortjournal = {Neurocomputing}, title = {Label self-correction intelligent diagnosis method and embedded system for axle box bearings of high-speed trains with noisy labels}, volume = {635}, year = {2025}, }
• (2025). SR-CIBN: Semantic relationship-based consistency and inconsistency balancing network for multimodal fake news detection. NEUCOM, 635, 129997. (www)

  @article{ , abstract = {The fast dissemination of online information has facilitated the evolution of multimodal fake news, thereby rendering the trustworthiness of its content difficult to identify. Some existing studies, although capturing the consistency and inconsistency features between different modalities, neglect to dynamically balance these two types of features based on their contributions during the features fusion. Thus, we propose a S emantic R elationship-based C onsistency and I nconsistency B alancing N etwork for multimodal fake news detection (SR-CIBN). Specifically, the global features are thoroughly investigated by hierarchically penetrating and interacting between multimodal features that are aligned by contrastive learning at both the intra- and inter-modal views. Then, the global consistency and inconsistency features are obtained through the interaction between the selected key image patches features and the global features. Additionally, the fusion intensity of the global consistency and inconsistency features is adjusted based on the image–text matching degree, resulting in the final optimized features. Under the joint learning framework we proposed, confusion between semantically similar real and fake news is effectively avoided by training with triplet loss based on the image–text semantic relationship. Our model surpasses comparable approaches, as shown by comprehensive experiments on the Twitter and Weibo datasets.}, archive = {J_NEUCOM}, author = {Hongzhu Yu and Hongchen Wu and Xiaochang Fang and Meng Li and Huaxiang Zhang}, doi = {10.1016/j.neucom.2025.129997}, journal = {Neurocomputing}, month = {6}, pages = {129997}, shortjournal = {Neurocomputing}, title = {SR-CIBN: Semantic relationship-based consistency and inconsistency balancing network for multimodal fake news detection}, volume = {635}, year = {2025}, }
• (2025). Memory-augmented prototypical meta-learning method for bearing fault identification under few-sample conditions. NEUCOM, 635, 129996. (www)

  @article{ , abstract = {Recent developments in deep learning have enhanced the performance of rolling bearing fault diagnosis. However, two challenges persist: 1) These models typically require large amounts of labeled data, which limits their effectiveness when such data is scarce; 2) The decision-making process of these models is often opaque, making it difficult to understand the key factors they focus on. In this paper, the memory-augmented prototypical meta-learning (MAPML) is proposed for few-shot fault identification. In MAPML, the dynamic prototype adjustment module is proposed to dynamically and adaptively refine class prototypes through the memory-augmented mechanism, enhancing the model’s ability to accurately differentiate between classes. Additionally, the multi-scale convolutional architecture is employed to extract hidden features from signals at multiple scales in parallel, providing comprehensive feature representation across different scales. This proposed method is verified through two datasets of bearings, demonstrating robust generalization capabilities across different working conditions. By introducing Gaussian white noise to simulate real industrial environments, the robustness and practicality of MAPML in handling noisy data are further substantiated. The interpretability analysis indicates that MAPML effectively captures impact characteristics in the time domain, as well as resonance bands and fault characteristic frequencies in the frequency domain.}, archive = {J_NEUCOM}, author = {Xianze Li and Zhitai Xing and Ling Xiang and Yang Chen and Aijun Hu}, doi = {10.1016/j.neucom.2025.129996}, journal = {Neurocomputing}, month = {6}, pages = {129996}, shortjournal = {Neurocomputing}, title = {Memory-augmented prototypical meta-learning method for bearing fault identification under few-sample conditions}, volume = {635}, year = {2025}, }
• (2025). Few-shot SAR image classification via multiple prototypes ensemble. NEUCOM, 635, 129989. (www)

  @article{ , abstract = {Synthetic aperture radar (SAR) image classification has benefited significantly from deep learning techniques, which excel at automatically learning semantic features from data. However, compared with optical remote sensing images, SAR images face a more pronounced sample limitation problem due to sparsely distributed training samples and complex backgrounds. As a result, the performance of data-driven SAR image classification significantly degrades due to overfitting, especially when deep learning algorithms operate under data scarcity. To address these challenges, we present a novel few-shot SAR classification approach using a multiple-prototype ensemble network within the meta-learning paradigm. Specifically, to better capture the scattering feature distribution and enhance intra-category aggregation in SAR images, multiple prototypes are learned to fully leverage the semantic information embedded in the limited support samples. Furthermore, we introduce a subspace discriminative loss to improve the representational power of the learned prototypes by ensuring consistency in SAR feature representation while maintaining inter-class divergence. Extensive experiments conducted on three real-world datasets demonstrate the superiority of the proposed method compared to several state-of-the-art approaches.}, archive = {J_NEUCOM}, author = {Zhiqiang Zhao and Yuhui Tong and Meng Jia and Yuan Qiu and Xiaofan Wang and Xinhong Hei}, doi = {10.1016/j.neucom.2025.129989}, journal = {Neurocomputing}, month = {6}, pages = {129989}, shortjournal = {Neurocomputing}, title = {Few-shot SAR image classification via multiple prototypes ensemble}, volume = {635}, year = {2025}, }
• (2025). LSKN-MFIF: Large selective kernel network for multi-focus image fusion. NEUCOM, 635, 129984. (www)

  @article{ , abstract = {Multi-focus image fusion (MFIF) is an image enhancement technique that compensates for the limited depth of field of optical lenses. It has broad application prospects across multiple disciplines such as digital photography, biomedical imaging, and machine vision, similar to other advanced visual tasks. This study proposes a large selective kernel network, referred to as LSKN-MFIF, which is designed to improve multi-focus image fusion. This network has the capability to dynamically adjust its extensive spatial receptive field, effectively expanding the receptive field and capturing multi-scale global information, fully extracting and integrating multi-scale context to better identify various focused regions in the image. More specifically, LSKN-MFIF extracts multi-scale global features through a large selective kernel module (LKSB) that decomposes large kernel convolutions, utilizes spatial feature selection for feature aggregation, and then enhances representation capability by extracting local information through a gated differential convolution block (GDCB), thereby generating accurate decision maps. Experimental results show that the proposed methodology surpasses current multi-focus image fusion techniques in terms of both subjective visual quality and objective performance metrics.}, archive = {J_NEUCOM}, author = {Hao Zhai and Guochao Zhang and Zhi Zeng and Zhendong Xu and Aiqing Fang}, doi = {10.1016/j.neucom.2025.129984}, journal = {Neurocomputing}, month = {6}, pages = {129984}, shortjournal = {Neurocomputing}, title = {LSKN-MFIF: Large selective kernel network for multi-focus image fusion}, volume = {635}, year = {2025}, }
• (2025). Double-activation neural network for solving parabolic equations with time delay. NEUCOM, 635, 129978. (www)

  @article{ , abstract = {While physics-informed neural network (PINN) has shown promise in solving partial differential equations (PDEs), their performance is limited by the expressive capacity of standard neural network architectures, particularly when dealing with complex, nonlinear problems. To improve the nonlinear expressive capacity of the network and broaden the applications of PINN, we propose a double-activation neural network (DANN) for solving parabolic equations with time delay, where each neuron is equipped with two activation functions and a new parameter is introduced in one of the functions to form quadratic terms. To address the issue of low fitting accuracy caused by the discontinuity of solution’s derivative, a piecewise fitting approach is proposed by dividing the global solving domain into several subdomains according to the discontinuous points. The convergence of the loss function is proved. We conduct a series of numerical experiments to show the efficiency of the proposed DANN technique, including solving delay partial differential equations (DPDEs) with constant delay, time-dependent delay, and delay differential equations (DDEs) with state-dependent delay. The robustness of DANN is assessed by varying the number of training points, hidden layers, neurons per layer, and random seeds. We also evaluate the extended application of DANN by simulating a second-order rogue wave in the derivative nonlinear Schrödinger (DNLS) equation, to show its applicability beyond DPDEs.}, archive = {J_NEUCOM}, author = {Qiumei Huang and Qiao Zhu}, doi = {10.1016/j.neucom.2025.129978}, journal = {Neurocomputing}, month = {6}, pages = {129978}, shortjournal = {Neurocomputing}, title = {Double-activation neural network for solving parabolic equations with time delay}, volume = {635}, year = {2025}, }
• (2025). Energy and synchronization of multifunctional loop neural networks. NEUCOM, 635, 129973. (www)

  @article{ , abstract = {Multifunctional thermo-photoelectric neurons are newly proposed intelligent artificial neurons with expanded functionalities. They are capable of detecting and encoding multimodal thermo-photoelectric signals. This study constructs three loop neural networks by bridging photosensitive, thermosensitive, and thermo-photoelectric neurons with linear resistors, inductive coils, and memristors, respectively. These networks are designated as Voltage, Magnetic, and Memory Coupling Functional Networks (VCFN, MACFN, and MECFN). In the adaptive VCFN model, the system's diverse energy cannot be effectively balanced by voltage coupling, resulting in an inability to achieve a stable synchronization state. MACFN and MECFN models perform energy pumping and extraction among neurons, resulting in a self-organizing behavior. This behavior spontaneously adjusts the membrane sequences of neurons within the model, ultimately achieving a stable equilibrium of the system's intrinsic field energy. The acceleration of this process can be facilitated by an increase in the coupling gain ratio. The MECFN model quickly achieves phase synchronization, and due to its adaptive mechanism, the phase synchronization state is further stabilized. This indicates memristors can serve as bridges for communication and encoding between heterogeneous functional neural networks. The effects of chaotic currents on isolated neurons and models were also considered. High-intensity chaotic currents will cause resonance phenomena in neurons under single-peaking oscillation mode, exhibiting chaotic geometric characteristics but without changing their periodic characteristics. In neurons with multi-spiking oscillation modes, chaotic currents may induce extremely narrow chaotic windows. VCFN, MACFN, and MECFN models demonstrate good robustness in spike periodic modes but are more susceptible to external chaotic current interference in chaotic modes. Appropriate intensities of chaotic current stimulation can promote synchronization in models, while excessive intensities may suppress it. This research provides insights into the synchronization processes of multifunctional neural networks and the design of multimodal and high-robustness intelligent sensor systems, providing a theoretical foundation for advancements in artificial intelligence.}, archive = {J_NEUCOM}, author = {Zebang Cheng and Shu Zhou and Jiajun Jiang and Shunwei Yao and Lin Peng and Tingting Shi and Xiaolin Liu and Jia Lin}, doi = {10.1016/j.neucom.2025.129973}, journal = {Neurocomputing}, month = {6}, pages = {129973}, shortjournal = {Neurocomputing}, title = {Energy and synchronization of multifunctional loop neural networks}, volume = {635}, year = {2025}, }
• (2025). Student behavior detection model based on multilevel residual networks and hybrid attention mechanisms. NEUCOM, 635, 129965. (www)

  @article{ , abstract = {Accurately analyzing student behaviors allows for better evaluation of student engagement, which in turn can improve teaching quality. To address challenges such as multi-scale scenes, occluded targets, and subtle fine features in classroom environments, while also considering model implementability, we propose an efficient student behavior detection model, RSAY. This model leverages multi-scale information extraction and a hybrid attention mechanism to support teaching. Both the backbone and feature fusion networks of the model integrate our designed Rep_SC_Atten module, which incorporates our novel multi-level residual network architecture and a lightweight hybrid attention mechanism. This hybrid architecture enhances the model’s sensitivity and ability to extract multi-scale information, while ensuring effective extraction of fine-grained features via the attention mechanism. Additionally, the DDetect strategy is introduced in the detection head to reduce model size without sacrificing accuracy. We evaluated our model using the SCB-Dataset and a custom student behavior dataset, demonstrating a 6.3% improvement in accuracy over the baseline model.}, archive = {J_NEUCOM}, author = {Wenbin Lu and Songyan Liu and Boyang Ding and Peng Chen and Fangpeng Lu}, doi = {10.1016/j.neucom.2025.129965}, journal = {Neurocomputing}, month = {6}, pages = {129965}, shortjournal = {Neurocomputing}, title = {Student behavior detection model based on multilevel residual networks and hybrid attention mechanisms}, volume = {635}, year = {2025}, }
• (2025). Reliable ℓ2−ℓ∞ state estimation for delayed neural networks under weighted try-one-discard protocol. NEUCOM, 635, 129923. (www)

  @article{ , abstract = {In this paper, the problem of reliable ℓ 2 - ℓ ∞ state estimation is addressed for discrete-time artificial neural networks with switched time-delays under weighted try-once-discard (WTOD) protocol. To mitigate data congestion and transmission burdens, the WTOD protocol is implemented in the transmission channels to optimize data communication, where the transmission priority is dynamically determined based on mission importance. A Bernoulli-distributed stochastic variable with known statistical properties is introduced to model the switching behavior between the presence and absence of time-delays and a failure matrix is constructed to characterize potential failures affecting the received measurement data. The primary objective of this paper is to develop a state estimator that effectively performs the desired estimation task by thoroughly accounting for the combined effects of switched time-delays and the WTOD protocol. Specifically, by utilizing Lyapunov theory and matrix inequality techniques, the estimator parameters are meticulously derived to ensure exponentially mean-square stability and ℓ 2 - ℓ ∞ performance. Finally, the efficacy and validity of the proposed algorithm are demonstrated through an illustrative example.}, archive = {J_NEUCOM}, author = {Yuqiang Luo and Siyu Guo and Di Zhao and Hong Lin}, doi = {10.1016/j.neucom.2025.129923}, journal = {Neurocomputing}, month = {6}, pages = {129923}, shortjournal = {Neurocomputing}, title = {Reliable ℓ2−ℓ∞ state estimation for delayed neural networks under weighted try-one-discard protocol}, volume = {635}, year = {2025}, }
• (2025). Self-paced learning for anchor-based multi-view clustering: A progressive approach. NEUCOM, 635, 129921. (www)

  @article{ , abstract = {In the field of multi-view clustering (MVC), the surge in data has led to a significant increase in both the number of samples and the complexity of feature spaces, posing considerable challenges to the domain. Traditional anchor-based clustering methods effectively reduce the time and space complexity of algorithms by selecting representative samples to reconstruct similarity matrices. However, these methods are susceptible to the influence of low-quality anchors. To address this issue, we propose a novel self-paced learning for anchor-based MVC method, termed MSPA. This approach begins by constructing an anchor alternative pool, a novel strategy for selecting anchors that captures both intra-view and inter-view structural information. Subsequently, the concept of self-paced learning (SPL) is employed to progressively integrate anchors of varying quality into the model learning process, thereby constructing an anchor graph. Finally, the K-Means algorithm is applied to the resulting feature matrix to infer the final clustering results. Comprehensive comparative analyses conducted on eight benchmark datasets demonstrate that our proposed method outperforms existing state-of-the-art MVC algorithms in terms of efficiency.}, archive = {J_NEUCOM}, author = {Xia Ji and Xinran Cheng and Peng Zhou}, doi = {10.1016/j.neucom.2025.129921}, journal = {Neurocomputing}, month = {6}, pages = {129921}, shortjournal = {Neurocomputing}, title = {Self-paced learning for anchor-based multi-view clustering: A progressive approach}, volume = {635}, year = {2025}, }
• (2025). LGKGR: A knowledge graph reasoning model using LLMs augmented GNNs. NEUCOM, 635, 129919. (www)

  @article{ , abstract = {Knowledge graph reasoning (KGR) aims to infer new factual knowledge based on existing structured factual data, and plays a vital role in various applications. Graph neural networks (GNNs)-based methods have garnered attention due to their exceptional capabilities in learning graph structures. However, they cannot effectively leverage rich text semantics within KG for reasoning. Given the remarkable semantic understanding capabilities of large language models (LLMs), this paper proposes a novel KGR model using LLMs augmented GNNs (LGKGR), which aims to utilize LLMs to enhance the graph structure learning of GNNs. Each round of reasoning includes three stages: path search, path pruning, and path decision. The first stage adopts an incremental path search strategy to identify adjacent entities of current query entity and extract features. The second stage adopts GNNs as a pruning tool to filter out semantically irrelevant reasoning paths. The third stage exploits LLMs for semantic analysis of candidate reasoning paths, and then selects the most possible reasoning paths. In the end, LLMs are further exploited to analyze the semantic information of reasoning paths and generate final reasoning results. Experimental results on public datasets demonstrate that the proposed method achieves an average improvement of 2.1% in the MRR metric and 2.68% in the Hits@1 metric compared to existing SOTA methods. Explainable reasoning justifications are also generated during the reasoning process.}, archive = {J_NEUCOM}, author = {Yuanming Zhang and Wenbo Zheng and Jiacheng Huang and Gang Xiao}, doi = {10.1016/j.neucom.2025.129919}, journal = {Neurocomputing}, month = {6}, pages = {129919}, shortjournal = {Neurocomputing}, title = {LGKGR: A knowledge graph reasoning model using LLMs augmented GNNs}, volume = {635}, year = {2025}, }
• (2025). SARFormer: Segmenting anything guided transformer for semantic segmentation. NEUCOM, 635, 129915. (www)

  @article{ , abstract = {Semantic segmentation plays a crucial role in robotic systems. Despite advances, we find that current state-of-the-art methods are hard to apply in practice due to their weak generalization ability. Especially, diffusion-based segmentation methods struggle with over-reliance on noisy Ground Truth (GT) annotations, which are corrupted with noise and directly fed into the model’s forward propagation process during training, limiting the model’s ability to generalize. While the Segment Anything Model (SAM) excels at instance segmentation, it faces challenges in controlling granularity and lacks semantic information. To address these issues, we propose SARFormer, a semantic segmentation algorithm guided by SAM. Unlike conventional methods, SARFormer uses GT solely for supervision and replaces noisy GT with SAM guidance, enabling better generalization. The key innovations include a region-based SAM optimizer to refine granularity and a feature aggregation method for enhanced deep feature extraction. Experimental results show SARFormer achieves competitive accuracy, demonstrating the effectiveness of SAM in improving segmentation performance}, archive = {J_NEUCOM}, author = {Lixin Zhang and Wenteng Huang and Bin Fan}, doi = {10.1016/j.neucom.2025.129915}, journal = {Neurocomputing}, month = {6}, pages = {129915}, shortjournal = {Neurocomputing}, title = {SARFormer: Segmenting anything guided transformer for semantic segmentation}, volume = {635}, year = {2025}, }
• (2025). DACFusion: Dual asymmetric cross-attention guided feature fusion for multispectral object detection. NEUCOM, 635, 129913. (www)

  @article{ , abstract = {Effective fusion of unique features from different spectra plays a crucial role in multispectral object detection. Recent research has focused on transplanting advanced methods from other multimodal fusion fields to multispectral object detection tasks. These fusion methods focus on the fusion of features and ignore the spatial correspondence between multispectral images. This lack of correspondence in turn limits the full utilization of the complementarities between the different modalities, which affects the accuracy of object detection. To address this problem, we creatively propose a dual asymmetric cross-attention multispectral fusion (DACFusion) method, which is able to process features interactively based on the positional correspondence between two spectra, and then asymmetrically fuses the multispectral data according to the characteristics of each spectrum to take advantage of their complementary strengths. Meanwhile, we introduce a large selective kernel network to expand the receptive field for object detection, which further improves the detection accuracy. Experimental results on the VEDAI and LLVIP datasets validate the significant performance advantages of our proposed method and show its applicability to a variety of practical application scenarios. Code will be available at https://github.com/wood-fish/DACFusion .}, archive = {J_NEUCOM}, author = {Jingchen Qian and Baiyou Qiao and Yuekai Zhang and Tongyan Liu and Shuo Wang and Gang Wu and Donghong Han}, doi = {10.1016/j.neucom.2025.129913}, journal = {Neurocomputing}, month = {6}, pages = {129913}, shortjournal = {Neurocomputing}, title = {DACFusion: Dual asymmetric cross-attention guided feature fusion for multispectral object detection}, volume = {635}, year = {2025}, }
• (2025). From separation to fusion: Screening-assisted bilevel collaborative evolutionary optimization for railway freight allocation. NEUCOM, 635, 129910. (www)

  @article{ , abstract = {Efficient freight space allocation and stowage planning are critical for optimizing transportation efficiency and minimizing operational costs in railway transportation systems of large-scale enterprises. Traditional methods typically handle freight space allocation and stowage decisions in isolation or by simply layering these processes, leading to suboptimal results in terms of transportation efficiency and operational costs. To address this issue, this paper proposes a novel screening-assisted bilevel collaborative evolutionary optimization (Sa-BCEO) algorithm to explore and fusion the interdependencies between freight space allocation and stowage problems, thereby improving transportation efficiency. First, a screening-assisted mechanism (SAM) is designed to alleviate the complexity of the nested structure of bilevel optimization. This mechanism narrows the search space by retaining individuals with higher potential in the upper-level optimization, thereby enhancing efficiency in solving the lower-level optimization problem. Then, a bilevel framework is constructed to optimize the freight allocation and stowage. The effectiveness of the Sa-BCEO algorithm is validated through extensive experiments on a real-world enterprise dataset and two random datasets. Extensive results demonstrate significant improvements in transportation efficiency and cost reduction compared to traditional optimization methods.}, archive = {J_NEUCOM}, author = {Yiyin Tang and Yalin Wang and Chenliang Liu and Yong Wang and Weihua Gui}, doi = {10.1016/j.neucom.2025.129910}, journal = {Neurocomputing}, month = {6}, pages = {129910}, shortjournal = {Neurocomputing}, title = {From separation to fusion: Screening-assisted bilevel collaborative evolutionary optimization for railway freight allocation}, volume = {635}, year = {2025}, }
• (2025). RA2SP: Transform readability assessment to sequence prediction with a loss function incorporating penalty mechanism. NEUCOM, 635, 129909. (www)

  @article{ , abstract = {In the field of text readability assessment, traditional methods often rely on complex linguistic feature engineering, which increases the complexity of feature extraction and analysis while not necessarily improving model performance. To address these challenges, this work proposes a novel sequence prediction framework named RA2SP, which transforms the text readability classification task into a sequence prediction task. By introducing a penalty mechanism in the loss function, RA2SP reduces the reliance on burdensome linguistic feature engineering. The framework solely depends on deep features extracted from large pre-trained language models as input. Experimental results across various public datasets demonstrate that RA2SP achieves state-of-the-art or competitive performance across various evaluation metrics compared to existing baseline models. Additionally, this work provides a theoretical analysis of the role of the proposed penalty mechanism during model training, offering theoretical support and validating its effectiveness through experimental results. Finally, the potential applicability and superiority of the RA2SP framework are showcased in other text classification tasks, highlighting its adaptability and scalability.}, archive = {J_NEUCOM}, author = {Yijun Chen and Yurui Zheng and Jianhui Xu and Shaohong Zhang}, doi = {10.1016/j.neucom.2025.129909}, journal = {Neurocomputing}, month = {6}, pages = {129909}, shortjournal = {Neurocomputing}, title = {RA2SP: Transform readability assessment to sequence prediction with a loss function incorporating penalty mechanism}, volume = {635}, year = {2025}, }
• (2025). Handling language prior and compositional reasoning issues in visual question answering system. NEUCOM, 635, 129906. (www)

  @article{ , abstract = {Visual Question Answering (VQA) models often suffer from language bias, favoring common but incorrect answers, and struggle with compositional reasoning in complex queries. This paper proposes a unified approach using a multimodal large language model enhanced with adaptive prompts designed for specific tasks. Our method directly addresses these issues by reducing language bias and improving compositional reasoning. Extensive evaluations on benchmark datasets, including VQA v2.0, VQACP, TDIUC, GQA, Visual7 W, TextVQA, and STVQA show that our approach outperforms state-of-the-art models, achieving accuracy improvements of 8% to 9%. These results demonstrate the effectiveness of our method in enhancing VQA accuracy, making it a significant advancement for more reliable and robust applications in real-world scenarios.}, archive = {J_NEUCOM}, author = {Souvik Chowdhury and Badal Soni}, doi = {10.1016/j.neucom.2025.129906}, journal = {Neurocomputing}, month = {6}, pages = {129906}, shortjournal = {Neurocomputing}, title = {Handling language prior and compositional reasoning issues in visual question answering system}, volume = {635}, year = {2025}, }
• (2025). Event-based optical flow: Method categorisation and review of techniques that leverage deep learning. NEUCOM, 635, 129899. (www)

  @article{ , abstract = {Developing new convolutional neural network architectures and event-based camera representations could play a crucial role in autonomous navigation, pose estimation, and visual odometry applications. This study explores the potential of event cameras in optical flow estimation using convolutional neural networks. We provide a detailed description of the principles of operation and the software available for extracting and processing information from event cameras, along with the various event representation methods offered by this technology. Likewise, we identify four method categories to estimate optical flow using event cameras: gradient-based, frequency-based, correlation-based and neural network models. We report on these categories, including their latest developments, current status and challenges. We provide information on existing datasets and identify the appropriate dataset to evaluate deep learning-based optical flow estimation methods. We evaluate the accuracy of the implemented methods using the average endpoint error metric; meanwhile, the efficiency of the algorithms is evaluated as a function of execution time. Finally, we discuss research directions that promise future advances in this field.}, archive = {J_NEUCOM}, author = {Robert Guamán-Rivera and Jose Delpiano and Rodrigo Verschae}, doi = {10.1016/j.neucom.2025.129899}, journal = {Neurocomputing}, month = {6}, pages = {129899}, shortjournal = {Neurocomputing}, title = {Event-based optical flow: Method categorisation and review of techniques that leverage deep learning}, volume = {635}, year = {2025}, }
• (2025). Multi-level sparse network lasso: Locally sparse learning with flexible sample clusters. NEUCOM, 635, 129898. (www)

  @article{ , abstract = {Traditional learning usually assumes that all samples share the same global model, which fails to preserve critical local information for heterogeneous data. It can be tackled by detecting sample clusters and learning sample-specific models but is limited to sample-level clustering and sample-specific feature selection. In this paper, we propose multi-level sparse network lasso (MSN Lasso) for flexible local learning. It multiplicatively decomposes model parameters into two components: One component is for coarse-grained group-level, and another is for fine-grained entry-level. At the clustering stage, MSN Lasso simultaneously groups samples (group-level) and clusters specific features across samples (entry-level). At the feature selection stage, it enables both across-sample (group-level) and sample-specific (entry-level) feature selection. Theoretical analysis reveals a potential equivalence to a jointly regularized local model, which informs the development of an efficient algorithm. A divide-and-conquer optimization strategy is further introduced to enhance the algorithm’s efficiency. Extensive experiments across diverse datasets demonstrate that MSN Lasso outperforms existing methods and exhibits greater flexibility.}, archive = {J_NEUCOM}, author = {Luhuan Fei and Xinyi Wang and Jiankun Wang and Lu Sun and Yuyao Zhang}, doi = {10.1016/j.neucom.2025.129898}, journal = {Neurocomputing}, month = {6}, pages = {129898}, shortjournal = {Neurocomputing}, title = {Multi-level sparse network lasso: Locally sparse learning with flexible sample clusters}, volume = {635}, year = {2025}, }
• (2025). Advancements in few-shot nested named entity recognition: The efficacy of meta-learning convolutional approaches. NEUCOM, 635, 129893. (www)

  @article{ , abstract = {Few-shot Named Entity Recognition (NER) involves the identification of new entities using a limited amount of labeled data, which may contain nested entities. Currently, mainstream few-shot NER methods are not designed to handle nested entities. This study introduces a novel span-based meta-learning framework that uses meta-learning convolution to address the challenges of few-shot nested NER. Our proposed method, called M eta-Learning C onvolution for F ew- S hot N ested NER ( MCFSN ), is the first to integrate meta-learning with convolutional neural networks, effectively handling nested entities with limited training examples. This study presents a two-stage processing approach: extracting span features using CNN combined with the Biaffine attention mechanism, followed by entity span classification utilizing ProtoNet and the Biaffine classifier. Our experiments demonstrate consistently superior performance across three diverse language datasets, outperforming several competing baseline models in terms of F1 scores. Specifically, our approach achieves 6.9% F1 score improvement on the Genia, 5.2% F1 value improvement on the GermEval, and 4.5% F1 value enhancement on the NEREL, thus validating the effectiveness of our proposed approach.}, archive = {J_NEUCOM}, author = {Shuaichen Zhu and Yang Yuan and Lin Shi and Shoukun Xu}, doi = {10.1016/j.neucom.2025.129893}, journal = {Neurocomputing}, month = {6}, pages = {129893}, shortjournal = {Neurocomputing}, title = {Advancements in few-shot nested named entity recognition: The efficacy of meta-learning convolutional approaches}, volume = {635}, year = {2025}, }
• (2025). DCMVC: Dual contrastive multi-view clustering. NEUCOM, 635, 129889. (www)

  @article{ , abstract = {Multi-view clustering, which aims to divide data into different categories that are unsupervised in respect to information from different views, plays an important role in the field of computer vision. Contrastive learning is widely used in deep multi-view clustering methods to learn more discriminative representations. However, most existing multi-view clustering methods based on contrastive learning use only a single positive sample and do not fully utilize the category information in the learning process. To address the above issues, we propose a novel dual contrastive multi-view clustering (DCMVC) method, which uses pseudo-labels to refine the embedded features to make them more suitable for clustering tasks. Specifically, an inter-view correlation contrastive module is designed to learn more compact clustering assignments through a shared clustering prediction layer. Then, on the basis of the clustering predictions, we propose an intra-view consistency contrastive module, which dynamically selects the samples with the same pseudo label as positive samples and sets the other samples as negative samples to construct contrastive learning. The proposed model can alleviate the constraints of a single positive sample on contrastive learning by fully considering the latent category information to regularize the representation structure. Extensive experiments conducted on nine real datasets demonstrate the superiority of the proposed method.}, archive = {J_NEUCOM}, author = {Pengyuan Li and Dongxia Chang and Zisen Kong and Yiming Wang and Yao Zhao}, doi = {10.1016/j.neucom.2025.129889}, journal = {Neurocomputing}, month = {6}, pages = {129889}, shortjournal = {Neurocomputing}, title = {DCMVC: Dual contrastive multi-view clustering}, volume = {635}, year = {2025}, }
• (2025). MAD-DGTD: Multivariate time series anomaly detection based on dynamic graph structure learning with time delay. NEUCOM, 635, 129887. (www)

  @article{ , abstract = {Anomaly detection of multivariate time series data is extremely important in the industrial operation maintenance of Internet of Things (IoT). Researchers have found that the relationship between multiple sensors can be modeled as graph structure, and most researchers expresses this relationship by learning static graph structures which only contains the information of single modal. However, in actual IoT, the relationship between sensors will change with the changes of operating conditions, and this fixed graph structure cannot capture the relationship between sensors when working mode changes. To compensate the shortage of static graph, we propose a Multivariate time series Anomaly Detection framework based on Dynamic Graph learning with Time Delay (MAD-DGTD). Firstly, time-delay dynamic graph learning module (TDDG) is proposed to learn the changed mutual relationship between sensors over time and model it as a dynamic graph structure. In TDDG, a delay impact learning mechanism was designed to reconfigure the similarity calculation of node embeddings, which is designed to handle the temporal asynchrony of interactions between sensors in IoT. Secondly, we designed a stacked time dimension information extraction module (TDIE) and graph convolution information propagation module (GCIP) to capture information of different fine-grained sizes through multi-scale feature extraction. Finally, experimental research on three real-world datasets shows that our method outperforms the existing 10 competitive baselines in terms of overall performance.}, archive = {J_NEUCOM}, author = {Kang Wang and Jun Kong and Meicheng Zhang and Min Jiang and Tianshan Liu}, doi = {10.1016/j.neucom.2025.129887}, journal = {Neurocomputing}, month = {6}, pages = {129887}, shortjournal = {Neurocomputing}, title = {MAD-DGTD: Multivariate time series anomaly detection based on dynamic graph structure learning with time delay}, volume = {635}, year = {2025}, }
• (2025). Joint subspace learning and subspace clustering based unsupervised feature selection. NEUCOM, 635, 129885. (www)

  @article{ , abstract = {Unsupervised feature selection (UFS) has become a focal point of extensive research due to its ability to reduce the dimensionality of unlabeled data. Currently, many UFS methods based on subspace learning embed multiple graph regularization terms to preserve the local similarity structure of samples or features and rarely consider exploring global structure simultaneously, such as the self-representation structure between features and the potential clustering structure of samples. We propose a novel UFS model based on subspace learning and subspace orthogonal basis clustering (JSLSC) to address this problem. First, through robust subspace learning, JSLSC explores the self-representation information between the selected features and the original feature space. Features’ local and global structures are learned through feature selection and self-representation structure learning. Secondly, orthogonal basis clustering is introduced to learn the potential clustering structure in the low-dimensional sample space, thus enabling subspace clustering. Thirdly, hard-constrained graph structure learning is introduced to adaptively maintain the local structural consistency between low-dimensional samples and original samples. Finally, an optimization algorithm and convergence proof are proposed, and the superiority of the JSLSC is demonstrated through comparative experiments on nine real datasets.}, archive = {J_NEUCOM}, author = {Zijian Xiao and Hongmei Chen and Yong Mi and Chuan Luo and Shi-Jinn Horng and Tianrui Li}, doi = {10.1016/j.neucom.2025.129885}, journal = {Neurocomputing}, month = {6}, pages = {129885}, shortjournal = {Neurocomputing}, title = {Joint subspace learning and subspace clustering based unsupervised feature selection}, volume = {635}, year = {2025}, }
• (2025). The consistency analysis of gradient learning under independent covariate shift. NEUCOM, 635, 129883. (www)

  @article{ , abstract = {Many issues have drawn significant attention to gradient learning (GL), which seeks to approximate the gradient of the target function. Despite rapid progress, the existing methods on GL are almost based on the strict assumption that the samples are independent and identically distributed (i.i.d.) drawn. In this paper, we go beyond the classical i.i.d. framework and propose to investigate the GL under the independent covariate shift (i.c.s.) assumption. To be specific, we establish the upper bound of generalization error from the viewpoint of function approximation and show its theoretical consistency under a mild regularity condition on the bounded density-ratio, which generalizes the classical GL results under the i.i.d. framework. In addition, we have discovered a real-world example which meets the i.c.s. assumption. The numerical studies on the synthetic and real-world examples validate the effectiveness of proposed approach on the i.c.s. setting.}, archive = {J_NEUCOM}, author = {Liyuan Liu and Hong Chen and Chi Xiao and Weifu Li}, doi = {10.1016/j.neucom.2025.129883}, journal = {Neurocomputing}, month = {6}, pages = {129883}, shortjournal = {Neurocomputing}, title = {The consistency analysis of gradient learning under independent covariate shift}, volume = {635}, year = {2025}, }
• (2025). EKCA-cap: Improving dense captioning via external knowledge and context awareness. NEUCOM, 635, 129867. (www)

  @article{ , abstract = {Dense captioning is a image-to-text task that aims to locate the key semantic areas of an image and describe them in natural language. Previous researches have made great progresses, but this task remains challenging as it often generates ambiguous sentences due to weak perception of the attributes and semantic relationships of visual objects ( e.g. , holding something green in hands rather than cutting green vegetables). Furthermore, there is a lack of coherence between the sentences generated by each Region of Interest (RoI) and the context in the input image. Due to dense captioning being a multimodal task, the lack of interaction between text and visual modalities in the network can also lead to low accuracy. To tackle these challenges, we propose a dense captioning architecture with External Knowledge and Context Awareness, namely EKCA-Cap. Specifically, we construct a novel Common-Sense Knowledge Graph (CSKG) to provide attribute information of visual objects and semantic relationships with other visual cues. In addition, a Contextual Extractor (CE) is designed to extract contextual features of the target region from both spatial and semantic levels. Finally, a Dual-Stream Interactive Module (DSIM) is introduced to maximize the interaction between visual and text knowledge modalities. Extensive comparison and ablation experiments on Visual Genome (VG) and RefCOCOg datasets demonstrate the effectiveness of our method.}, archive = {J_NEUCOM}, author = {Zhenwei Zhu and Fengyu Zhou and Saike Huang and Teng Li}, doi = {10.1016/j.neucom.2025.129867}, journal = {Neurocomputing}, month = {6}, pages = {129867}, shortjournal = {Neurocomputing}, title = {EKCA-cap: Improving dense captioning via external knowledge and context awareness}, volume = {635}, year = {2025}, }
• (2025). GSNet: A new small object attention based deep classifier for presence of gun in complex scenes. NEUCOM, 635, 129855. (www)

  @article{ , abstract = {The motivation for focusing on weapon-based scene classification stems from the critical need to enhance public safety by enabling automated systems to quickly and accurately detect firearms in various environments. In contrast to common surveillance scenario classification based on intruders, weapon-based scenario classification often involves small weapons distributed throughout the scene or image. This requires more discriminative features and local semantics for effective classification. However, when deep convolutional neural networks (CNNs) are applied to scene classification, the loss of low- and mid-level features cannot be avoided. Furthermore, most existing networks tend to emphasize the global semantics of images. The low inter-class variability and high intra-class variability present specific challenges in weapon-based scene classification. To address these challenges, we propose a small object attention-based architecture in this work, with DenseNet serving as the backbone of our classification model. We modified the original DenseNet architecture to obtain more structured features. Additionally, we introduce a Small Object Attention (SAN) module after each dense block and an enhancement layer after each transition layer. Furthermore, we propose an enhanced classification layer in place of the traditional softmax layer, which helps retain relevant semantic features during classification. Consequently, the proposed classification model processes small patches of the image, preserving the relevant features of weapons. Experiments on six widely used benchmark datasets for weapon-based scenes demonstrate that our GSNet outperforms state-of-the-art methods by a significant margin while utilizing considerably fewer parameters. On average, the DenseNet model achieved an accuracy of 94.2%, whereas the proposed network attained an average accuracy of 98% across the six datasets.}, archive = {J_NEUCOM}, author = {Rajib Debnath and Kakali Das and Mrinal Kanti Bhowmik}, doi = {10.1016/j.neucom.2025.129855}, journal = {Neurocomputing}, month = {6}, pages = {129855}, shortjournal = {Neurocomputing}, title = {GSNet: A new small object attention based deep classifier for presence of gun in complex scenes}, volume = {635}, year = {2025}, }
• (2025). A survey on dynamic scene understanding using temporal knowledge graphs: From scene knowledge representation to extrapolation. NEUCOM, 635, 129854. (www)

  @article{ , abstract = {Dynamic scene understanding is the process of extracting information from video, identifying and inferring entities and relations within the scene, with the aim of thoroughly analyzing complex scenes that evolve over time. This process leverages temporal knowledge graphs to achieve a deep and comprehensive understanding of dynamic environments and is widely applied in areas such as autonomous driving, surveillance, and video analysis. Initially, scene knowledge representation is explored as the foundational step in dynamic scene understanding, achieved through the generation of temporal knowledge graphs. These graphs are categorized based on temporal granularity. Temporal knowledge graphs are divided into multiple-frame dynamic graphs and single-frame dynamic graphs. The generation methods for multiple-frame dynamic graphs are categorized into fragment-based and sliding-window approaches, while single-frame dynamic graphs primarily utilize transformer-based methods. This section provides an overview of the generation models for temporal knowledge graphs. Subsequently, dynamic scenes are further analyzed using extrapolation methods, which are classified into entity-based and relation-based modeling approaches. Entity-based modeling methods mainly include temporal point processes and graph neural network techniques, while relation-based modeling focuses on reinforcement learning and meta-learning techniques. This section summarizes various existing extrapolation techniques within these categories. Finally, the paper discusses the challenges associated with temporal knowledge graphs and explores potential research directions, offering insights into future advancements in dynamic scene understanding.}, archive = {J_NEUCOM}, author = {Linnan Lu and Guannan Si and Xinyu Liang and Mingshen Li and Fengyu Zhou}, doi = {10.1016/j.neucom.2025.129854}, journal = {Neurocomputing}, month = {6}, pages = {129854}, shortjournal = {Neurocomputing}, title = {A survey on dynamic scene understanding using temporal knowledge graphs: From scene knowledge representation to extrapolation}, volume = {635}, year = {2025}, }
• (2025). MFP-DETR: Marine UAV target detection based on multi-scale fuzzy perception. NEUCOM, 635, 129843. (www)

  @article{ , abstract = {An end-to-end object detection algorithm based on multi-scale fuzzy perception and feature enhancement, MFP-DETR, is proposed to solve the problem of image blur caused by the fast motion of Unmanned Aerial Vehicle (UAV) and pixel limitation of cameras in Unmanned Surface Vehicle (USV) images. Firstly, a learnable image mapping network — LPNet is designed to improve the representation of the target region and the quality of the input image. Secondly, the features of S3 and S4 and the features of S5 in the backbone Network were extracted, and they were input into the Squeeze and Excitation Network V2 (SENetV2) feature fusion module for feature fusion to enhance the detection ability of small targets. Finally, replace the Attention on the Intermediate Feature Interaction (AIFI) module in the Real-Time DEtection TRansformer (RT-DETR) with a Contextual Transformer(CoT) module. Context information is fully used to guide the learning of dynamic attention matrix and enhance the ability of global visual representation, to improve the accuracy of the model for different targets. The experimental results show that this method can enhance the fuzzy image effectively and detect the UAV accurately under background interference in the natural environment.}, archive = {J_NEUCOM}, author = {Ting Zou and Quanbo Ge and Yanjun Huang}, doi = {10.1016/j.neucom.2025.129843}, journal = {Neurocomputing}, month = {6}, pages = {129843}, shortjournal = {Neurocomputing}, title = {MFP-DETR: Marine UAV target detection based on multi-scale fuzzy perception}, volume = {635}, year = {2025}, }
• (2025). RagNet3D: Learning distinguishable representation for pooled grids in 3D object detection. NEUCOM, 635, 129841. (www)

  @article{ , abstract = {3D object detection plays a crucial role in autonomous driving. Previous two-stage 3D detectors have developed Grid-based Region-of-Interest (RoI) Pooling techniques, such as RoI-Grid Pooling and Voxel Pooling, both of which quantify RoI into grids. However, the grids are usually ambiguous since parts of them pertain to multiple RoIs that identify a single object. To address this issue, we propose a RoI-Aware Grids Pooling Network (RagNet3D), which introduces RoI-View context to create distinguishable grid representations. Specifically, we present a RoI-View Prediction module that predicts RoI-View context via the guidance of the distance between RoIs and objects. Meanwhile, we propose a Couple-View Fusion module that propagates the probabilistic distribution, calculated from the RoI-View context, into the RoI-irrelevant grid features for further box refinement. Extensive experiments on KITTI and Waymo Open Dataset show that our method achieves remarkable improvements against the baselines.}, archive = {J_NEUCOM}, author = {Jiaxin Chen and Yuehui Han and Zhiqiang Yan and Jianjun Qian and Jun Li and Jian Yang}, doi = {10.1016/j.neucom.2025.129841}, journal = {Neurocomputing}, month = {6}, pages = {129841}, shortjournal = {Neurocomputing}, title = {RagNet3D: Learning distinguishable representation for pooled grids in 3D object detection}, volume = {635}, year = {2025}, }
• (2025). Machine learning for drone detection from images: A review of techniques and challenges. NEUCOM, 635, 129823. (www)

  @article{ , abstract = {Unmanned Aerial Vehicles (UAVs), popularly known as drones, have revolutionized many sectors. In precision agriculture, they are used to effectively sprinkle water, fertilizers, and pesticides. In cinematography, UAVs snap aerial images that were difficult or impossible to obtain in the past. However, just as they can be used for good, they also have the potential for malicious uses. For example, drug smugglers use drones to evade border surveillance and push their goods across countries. Additionally, in the hands of militants, drones can be used to launch ballistics on targets, leading to the loss of lives and properties. Thus, researchers have recently focused on designing automated tools to detect friendly from unfriendly drones. One effective tool for such is Machine Learning (ML). This paper reviews works that use ML to detect drones from images. The images include visible light, infrared, and thermal. After studying the papers, we present the taxonomy and trends in the field. In addition, we also provide open research issues: the development of lightweight models, the use of synthetic data, the adoption of auto-annotation models, and the employment of transformer-based models.}, archive = {J_NEUCOM}, author = {Abubakar Bala and Ali H. Muqaibel and Naveed Iqbal and Mudassir Masood and Diego Oliva and Mujaheed Abdullahi}, doi = {10.1016/j.neucom.2025.129823}, journal = {Neurocomputing}, month = {6}, pages = {129823}, shortjournal = {Neurocomputing}, title = {Machine learning for drone detection from images: A review of techniques and challenges}, volume = {635}, year = {2025}, }
• (2025). A unified concept-based system for local, global, and misclassification explanations. NEUCOM, 635, 129761. (www)

  @article{ , abstract = {Explainability of Deep Neural Networks (DNNs) has garnered increasing attention in recent years. Of the various explainability approaches, concept-based techniques stand out for their ability to utilize human-meaningful concepts instead of focusing solely on individual pixels. However, there is a scarcity of methods that consistently provide both local and global explanations. Moreover, most of the methods have no offer to explain misclassification cases. Considering these challenges, we present a unified concept-based system for unsupervised learning of local and global concepts. Our primary objective is to uncover the intrinsic concepts underlying each data category by training a surrogate explainer network to estimate the concepts’ importance. Our experimental results substantiated the efficacy of the discovered concepts through diverse quantitative and qualitative assessments, encompassing faithfulness, completeness, and generality. Furthermore, our approach facilitates the explanation of both accurate and erroneous predictions, rendering it a valuable tool for comprehending the characteristics of the target objects of different classes.}, archive = {J_NEUCOM}, author = {Fatemeh Aghaeipoor and Dorsa Asgarian and Mohammad Sabokrou}, doi = {10.1016/j.neucom.2025.129761}, journal = {Neurocomputing}, month = {6}, pages = {129761}, shortjournal = {Neurocomputing}, title = {A unified concept-based system for local, global, and misclassification explanations}, volume = {635}, year = {2025}, }
• (2025). Neural headline generation: A comprehensive survey. NEUCOM, 635, 129633. (www)

  @article{ , abstract = {Automatic headline generation (HG) is an important natural language processing (NLP) task that aims to obtain a highly compressed text snippet from a document, to exhibit the core concept. Traditional headline generation (HG) techniques predominantly employ text summarization methods to generate short texts, by selecting important information from original documents. In recent years, with the rapid development of deep learning techniques, research on HG has leaned toward neural network-based end-to-end modeling approaches. Pretrained schemes and large language models (LLMs) demonstrate superior capability in generating natural language texts, thereby promoting further exploration on HG studies. However, a quality gap remains between machine-generated and human-written texts, making the generation of attractive and faithful headlines worthy of in-depth research. Therefore, this study presents a review of the most recent technologies on HG, including methods, datasets, and evaluation strategies. Future research directions are outlined, which provide a valuable reference point for HG and other summarization tasks. A collection of reference papers and code sources is available at: https://github.com/xiaona-chang/HGSurvey .}, archive = {J_NEUCOM}, author = {Han Ren and Xiaona Chang and Xia Li}, doi = {10.1016/j.neucom.2025.129633}, journal = {Neurocomputing}, month = {6}, pages = {129633}, shortjournal = {Neurocomputing}, title = {Neural headline generation: A comprehensive survey}, volume = {635}, year = {2025}, }
• (2025). Masked graph autoencoder-based multi-agent dynamic relational inference model for trajectory prediction. NEUCOM, 634, 129922. (www)

  @article{ , abstract = {Dynamic relational inference models uncover potential complex system interactions, enabling trajectory prediction and improving the interpretability of underlying system dynamics. However, the existing models cannot accurately infer the structural evolution trends and complete dynamic processes of temporal networks. Additionally, when uncertain noisy data are input, more serious graph noise problems, including redundant and noisy edges, occur, undermining the stability of interaction inference and reducing the accuracy of trajectory prediction. Therefore, a masked graph autoencoder-based multi-agent dynamic relational inference (MGAE-MDRI) trajectory prediction model is proposed herein. The mask reconstruction module is integrated into MDRI, where the partial edges of the interaction graph, representing multi-agent dynamic evolution, are masked through sampling. The reconstruction strategy leverages path and degree considerations to mitigate the impact of graph noise on the network topology. Furthermore, a graph attention network-based path sampler with a preference random walk is introduced, effectively combining network topology and node attribute features to construct a topologically weighted degree matrix and assign optimal mask sampling weights to neighboring nodes. Experiments conducted on four standard public datasets demonstrate that MGAE-MDRI outperforms the state-of-the-art models, achieving better trajectory prediction robustness and for complex multi-agent systems.}, archive = {J_NEUCOM}, author = {Fuyuan Zhao and Xiangang Cao and Jiangbin Zhao and Yong Duan and Xin Yang and Xinyuan Zhang}, doi = {10.1016/j.neucom.2025.129922}, journal = {Neurocomputing}, month = {6}, pages = {129922}, shortjournal = {Neurocomputing}, title = {Masked graph autoencoder-based multi-agent dynamic relational inference model for trajectory prediction}, volume = {634}, year = {2025}, }
• (2025). Brain state model: A novel method to represent the rhythmicity of object-specific selective attention from magnetoencephalography data. NEUCOM, 634, 129920. (www)

  @article{ , abstract = {Object-specific selective attention can be achieved either by simultaneously splitting attention to multiple objects, or by sequentially shifting spatial attention among objects. A growing body of research show that object-specific selective attention can be implemented using the second way and that the sequential movement of attention exhibits specific rhythmicity. However, the neurocomputing mechanisms underlying this phenomenon are still not fully understood. To clarify this issue, we conducted magnetoencephalography experiments on healthy participants and subsequently proposed a computational framework based on time-series decomposition and rhythmic analysis to delve into the neural mechanisms of object-specific selective attention. Our investigation reveals that the four single-object attention states are decodable on the level of magnetoencephalography (MEG) sensor signals. Furthermore, these states manifest dynamically and rhythmically during object-specific selective attention. These findings suggest that the attentional rhythm exhibited by neural activity during object-specific selective attention is fundamentally characterized by a set of basic attentional units. This research provides valuable information for future investigations into the brain model of object-specific selective attention.}, archive = {J_NEUCOM}, author = {Chunyu Liu and Xin-Yue Yang and Xueyuan Xu}, doi = {10.1016/j.neucom.2025.129920}, journal = {Neurocomputing}, month = {6}, pages = {129920}, shortjournal = {Neurocomputing}, title = {Brain state model: A novel method to represent the rhythmicity of object-specific selective attention from magnetoencephalography data}, volume = {634}, year = {2025}, }
• (2025). GDoT: A gated dual domain transformer for enhanced MRI off-resonance correction. NEUCOM, 634, 129918. (www)

  @article{ , abstract = {Deep learning-based MRI reconstruction methods have gained significant attention recently due to the need for accelerated MRI scans. However, existing deep learning-based methods for off-resonance correction rely on simple CNNs, resulting in suboptimal solutions. In this paper, we propose a gated dual domain transformer with gated spatial projection and gated frequency projection to effectively handle complex-valued MRI, as the first attempt to utilize transformer-based model for off-resonance correction. Additionally, we introduce a selective perceptual loss with a novel test-time translation-merger to reconstruct perceptually high-quality images without checkerboard artifacts. Experiments on both simulated and real off-resonance MRI datasets demonstrate the effectiveness of our approach. Furthermore, we also present ablation studies to determine the optimal design choices.}, archive = {J_NEUCOM}, author = {Jaesin Ahn and Heechul Jung}, doi = {10.1016/j.neucom.2025.129918}, journal = {Neurocomputing}, month = {6}, pages = {129918}, shortjournal = {Neurocomputing}, title = {GDoT: A gated dual domain transformer for enhanced MRI off-resonance correction}, volume = {634}, year = {2025}, }
• (2025). Solving spatial-temporal PDEs with arbitrary boundary conditions using physics-constrained convolutional recurrent neural networks. NEUCOM, 634, 129917. (www)

  @article{ , abstract = {The inception of physics-constrained or physics-informed machine learning represents a paradigm shift, addressing the challenges associated with data scarcity and enhancing model interpretability. This innovative approach incorporates the fundamental laws of physics as constraints, guiding the training process of machine learning models. In this work, the physics-constrained convolutional recurrent neural network is further extended for solving spatial-temporal partial differential equations with arbitrary boundary conditions. Two notable advancements are introduced: the implementation of boundary conditions as soft constraints through finite difference-based differentiation, and the establishment of an adaptive weighting mechanism for the optimal allocation of weights to various losses. These enhancements significantly augment the network's ability to manage intricate boundary conditions and expedite the training process. The efficacy of the proposed model is validated through its application to two-dimensional phase transition, fluid dynamics, and reaction-diffusion problems, which are pivotal in materials modeling. Compared to traditional physics-constrained neural networks, the physics-constrained convolutional recurrent neural network demonstrates a tenfold increase in prediction accuracy within a similar computational budget. Moreover, the model's exceptional performance in extrapolating solutions for the Burgers' equation underscores its utility. Therefore, this research establishes the physics-constrained recurrent neural network as a viable surrogate model for sophisticated spatial-temporal PDE systems, particularly beneficial in scenarios plagued by sparse and noisy datasets.}, archive = {J_NEUCOM}, author = {Guangfa Li and Yanglong Lu and Dehao Liu}, doi = {10.1016/j.neucom.2025.129917}, journal = {Neurocomputing}, month = {6}, pages = {129917}, shortjournal = {Neurocomputing}, title = {Solving spatial-temporal PDEs with arbitrary boundary conditions using physics-constrained convolutional recurrent neural networks}, volume = {634}, year = {2025}, }
• (2025). DSQN: Robust path planning of mobile robot based on deep spiking Q-network. NEUCOM, 634, 129916. (www)

  @article{ , abstract = {With the rapid advancement of science and technology, the field of mobile robot applications continues to expand, with path planning emerging as a fundamental yet challenging task. While traditional path planning techniques have developed into a relatively complete theoretical system, their performance in uncertain environments remains a critical area of research. To address this, we propose a novel Deep Spiking Q-Network (DSQN) algorithm that significantly enhances path planning performance by leveraging the unique advantages of spiking neural networks (SNNs). Unlike classic Q-learning and its contemporary variants, the DSQN algorithm integrates global and local information simultaneously, resulting in superior overall performance. As the third generation of neural networks, SNNs offer unparalleled robustness and energy efficiency by mimicking biological neural systems. By introducing spiking neurons into the conventional Deep Q-learning (DQN) framework, the DSQN algorithm overcomes key challenges in deep reinforcement learning (DRL), such as limited robustness and high energy consumption. The DSQN training process incorporates both surrogate gradient learning (SGL) and ANN-to-SNN conversion techniques, with SGL demonstrating remarkable effectiveness in mobile robot path planning tasks. Experimental results validate the practicality and efficiency of DSQN, showcasing improved performance across diverse test scenarios compared to the original DQN algorithm. These findings highlight the potential of DSQN to advance path planning in complex and uncertain environments, establishing it as a robust and energy-efficient solution for mobile robotics.}, archive = {J_NEUCOM}, author = {Aakash Kumar and Lei Zhang and Hazrat Bilal and Shifeng Wang and Ali Muhammad Shaikh and Lu Bo and Avinash Rohra and Alisha Khalid}, doi = {10.1016/j.neucom.2025.129916}, journal = {Neurocomputing}, month = {6}, pages = {129916}, shortjournal = {Neurocomputing}, title = {DSQN: Robust path planning of mobile robot based on deep spiking Q-network}, volume = {634}, year = {2025}, }
• (2025). Recent progress in digital twin-driven fault diagnosis of rotating machinery: A comprehensive review. NEUCOM, 634, 129914. (www)

  @article{ , abstract = {The development of the Internet of Things (IoT) and industrial automation has propelled diagnostics into the age of intelligence and digitization. Equipment fault diagnosis is predicted to shift from "preventive" to "predictive" through digital twin (DT) technology, which creates a virtual mirror of the physical entity and simulates its operating state in a genuine operating environment. The operation, development history, and current state of application of DT technology are investigated through a literature review, highlighting the technology's enormous promise in intelligent fault diagnostics. Then, from the four fields of aerospace, transportation, industrial machinery, and energy equipment, the current status of domestic and international research and the latest research developments on the application of DT technology to implement intelligent fault diagnosis (IFD) of equipment are reviewed. On this basis, the limitations of the current application of DT technology in the field of fault diagnosis are analyzed, and the challenges and development trends of the future application of DT technology in the implementation of IFD are discussed and pointed out, providing a clear direction for intelligent fault diagnosis of machinery.}, archive = {J_NEUCOM}, author = {Pengbo Zhang and Renxiang Chen and Lixia Yang and Ye Zou and Liang Gao}, doi = {10.1016/j.neucom.2025.129914}, journal = {Neurocomputing}, month = {6}, pages = {129914}, shortjournal = {Neurocomputing}, title = {Recent progress in digital twin-driven fault diagnosis of rotating machinery: A comprehensive review}, volume = {634}, year = {2025}, }
• (2025). Hybrid safe reinforcement learning: Tackling distribution shift and outliers with the student-t’s process. NEUCOM, 634, 129912. (www)

  @article{ , abstract = {Safe reinforcement learning (SRL) aims to optimize control policies that maximize long-term reward, while adhering to safety constraints. SRL has many real-world applications such as, autonomous vehicles, industrial robotics, and healthcare. Recent advances in offline reinforcement learning (RL) — where agents learn policies from static datasets without interacting with the environment — have made it a promising approach to derive safe control policies. However, offline RL faces significant challenges, such as covariate shift and outliers in the data, which can lead to suboptimal policies. Similarly, online SRL, which derives safe policies through real-time environment interaction, struggles with outliers and often relies on unrealistic regularity assumptions, limiting its practicality. This paper addresses these challenges by proposing a hybrid-offline–online approach. First, prior knowledge from offline learning guides online exploration. Then, during online learning, we replace the popular Gaussian Process (GP) with the Student-t’s Process (TP) to enhance robustness to covariate shift and outliers.}, archive = {J_NEUCOM}, author = {Xavier Hickman and Yang Lu and Daniel Prince}, doi = {10.1016/j.neucom.2025.129912}, journal = {Neurocomputing}, month = {6}, pages = {129912}, shortjournal = {Neurocomputing}, title = {Hybrid safe reinforcement learning: Tackling distribution shift and outliers with the student-t’s process}, volume = {634}, year = {2025}, }
• (2025). Efficient motor imagery electroencephalogram classification via cross tensor coupling decomposition based on augmented covariance networks. NEUCOM, 634, 129911. (www)

  @article{ , abstract = {A major challenge in fully harnessing the potential of motor imagery-based brain-computer interfaces (MI-BCIs) is the accurate classification of MI electroencephalography (MI-EEG) signals. Traditional methods of analyzing MI-EEG signals, such as spatial pattern analysis, focus solely on the average linear correlation within a specific time frame of the signals. To address these limitations, we introduce an effective classification approach named "Cross Tensor Coupling Decomposition (CTCD) based on Augmented Covariance Networks (ACNs)". Based on a specific embedding of the original system, we introduce the ACN, which can be conceptualized as a standard covariance network computed within a high-dimensional space. It not only integrates the spatiotemporal characteristics of the signals but also enhances the representation of nonlinear information. By constructing ACN tensors and applying CTCD, we rapidly and accurately extract multidimensional deep features from different categories of MI tasks, significantly enhancing decoding accuracy. Validating on public datasets, we achieved an average accuracy of 91.75 % ± 3.25 % on Dataset 1 and 85.86 % ± 7.93 % on Dataset 2 in subject-independent binary classification experiments, and for cross-subject binary classification experiments, we obtained an average accuracy of 93.91 % ± 8.71 % on Dataset 1 and 84.36 % ± 16.63 % on Dataset 2. These results highlight the robustness and superiority of our approach. Furthermore, our method enables rapid decoding and real-time feedback, enhancing the practical application of MI-BCIs.}, archive = {J_NEUCOM}, author = {Hechong Su and Jieren Xie and Zengyao Yang and Yuncheng Ge and Jingya Fu and Chengxi Xie and Kai Zhang and Xinyi Hu and Sicong Zhang and Guanghua Xu}, doi = {10.1016/j.neucom.2025.129911}, journal = {Neurocomputing}, month = {6}, pages = {129911}, shortjournal = {Neurocomputing}, title = {Efficient motor imagery electroencephalogram classification via cross tensor coupling decomposition based on augmented covariance networks}, volume = {634}, year = {2025}, }
• (2025). Single-shot phase-shifting composition fringe projection profilometry by multi-attention fringe restoration network. NEUCOM, 634, 129908. (www)

  @article{ , abstract = {3D measurement techniques have permeated sectors like intelligent manufacturing and online inspection, demanding higher real-time performance in industrial applications. Recently, deep-learning-based profilometry (DLP) has gained significant attention due to the powerful feature extraction capabilities of convolutional neural networks, enabling single-shot projection to meet both real-time and robustness requirements in dynamic measurement scenarios. However, while these methods yield promising results, they predominantly emphasize model improvements (e.g., optimizing module functionality or expanding network parameters) often neglecting the parallel integration with traditional physical models and the optimization of input features. To overcome these limitations, a novel DLP that combines parallel phase-shifting theory with a fringe restoration network has been proposed and leverages a high-quality dataset, balancing the trade-offs between projection efficiency and reconstruction accuracy: (1) A learning-based spatial composite phase-shifting profilometry (LSCPP) is proposed, marking the first application of equal thickness interference theory from digital holography to DLP. This mitigates pixel loss caused by frequency-domain filtering in FTP while ensuring high-precision fringe reconstruction without auxiliary templates; (2) A multi-exposure dataset is constructed, accompanied by the introduction of a novel evaluation metric, modulation error rate (MER), for performance assessment; (3) Extensive experiments are conducted across both static and dynamic scenarios, utilizing multi-exposure and public datasets, and are compared with state-of-the-art DLPs. It demonstrates the proposed LSCPP’s real-time capability and generalization performance in dynamic measurement environments.}, archive = {J_NEUCOM}, author = {Jiayi Qin and Yansong Jiang and Yiping Cao and Haitao Wu}, doi = {10.1016/j.neucom.2025.129908}, journal = {Neurocomputing}, month = {6}, pages = {129908}, shortjournal = {Neurocomputing}, title = {Single-shot phase-shifting composition fringe projection profilometry by multi-attention fringe restoration network}, volume = {634}, year = {2025}, }
• (2025). A parallel neural network based structural anomaly detection: Leveraging time-frequency domain features. NEUCOM, 634, 129907. (www)

  @article{ , abstract = {Structural anomaly detection is essential for ensuring the safety, reliability, and longevity of building engineering. By identifying deviations from normal patterns, it enables early intervention and prevents potential failures. However, most existing methods rely on a single feature extracted from the time domain or the frequency domain. Time-domain features alone are insufficient to capture variations in the frequency components, while frequency-domain features fail to account for transient behaviours in the time domain. This limitation significantly reduces detection performance, particularly when dealing with nonlinear and non-stationary signals. To address the issues, this study proposes a new framework for anomaly detection using parallel time convolutional networks (TCN) and wavelet decomposition based convolutional neural networks (WD-CNN), termed PTWC. In this framework, time-frequency domain features are achieved by utilizing TCN to capture time-domain features and WD-CNN to capture frequency-domain features, distinguishing structural anomaly patterns. The proposed PTWC framework is validated on an actual three-story frame structure. Compared with ten baseline methods, the experimental results demonstrate that PTWC has high accuracy and achieves at least 7 % improvement in area under the curve (AUC) scores, thus confirming its superior performance in structural anomaly detection.}, archive = {J_NEUCOM}, author = {Yingying He and Bo Yang and Weihong Jin and Likai Zhang and Hongyang Chen}, doi = {10.1016/j.neucom.2025.129907}, journal = {Neurocomputing}, month = {6}, pages = {129907}, shortjournal = {Neurocomputing}, title = {A parallel neural network based structural anomaly detection: Leveraging time-frequency domain features}, volume = {634}, year = {2025}, }
• (2025). Enhancing fourier neural operators with CNNs architectures: Pooling, groupwise convolution and inverted block. NEUCOM, 634, 129905. (www)

  @article{ , abstract = {In this work, we hypothesize that effective architectural configurations from Convolutional Neural Networks (CNNs) can significantly enhance the performance of the Fourier Neural Operator (FNO). Pooling layers, such as max-pooling and average-pooling in CNNs, play a key role in improving the learnability of nonlocal features in images. However, despite its resemblance to CNNs, FNO lacks such mechanisms due to two primary challenges: (1) applying traditional pooling layers violates discretization invariance, and (2) the fixed size of pooling windows cannot be pre-defined because of variable data resolutions. To address these issues, we propose a novel “pooling operator” for the FNO architecture that preserves discretization invariance and is differentiable with respect to window size. Additionally, we adapt groupwise convolution and the inverted residual block from CNNs to FNOs. Together, the pooling operator, groupwise convolution, and inverted residual block play a pivotal role in achieving superior performance. Benchmark experiments on the 2D Navier–Stokes and 1D Burger’s equation datasets demonstrate the effectiveness of our proposed architecture.}, archive = {J_NEUCOM}, author = {Seungtae Park and Heejoon Jeon and Hyung Ju Hwang}, doi = {10.1016/j.neucom.2025.129905}, journal = {Neurocomputing}, month = {6}, pages = {129905}, shortjournal = {Neurocomputing}, title = {Enhancing fourier neural operators with CNNs architectures: Pooling, groupwise convolution and inverted block}, volume = {634}, year = {2025}, }
• (2025). Label-only model inversion attacks: Adaptive boundary exclusion for limited queries. NEUCOM, 634, 129902. (www)

  @article{ , abstract = {Recent studies have shown that privacy data in deep learning models is vulnerable to attacks by adversaries who can restore private images even with access only to the labels of DNN models. However, existing similar attacks rely heavily on extensive access to the target model, making them easily detectable by defenders. This paper introduces an Adaptive Boundary Exclusion method, called ABE-MI, that achieves a high success rate with minimal access attempts. The core idea of our algorithm is to estimate the gradient of the target point through Gaussian sampling that progressively expands outward, dynamically adjusting the movement step of the data point to reach the centroid of the decision area quickly. Experiments show that our algorithm can attack both CNN and ViT models, and its performance on ViT models far surpasses that of the state-of-the-art (SOTA) algorithms. Moreover, compared to the best current label-only algorithms, our approach requires significantly fewer queries and even exceeds the performance of white-box attacks on other metrics. Furthermore, our algorithm remains effective on tabular data.}, archive = {J_NEUCOM}, author = {Jiayuan Wu and Chang Wan and Hao Chen and Zhonglong Zheng and Yaxin Sun}, doi = {10.1016/j.neucom.2025.129902}, journal = {Neurocomputing}, month = {6}, pages = {129902}, shortjournal = {Neurocomputing}, title = {Label-only model inversion attacks: Adaptive boundary exclusion for limited queries}, volume = {634}, year = {2025}, }
• (2025). Feature representation fidelity preservation during neural network pruning for enhanced compression efficiency. NEUCOM, 634, 129901. (www)

  @article{ , abstract = {Channel pruning is a widely used technique for compressing convolutional neural networks (CNNs), primarily by identifying and eliminating redundant filters or channels. However, existing pruning criteria often fail to consider inter-channel interactions, leading to inaccurate importance scores. Furthermore, many criteria focus solely on the immediate influence of channels, neglecting their contributions to higher-level feature representations. These limitations hinder CNNs' ability to retain critical information after pruning. In this paper, we propose a novel criterion to address these issues. Our method highlights how feature information loss occurs due to inter-channel dependencies during forward propagation, which affects the accuracy of feature-based importance scores. To mitigate this, we introduce a per-layer channel contribution analysis that captures channel-layer interactions, ensuring more precise pruning decisions. Extensive experiments on image classification (CIFAR-10, ImageNet), object detection (PASCAL VOC, COCO 2017), and 3D point cloud analysis demonstrate the effectiveness of our approach. For instance, our method achieves a 52.6 % reduction in floating-point operations (FLOPs) in ResNet-56 on CIFAR-10 while improving accuracy by 0.4 %, outperforming existing pruning techniques. On ImageNet, pruning ResNet-50 results in a 55.5 % FLOPs reduction with only a 0.72 % top-1 accuracy drop, surpassing prior methods. Additionally, pruning MobileNet-V2 achieves a 31.5 % FLOPs reduction with a higher top-1 accuracy than alternative approaches. These results confirm that our method enhances compression efficiency while preserving feature representation fidelity, making it a robust solution for CNN model compression.}, archive = {J_NEUCOM}, author = {Ambuj}, doi = {10.1016/j.neucom.2025.129901}, journal = {Neurocomputing}, month = {6}, pages = {129901}, shortjournal = {Neurocomputing}, title = {Feature representation fidelity preservation during neural network pruning for enhanced compression efficiency}, volume = {634}, year = {2025}, }
• (2025). Dynamic neighborhood selection for regularized local linear embedding with elastic networks. NEUCOM, 634, 129900. (www)

  @article{ , abstract = {The processing of high-dimensional data is of paramount importance in the fields of machine learning and pattern recognition. Local linear embedding (LLE), a popular nonlinear dimensionality reduction method, has gained attention for maintaining the local structure of data in low-dimensional space. However, traditional LLE algorithms have limitations in dealing with datasets that are non-uniformly distributed or noisy. To address this issue, we propose an elastic network regularized LLE algorithm with dynamic neighborhood selection (DNS-ENRLLE). The objective is to enhance the adaptability and robustness of the algorithm through a dynamic neighborhood selection mechanism and a regularization strategy. A series of experiments on synthetic and real datasets demonstrate that our method outperforms existing LLE and other nonlinear dimension reduction methods. In particular, the improved method demonstrates significant performance improvements in the analysis of non-uniformly distributed data, effectively enhancing the accuracy and reliability of data analysis results.}, archive = {J_NEUCOM}, author = {Jiaqi Cheng and Chaojie Wang}, doi = {10.1016/j.neucom.2025.129900}, journal = {Neurocomputing}, month = {6}, pages = {129900}, shortjournal = {Neurocomputing}, title = {Dynamic neighborhood selection for regularized local linear embedding with elastic networks}, volume = {634}, year = {2025}, }
• (2025). Improving GBDT performance on imbalanced datasets: An empirical study of class-balanced loss functions. NEUCOM, 634, 129896. (www)

  @article{ , abstract = {Class imbalance poses a persistent challenge in machine learning, particularly for tabular data classification tasks. While Gradient Boosting Decision Trees (GBDT) models are widely regarded as state-of-the-art for these tasks, their effectiveness diminishes in the presence of imbalanced datasets. This paper is the first to comprehensively explore the integration of class-balanced loss functions into three popular GBDT algorithms, addressing binary, multi-class, and multi-label classification. We present a novel benchmark, derived from extensive experiments across diverse datasets, to evaluate the performance gains from class-balanced losses in GBDT models. Our findings establish the efficacy of these loss functions in enhancing model performance under class imbalance, providing actionable insights for practitioners tackling real-world imbalanced data challenges. To bridge the gap between research and practice, we introduce an open-source Python package that simplifies the application of class-balanced loss functions within GBDT workflows, democratizing access to these advanced methodologies. The code is available at https://github.com/Luojiaqimath/ClassbalancedLoss4GBDT .}, archive = {J_NEUCOM}, author = {Jiaqi Luo and Yuan Yuan and Shixin Xu}, doi = {10.1016/j.neucom.2025.129896}, journal = {Neurocomputing}, month = {6}, pages = {129896}, shortjournal = {Neurocomputing}, title = {Improving GBDT performance on imbalanced datasets: An empirical study of class-balanced loss functions}, volume = {634}, year = {2025}, }
• (2025). Unsupervised infrared–visible person re-identification by multi-level dual-stream contrastive learning. NEUCOM, 634, 129895. (www)

  @article{ , abstract = {Unsupervised Visible–Infrared Person Re-identification (USL-VI-ReID) aims to learn a label-free cross-modal retrieval model to reduce reliance on expensive manual annotations for cross-modal data. Previous works on USL-VI-ReID tend to generate sample labels from the perspective of inter-modality contrastive learning, while neglecting the information differences between cameras, resulting in low credibility of the generated sample labels. Additionally, traditional inter-modality contrastive learning methods use shared feature encoders, which cannot learn personalized modality features. To address these issues, this paper proposes a Dual-Stream Contrastive Learning Method (DCM) that jointly learns personalized feature encoders from both inter-modality and inter-camera perspectives, generating more reliable sample labels, which we refer to as a Multi-level Dual-Stream Contrastive Learning Method. We first establish a Dual-Stream Contrastive Learning (DCL) pre-training framework, which includes three different feature encoders for learning pedestrian representations within specific modalities, and optimize these encoders through contrastive learning using all training data. Then, based on these three feature encoders, we construct a Multi-level Dual-Stream Contrastive Learning framework (MLA) for cross-camera to cross-modal feature clustering and pseudo-label generation. This enables MLA to focus more on learning hierarchical domain information during subsequent model training, combining multi-level feature information aggregation to obtain more accurate cross-modal pseudo-labels. Finally, to further ensure the semantic consistency of the two modal labels, we insert a Cross-Modality Alignment Module (CMAM) between modalities to further improve the recognition accuracy of USL-VI-ReID. Numerous experiments demonstrate that the proposed method stands out among various USL-VI-ReID approaches, exhibiting promising performance.}, archive = {J_NEUCOM}, author = {Yifeng Zhang and Canlong Zhang and Haifei Ma and Zhixin Li and Zhiwen Wang and Chunrong Wei}, doi = {10.1016/j.neucom.2025.129895}, journal = {Neurocomputing}, month = {6}, pages = {129895}, shortjournal = {Neurocomputing}, title = {Unsupervised infrared–visible person re-identification by multi-level dual-stream contrastive learning}, volume = {634}, year = {2025}, }
• (2025). MSPL: Multi-granularity semantic prototype learning for occluded person re-identification. NEUCOM, 634, 129894. (www)

  @article{ , abstract = {Although person re-identification has made remarkable progress in recent years, it remains a challenging problem in occluded scenes. Some existing popular methods attempt to decouple the human local features directly by clustering semantics to counter occlusion. However, due to the lack of good semantic learning strategies, the network will inevitably acquire non-discriminative local features. To effectively separate body parts and eliminate occlusion interference, we propose a Multi-granularity Semantic Prototype Learning (MSPL) network based on a superior semantic learning strategy. MSPL mainly consists of a prototype embedding encoder and a multi-granularity prototype decoder. With the help of our proposed Dual Parallel Attention (DPA) module and easy-to-hard learning strategy, the encoder can gradually aggregate semantics and effectively separate discriminative fine-grained features. We also propose a Pose-guided Feature Enhancement (PFE) module and a Prototype Cyclic Grouping (PCG) module in the decoder to resist background clutter and enhance network robustness for complex occlusion scenes. Finally, we design a Prototype Discriminability (PD) loss to reduce the focused redundancy to increase the discrepancy between the focused areas of prototypes. Extensive experimental results on several challenging benchmark datasets show that our MSPL achieves excellent performance and good generalization ability.}, archive = {J_NEUCOM}, author = {Zhihao Li and Huaxiang Zhang and Lei Zhu and Jiande Sun and Li Liu}, doi = {10.1016/j.neucom.2025.129894}, journal = {Neurocomputing}, month = {6}, pages = {129894}, shortjournal = {Neurocomputing}, title = {MSPL: Multi-granularity semantic prototype learning for occluded person re-identification}, volume = {634}, year = {2025}, }
• (2025). Learning with noisy labels for classifying biological echoes in polarimetric weather radar observations using artificial neural networks. NEUCOM, 634, 129892. (www)

  @article{ , abstract = {The identification of biological echoes in radar data has revolutionized research into airborne migratory species. Deep learning applied to polarimetric weather radar observations can reveal signature patterns of mass movement by bio-scatterers such as birds, bats, and insects. However, due to the difficulties in labelling bio-scatterers in these data, threshold approaches have been proposed in the literature. In this research, we used the depolarization ratio (DR) based on differential reflectivity (zDR) and the cross-correlation coefficient (pHV), along with citizen scientist-reported data, to label bio-scatterers for deep learning. This method of labelling biological echoes in radar signatures is prone to noise, which impacts the accuracy of any model that relies on it. We introduce a novel semi-supervised co-training approach that uses a bootstrap ensemble with a confidence threshold. Our ensemble consists of the newly proposed STNet and two modified FNet models, which incorporate co-learning through bootstrap sampling for label correction. This innovative method significantly improves classification accuracy across all three multivariate numerical datasets compared to baseline models that lack co-learning with bootstrap-based label correction.}, archive = {J_NEUCOM}, author = {John Atanbori and Christos A. Frantzidis and Mohammed Al-Khafajiy and Aliyu Aliyu and Behnaz Sohani and Kofi Appiah and Harriet Moore and Catherine Sanders and Alastair I. Ward}, doi = {10.1016/j.neucom.2025.129892}, journal = {Neurocomputing}, month = {6}, pages = {129892}, shortjournal = {Neurocomputing}, title = {Learning with noisy labels for classifying biological echoes in polarimetric weather radar observations using artificial neural networks}, volume = {634}, year = {2025}, }
• (2025). Multilingual neural machine translation for low-resource languages by twinning important nodes. NEUCOM, 634, 129890. (www)

  @article{ , abstract = {Multilingual neural machine translation (MNMT) is a novel machine translation approach that benefits from large multilingual resources. However, its performance drops significantly when training with low-resource languages due to the reliance on parameter sharing and data size. In this paper, a new method is proposed to improve the performance of MNMT for a pair of languages where the target language is low-resource. The main idea of this study is to find important nodes that have parameters connected to them that negatively affect an MNMT model and then split those nodes into two sub nodes. Then, the model selects the important sub node that has an effect on the specific language pair to create a twin sub node. This twin sub node helps to strengthen the translation quality of the specific language pair without having a negative effect on other languages. The proposed method works in four steps as: 1) training an MNMT model with parameter sharing over multiple languages, 2) selecting important nodes which negatively affect the MNMT, 3) splitting important nodes into sub nodes, and 4) Twining important sub nodes. The proposed method has been evaluated using several multilingual datasets, including TED 2013, TED 2020, and BIBLE, by examining English-Persian language as a case study. The obtained results show that the proposed method yields the best results for one-to-many and many-to-many models according to the average BLEU value and semantic similarity . The results also show that the proposed method has given better results than other well-known large language models, such as ChatGPT, BING GPT4, and the Google Neural Machine Translation (GNMT) model, when applied to a low-resource language.}, archive = {J_NEUCOM}, author = {Abouzar Qorbani and Reza Ramezani and Ahmad Baraani and Arefeh Kazemi}, doi = {10.1016/j.neucom.2025.129890}, journal = {Neurocomputing}, month = {6}, pages = {129890}, shortjournal = {Neurocomputing}, title = {Multilingual neural machine translation for low-resource languages by twinning important nodes}, volume = {634}, year = {2025}, }
• (2025). Causal and local correlations based network for multivariate time series classification. NEUCOM, 634, 129884. (www)

  @article{ , abstract = {Recently, time series classification has attracted the attention of a large number of researchers, and hundreds of methods have been proposed. However, these methods often ignore the spatial correlations among dimensions and the local correlations among features. To address this issue, the causal and local correlations based network (CaLoNet) is proposed in this study for multivariate time series classification. First, pairwise spatial correlations between dimensions are modeled using causality modeling to obtain the graph structure. Then, a relationship extraction network is used to fuse local correlations to obtain long-term dependency features. Finally, the graph structure and long-term dependency features are integrated into the graph neural network. Experiments on the UEA datasets show that CaLoNet can obtain competitive performance compared with state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Mingsen Du and Yanxuan Wei and Xiangwei Zheng and Cun Ji}, doi = {10.1016/j.neucom.2025.129884}, journal = {Neurocomputing}, month = {6}, pages = {129884}, shortjournal = {Neurocomputing}, title = {Causal and local correlations based network for multivariate time series classification}, volume = {634}, year = {2025}, }
• (2025). MSSFN: A multi-scale sequence fusion network for CT-based diagnosis of pulmonary complications. NEUCOM, 634, 129878. (www)

  @article{ , abstract = {Pulmonary complications exhibit similar symptoms, such as ground-glass opacities and patchy consolidations, posing significant challenges for manual diagnosis. In this paper, a novel Multi-Scale Sequence Fusion Network (MSSFN) designed for a small-scale Computed Tomography (CT) database is proposed for computer-aided identification of pulmonary complications. The MSSFN first applies a transfer learning mechanism that utilizes the convolutional neural network weights pre-trained on a large dataset, enabling optimization of both feature extraction efficiency and model accuracy. Next, the multi-scale feature fusion module uses different types of convolutional layers in parallel to extract and integrate high-level features, enabling the network to capture complex lung lesion characteristics from multiple perspectives and enhance information interaction across different layers. Finally, a sequence feature fusion module establishes three-dimensional connections within the CT data, ensuring comprehensive data fusion by effectively integrating spatial features across the sequence. The comprehensive experimental results demonstrate that, under the same experimental conditions, the MSSFN achieves an accuracy of 82.54% for identifying pulmonary complications in CT images, which exhibits superior performance compared to nine similar network structures as well as other four state-of-the-art deep learning models. Consequently, the proposed MSSFN demonstrates practical significance by providing radiologists with reliable tools for accurately distinguishing various pulmonary complications in CT images. It also holds theoretical value by advancing methods for constructing robust features with enhanced representational capabilities.}, archive = {J_NEUCOM}, author = {Hongfu Zeng and Xinyu Li and Haipeng Xu and Keyi Yu and Huihua Hu and Peishu Wu and Nianyin Zeng}, doi = {10.1016/j.neucom.2025.129878}, journal = {Neurocomputing}, month = {6}, pages = {129878}, shortjournal = {Neurocomputing}, title = {MSSFN: A multi-scale sequence fusion network for CT-based diagnosis of pulmonary complications}, volume = {634}, year = {2025}, }
• (2025). Performance-oriented fault detection and fault-tolerant control for nonlinear uncertain systems: Improved stochastic configuration network-based methods. NEUCOM, 634, 129869. (www)

  @article{ , abstract = {This paper primarily focuses on performance-oriented fault detection (FD) and fault-tolerant control (FTC) for nonlinear uncertain systems. To achieve this, we initially develop an improved stochastic configuration network-based (ISCN-based) method for modeling nonlinear uncertain systems, leveraging the supervision mechanism and incremental construction techniques. It is followed by a performance indicator to enhance the robustness against the uncertainties and the sensitivity to the faults, which is further implemented for performance-oriented FD purpose. Subsequently, an adaptive FTC method is proposed to recover the nonlinear system performance, and associated with it, the configuration of the fault-tolerant controller parameters is investigated. Finally, a case study on the continuous stirred-tank reactor (CSTR) system is presented to showcase the efficacy and advantage of the proposed methods.}, archive = {J_NEUCOM}, author = {Jingjing Gao and Feng Gao and Zhengxuan Zhang and Xu Yang and Jian Huang and Kaixiang Peng}, doi = {10.1016/j.neucom.2025.129869}, journal = {Neurocomputing}, month = {6}, pages = {129869}, shortjournal = {Neurocomputing}, title = {Performance-oriented fault detection and fault-tolerant control for nonlinear uncertain systems: Improved stochastic configuration network-based methods}, volume = {634}, year = {2025}, }
• (2025). SPRInT: Scaling programmatic reasoning for INstruction tuning in mathematics. NEUCOM, 634, 129868. (www)

  @article{ , abstract = {We present SPRInT , a novel approach for large-scale, cost-effective synthesis of instruction-tuning datasets, leveraging Program-of-Thoughts (PoT) to enhance mathematical reasoning capabilities. Through the SPRInT framework, we synthesized data from seven high-quality open-source math datasets (including GSM8K, MATH, AQuA), and developed InfinityMATH -a dataset containing over 100,000 samples generated from QA pairs, offering extensive coverage across various mathematical domains. The SPRInT model series, fine-tuned on InfinityMATH using open-source language and code models such as Llama2-7B, Mistral-7B, and CodeLlama-7B, achieved remarkable improvements in mathematical reasoning, with performance gains between 184.7% and 514.3%. In zero-shot settings, our SPRInT -CodeLlama-7B model surpassed MAmmoTH-Coder on widely-used benchmarks, including GSM8K (65.80% vs. 56.86%) and MATH (34.06% vs. 29.88%). To assess logical consistency in numerical transformations, we created the GSM8K+ and MATH＋ test sets by modifying the numerical values in the original datasets. While traditional models struggled with these alterations, the SPRInT models exhibited superior robustness. The InfinityMATH dataset is publicly available at https://huggingface.co/datasets/BAAI/InfinityMATH .}, archive = {J_NEUCOM}, author = {Yan Yan and Lin Li and Bo-Wen Zhang}, doi = {10.1016/j.neucom.2025.129868}, journal = {Neurocomputing}, month = {6}, pages = {129868}, shortjournal = {Neurocomputing}, title = {SPRInT: Scaling programmatic reasoning for INstruction tuning in mathematics}, volume = {634}, year = {2025}, }
• (2025). SCSA: Exploring the synergistic effects between spatial and channel attention. NEUCOM, 634, 129866. (www)

  @article{ , abstract = {Channel and spatial attentions have respectively brought significant improvements in extracting feature dependencies and spatial structure relations for various downstream vision tasks. The combined use of both channel and spatial attentions is widely considered beneficial for further performance improvement; however, the synergistic effects between channel and spatial attentions, especially in terms of spatial guidance and mitigating semantic disparities, have not yet been thoroughly studied. This motivates us to propose a novel Spatial and Channel Synergistic Attention module (SCSA), entailing our investigation toward the synergistic relationship between spatial and channel attentions at multiple semantic levels. Our SCSA consists of two parts: the Shareable Multi-Semantic Spatial Attention (SMSA) and the Progressive Channel-wise Self-Attention (PCSA). SMSA integrates multi-semantic information and utilizes a progressive compression strategy to inject discriminative spatial priors into PCSA’s channel self-attention, effectively guiding channel recalibration. Additionally, the robust feature interactions based on the Channel-wise single-head self-attention mechanism in PCSA further mitigate the disparities in multi-semantic information among different sub-features within SMSA. We conduct extensive experiments on seven benchmark datasets, including classification on ImageNet-1K, object detection on MSCOCO, segmentation on ADE20K, and four other complex scene detection datasets. Our results demonstrate that our proposed SCSA not only surpasses the current plug-and-play state-of-the-art attention but also exhibits enhanced generalization capabilities across various task scenarios. The code and models are available at: https://github.com/HZAI-ZJNU/SCSA .}, archive = {J_NEUCOM}, author = {Yunzhong Si and Huiying Xu and Xinzhong Zhu and Wenhao Zhang and Yao Dong and Yuxing Chen and Hongbo Li}, doi = {10.1016/j.neucom.2025.129866}, journal = {Neurocomputing}, month = {6}, pages = {129866}, shortjournal = {Neurocomputing}, title = {SCSA: Exploring the synergistic effects between spatial and channel attention}, volume = {634}, year = {2025}, }
• (2025). Sentiment analysis applications using deep learning advancements in social networks: A systematic review. NEUCOM, 634, 129862. (www)

  @article{ , abstract = {Sentiment analysis is required to extract insights from social media content affecting decision-making and personalized services. The enormous volume of social network information has to be technically processed to extract relevant knowledge. Sentiment analysis is the most widely used method for this purpose. The current techniques of sentiment analysis have made significant progress in various fields. However, the potential of social networks to better understand human emotions and the recent advancements in deep learning necessitate the review and use of advanced sentiment analysis techniques that still require more attention from researchers in this field. In this regard, this review presents a systematic literature review (SLR) on the advancements of sentiment analysis using deep learning techniques in social networks from 2019 to May 2024. Furthermore, this review emphasizes that sentiment analysis can provide meaningful insights into information extracted from large and diverse datasets such as social media, which is extremely important for decision-making and personalized services. It also highlights mental health concerns as one of the windows into the emotional atmosphere of social networks. In addition, this SLR provides a technical taxonomy and comparison of various deep learning approaches. This SLR not only provides a comprehensive overview of the most advanced techniques and methodologies now used in sentiment analysis but also highlights forthcoming challenges and open issues that need to be addressed in the future. This study helps researchers and practitioners use deep learning to improve sentiment analysis applications and digital social well-being.}, archive = {J_NEUCOM}, author = {Erfan Bakhtiari Ramezani}, doi = {10.1016/j.neucom.2025.129862}, journal = {Neurocomputing}, month = {6}, pages = {129862}, shortjournal = {Neurocomputing}, title = {Sentiment analysis applications using deep learning advancements in social networks: A systematic review}, volume = {634}, year = {2025}, }
• (2025). Unsupervised visual similarity-based medical image retrieval via dual-encoder and metric learning. NEUCOM, 634, 129861. (www)

  @article{ , abstract = {Skin cancer stands as one of the most lethal cancers, with its prognosis heavily reliant upon timely diagnosis. Medical image retrieval (MIR) techniques aim to retrieve images similar to the query one from a dataset, assisting doctors in disease diagnosis and clinical treatment programming. However, a significant challenge in dermatoscopic image interpretation stems from the heterogeneity of lesion visual appearance, which hinders existing methods from effectively balancing visual similarity with disease category. To address this issue, we propose an unsupervised approach for medical image retrieval to enhance the visual congruity of retrieval while ensuring disease accuracy.Dual-Encoder is trained to extract image features via a self-distilling dual network, and followed by subsequent refinement of the feature representation utilizing unsupervised metric learning methodologies. Our method is tested on the skin cancer dataset ISIC2019, evaluated using dual-dimensional metrics of visual and disease similarity. To assess visual similarity performance, we constructed a test visual dataset and categorize into 27 visual types we fine-tuned the pre-trained model on multiple datasets encompassing ultrasound, MRI and CT images to evaluate its applicability, which outperforms other advanced generalized retrieval methods. The experiments demostrated the effectiveness and generalization of our proposed method.}, archive = {J_NEUCOM}, author = {Xiya Weng and Yan Zhuang and Rui Wang and Ke Chen and Lin Han and Zhan Hua and Jiangli Lin}, doi = {10.1016/j.neucom.2025.129861}, journal = {Neurocomputing}, month = {6}, pages = {129861}, shortjournal = {Neurocomputing}, title = {Unsupervised visual similarity-based medical image retrieval via dual-encoder and metric learning}, volume = {634}, year = {2025}, }
• (2025). Graph-based interactive knowledge distillation for social relation continual learning. NEUCOM, 634, 129860. (www)

  @article{ , abstract = {As multimedia advances, there is a growing need for machines to adeptly understand diverse social relations. Traditional methods for recognizing these relations, which are limited to a fixed number of classes, are ill-equipped for continual learning as new social interactions emerge. To address this prob-lem, we propose a pioneering Graph-based Interactive Knowledge Distillation (GI-KD) method for social relation continual learning. GI-KD, embedded in a class incremental learning structure, creates a balanced system where previously learned social relations and new knowledge are positioned at either end of the scale. The old and new knowledge is learned dynamically by adjusting the tilt of the balance. To achieve this balance, we propose a novel Libra loss function, which evaluate the relative contribution of old and new information and thus guides the adaptive fine-tuning of the model. We evaluate the GI-KD on three public social relation recognition (SRR) datasets, under different data distribution strategies. Our method shows a remarkable average 3.6% increase in incremental accuracy over current CIL techniques, effectively reducing catastrophic forgetting. Furthermore, GI-KD improves mAP and Acc by 4.6%, 5.4%, and 4.5%, respectively, compared to current CIL techniques, highlighting its strength in both continual learning and SRR.}, archive = {J_NEUCOM}, author = {Wang Tang and Linbo Qing and Pingyu Wang and Lindong Li and Yonghong Peng}, doi = {10.1016/j.neucom.2025.129860}, journal = {Neurocomputing}, month = {6}, pages = {129860}, shortjournal = {Neurocomputing}, title = {Graph-based interactive knowledge distillation for social relation continual learning}, volume = {634}, year = {2025}, }
• (2025). Sample-pair learning network for extremely imbalanced classification. NEUCOM, 634, 129859. (www)

  @article{ , abstract = {In data classification, class-balanced data is ideal, but real datasets are often imbalanced, necessitating rebalancing through methods like resampling. In recent years, some new generative model-based resampling methods have been proposed. However, when facing extreme class imbalance, where the minority class is strongly underrepresented and on its own does not contain enough information to conduct the generative process. Some deep learning methods have been proposed to solve extremely imbalanced classification problems, but some of them are only used for specific datasets. Therefore, we proposed a novel deep learning method that combines a generative strategy with multi-task joint learning, termed sample-pair learning network (SPLN), for extremely imbalanced classification. The network consists of data preprocessing and multi-task joint learning modules. During data preprocessing, the training set is expanded by constructing positive and negative sample-pairs, then rebalanced using a strategy combining attention and resampling, termed undersampling based on attention power values (APVUS). The multi-task joint learning module employs a Siamese convolutional subnetwork to measure the similarity between sample-pairs and a multi-layer perceptron to recognize the category of single samples. The module can reduce the risk of overfitting caused by excessive noise in the training set. Finally, we designed a voting model based on the Siamese convolutional subnetwork to infer the categories of test samples. Experimental results demonstrate that our approach outperforms state-of-the-art generative model-based methods and is effective and general for extremely imbalanced classification.}, archive = {J_NEUCOM}, author = {Linjun Chen and Xiao-Yuan Jing and Runhang Chen and Fei Wu and Yongchang Ding and Changhui Hu and Ziyun Cai}, doi = {10.1016/j.neucom.2025.129859}, journal = {Neurocomputing}, month = {6}, pages = {129859}, shortjournal = {Neurocomputing}, title = {Sample-pair learning network for extremely imbalanced classification}, volume = {634}, year = {2025}, }
• (2025). Distantly supervised relation extraction with multi-level contextual information integration. NEUCOM, 634, 129858. (www)

  @article{ , abstract = {The core objective of distantly Supervised relation extraction (DSRE) is to automatically and at scale extract relational information between two entities from unstructured text. However, the generation of noisy data in the automated labeling process is inevitable. To address the limitations of DSRE in terms of contextual information, current mainstream methods primarily focus on enriching entity representations through external knowledge bases. Nevertheless, there remains a wealth of untapped and rich information within the text itself, which plays a crucial role in helping the model to better understand the relationships between entities. In response to this, we propose a novel strategy that does not require the introduction of additional external knowledge. Instead, it cleverly leverages the principle of semantic similarity to construct contextually rich information that is highly semantically aligned with individual sentences. Furthermore, we incorporate information from Knowledge Graphs (KG) to achieve a deep integration of both textual and KG information. Specifically, we share contextual information at any level of the text with the KG information, merging them as the final bag-level representation for relation extraction. Experimental results demonstrate that updating the KG encoder using only the text encoder yields better performance. Furthermore, experiments conducted on two different domain-specific relation extraction datasets confirmed that our model achieved superior performance.}, archive = {J_NEUCOM}, author = {Danjie Han and Heyan Huang and Shumin Shi and Changsen Yuan and Cunhan Guo}, doi = {10.1016/j.neucom.2025.129858}, journal = {Neurocomputing}, month = {6}, pages = {129858}, shortjournal = {Neurocomputing}, title = {Distantly supervised relation extraction with multi-level contextual information integration}, volume = {634}, year = {2025}, }
• (2025). Multi-modal information fusion for multi-task end-to-end behavior prediction in autonomous driving. NEUCOM, 634, 129857. (www)

  @article{ , abstract = {Behavior prediction in autonomous driving is increasingly achieved through end-to-end frameworks that predict vehicle states from multi-modal information, streamlining decision-making and enhancing robustness in time-varying road conditions. This study proposes a novel multi-modal information fusion-based, multi-task end-to-end model that integrates RGB images, depth maps, and semantic segmentation data, enhancing situational awareness and predictive precision. Utilizing a Vision Transformer (ViT) for comprehensive spatial feature extraction and a Residual-CNN-BiGRU structure for capturing temporal dependencies, the model fuses spatiotemporal features to predict vehicle speed and steering angle with high precision. Through comparative, ablation, and generalization tests on the Udacity and self-collected datasets, the proposed model achieves steering angle prediction errors of MSE 0.012 rad, RMSE 0.109 rad, and MAE 0.074 rad, and speed prediction errors of MSE 0.321 km/h, RMSE 0.567 km/h, and MAE 0.373 km/h, outperforming existing driving behavior prediction models. Key contributions of this study include the development of a channel difference attention mechanism and advanced spatiotemporal feature fusion techniques, which improve predictive accuracy and robustness. These methods effectively balance computational efficiency and predictive performance, contributing to practical advancements in driving behavior prediction.}, archive = {J_NEUCOM}, author = {Guo Baicang and Liu Hao and Yang Xiao and Cao Yuan and Jin Lisheng and Wang Yinlin}, doi = {10.1016/j.neucom.2025.129857}, journal = {Neurocomputing}, month = {6}, pages = {129857}, shortjournal = {Neurocomputing}, title = {Multi-modal information fusion for multi-task end-to-end behavior prediction in autonomous driving}, volume = {634}, year = {2025}, }
• (2025). Model-free extended Q-learning method for h∞ output tracking control of networked control systems with network delays and packet loss. NEUCOM, 634, 129846. (www)

  @article{ , abstract = {In this paper, the extended Q-learning method is used to study the H ∞ output tracking control (HOTC) problem of networked control systems with state delay and data loss. Compared with the existing results, the network control system in this paper contains both network delays and packet loss, as well as external disturbances. To deal with the disturbances, the H ∞ control problem is transformed into the maximum and minimum value problem, which is solved by the method of zero-sum game. The packet loss and delay of the state make it difficult to obtain accurate current state information. Therefore, it is necessary to design a new smith predictor that contains delay and packet loss to predict the current state. Using the predicted state, the extended Q-learning algorithm is implemented to solve the H ∞ output tracking problem with unknown dynamics of the system. Then, the convergence of the extended Q-learning algorithm is proved. Moreover, the stability and optimality of the proposed method are analyzed in the theorems. Finally, numerical simulation is performed to verify the effectiveness of the proposed algorithm.}, archive = {J_NEUCOM}, author = {Longyan Hao and Chaoli Wang and Dong Liang and Shihua Li}, doi = {10.1016/j.neucom.2025.129846}, journal = {Neurocomputing}, month = {6}, pages = {129846}, shortjournal = {Neurocomputing}, title = {Model-free extended Q-learning method for h∞ output tracking control of networked control systems with network delays and packet loss}, volume = {634}, year = {2025}, }
• (2025). H∞ human-assistance fuzzy control of discrete-time nonlinear human-in-the-loop systems. NEUCOM, 634, 129845. (www)

  @article{ , abstract = {Reflecting the change of human behavior on control performance of the human–machine collaboration is a difficult task. In this paper, an H ∞ human-assistance fuzzy control is developed for discrete-time nonlinear Markov jump human-in-the-loop (HiTL) systems. More specifically, to improve the human internal state, giving control input to human is considered in a controlled hidden Markov model utilized to describe human behavior. Then, a Takagi–Sugeno fuzzy model is employed to represent the nonlinear HiTL system. Based on this model and by using a stochastic Lyapunov functional, sufficient conditions for the existence of an H ∞ human-assistance fuzzy controller are provided in terms of linear matrix inequalities guaranteeing closed-loop stochastic stability of HiTL fuzzy system with an H ∞ control performance. Finally, simulation results on a single-degree-of-freedom vehicle with HiTL demonstrate that both adding the control input to the driver via the warning system and applying the suggested human-assistance fuzzy control to vehicle via automation are effective.}, archive = {J_NEUCOM}, author = {Xiu-Mei Zhang and Huai-Ning Wu}, doi = {10.1016/j.neucom.2025.129845}, journal = {Neurocomputing}, month = {6}, pages = {129845}, shortjournal = {Neurocomputing}, title = {H∞ human-assistance fuzzy control of discrete-time nonlinear human-in-the-loop systems}, volume = {634}, year = {2025}, }
• (2025). Improved YOLOv7 for small object detection in airports: Task-oriented feature learning with gaussian wasserstein loss and attention mechanisms. NEUCOM, 634, 129844. (www)

  @article{ , abstract = {Small objects in the airport scene, such as Persons and Vehicles, can lead to low accuracy and robustness of the object detection task. To address the above problems, an improved YOLOv7 model is proposed to detect certain objects on the airport scene. Considering the perspective distortion of the monitoring camera on the airport surface, the deformable convolutional feature extractor (DCFE) is first designed to adaptively extract features from input images for irregular occlusion objects. To learn task-oriented features from different channels, the attention mechanism is incorporated into the backbone to focus on informative concepts in a data-driven manner, formulating an attention feature extractor (AttFE). During the model training, the Normalized Gaussian Wasserstein distance (NWD) is considered as the loss function to measure the prediction errors after converting the bounding boxes into Gaussian distribution, thereby enhancing the ability to fit the small objects. A real-world airport surface dataset (ASD) is constructed to validate the proposed model. Extensive experimental results demonstrate that the proposed model outperforms selective baselines, achieving a 1.2% absolute improvement in mAP over the original YOLOv7 network. Experiments conducted on multiple common datasets and the results demonstrate that the proposed model exhibits superior performance in terms of mAP. All proposed technical modules contribute to expected performance improvement. Most importantly, the proposed model achieves higher performance for small objects and has the desired robustness over occluded objects.}, archive = {J_NEUCOM}, author = {Ruijie Peng and Chuanlin Liao and Weijun Pan and Xiaolin Gou and Jianwei Zhang and Yi Lin}, doi = {10.1016/j.neucom.2025.129844}, journal = {Neurocomputing}, month = {6}, pages = {129844}, shortjournal = {Neurocomputing}, title = {Improved YOLOv7 for small object detection in airports: Task-oriented feature learning with gaussian wasserstein loss and attention mechanisms}, volume = {634}, year = {2025}, }
• (2025). Conv-enhanced transformer and robust optimization network for robust multimodal sentiment analysis. NEUCOM, 634, 129842. (www)

  @article{ , abstract = {Multimodal sentiment analysis seeks to utilize data from multiple modalities to enhance the overall performance of sentiment analysis. However, in real-world applications, missing modality data is an inevitable phenomenon. The semantic sparsity associated with missing data hinders existing methods from effectively capturing local dependencies among modalities, resulting in inefficiencies during cross-modal interactions. Furthermore, simplistic feature reconstruction strategies lead to suboptimal representation learning, rendering models susceptible to noise interference. To address these issues, we propose a novel Conv-Enhanced Transformer and Robust Optimization Network (CTRN) for robust multimodal sentiment analysis. Firstly, this method effectively models local dependencies by incorporating convolutional layers within the cross-modal Transformer, while exploring the intrinsic correlations between global multimodal context and local unimodal feature, thereby facilitating cross-modal fusion. Additionally, to enable effective representation learning and reduce the impact of noisy data, we implements Auxiliary Robust Optimization (ARO), integrating feature reconstruction and adversarial training to encourage the model to learn supplementary semantic information from the differences between complete and missing data. Finally, the introduction of momentum distillation further mitigates the influence of noisy data on model performance. Experimental results on three benchmark datasets demonstrate that our approach significantly enhances the performance of robust multimodal sentiment analysis.}, archive = {J_NEUCOM}, author = {Bin Sun and Li Jia and Yiming Cui and Na Wang and Tao Jiang}, doi = {10.1016/j.neucom.2025.129842}, journal = {Neurocomputing}, month = {6}, pages = {129842}, shortjournal = {Neurocomputing}, title = {Conv-enhanced transformer and robust optimization network for robust multimodal sentiment analysis}, volume = {634}, year = {2025}, }
• (2025). Memo-UNet: Leveraging historical information for enhanced wave height prediction. NEUCOM, 634, 129840. (www)

  @article{ , abstract = {Wave height prediction is a challenging spatiotemporal modeling task. It requires accurately capturing the dynamic evolution patterns of historical data. However, current methods show significant limitations in storing, retrieving, and utilizing historical information. These limitations hinder the learning ability of neural networks in critical temporal dynamics utilization of wave height prediction. Therefore, we propose a novel neural network Meme-Unet for wave height prediction. Specifically, to address the challenges in storing and retrieving historical information, we design a Memo module that adaptively stores historical data and retrieves key historical information through a minimum distance constraint. Additionally, to enhance the utilization effectiveness of historical information, we designed a TimeEncoding module with attention mechanisms to guide the model in better capturing temporal relationships. We conducted wave height prediction experiments in two different ocean regions, achieving the best MSE of 0.041 and a 6.8% improvement compared to state-of-the-art methods. This shows the advantage of the Memo-UNet’s design in capturing the dynamics of spatiotemporal modeling.}, archive = {J_NEUCOM}, author = {Teng Fang and Xiaojie Li and Canghong Shi and Xian Zhang and Wei Xiao and Yi Kou and Imran Mumtaz and Zhan ao Huang}, doi = {10.1016/j.neucom.2025.129840}, journal = {Neurocomputing}, month = {6}, pages = {129840}, shortjournal = {Neurocomputing}, title = {Memo-UNet: Leveraging historical information for enhanced wave height prediction}, volume = {634}, year = {2025}, }
• (2025). Hard-label adversarial attack with dual-granularity optimization on texts. NEUCOM, 634, 129839. (www)

  @article{ , abstract = {The advancement of artificial intelligence security research has led to the emergence of adversarial attack technology as a critical approach for identifying potential security vulnerabilities in artificial intelligence models. When targeting natural language processing models, conducting adversarial attacks in the hard-label setting presents a more practical and challenging black-box scenario due to the difficulty in computing gradients directly from discrete word sequences. Current textual adversarial attack methods are inefficient due to the lack of consideration for the limited number of queries available during the adversarial text generation process, creating a disparity between these approaches and real-world adversarial attack scenarios. To this end, this work proposes a dual-granularity optimization strategy that consists of a single-word semantic optimization and a multi-word joint semantic optimization procedure, and presents a query-efficient hard-label attack method called DualAttack by incorporating the proposed dual-granularity optimization strategy into the mutation and crossover process of the Genetic Algorithm framework. Extensive experimental results demonstrate that DualAttack can effectively produce high-quality adversarial texts with superior semantic similarity and minimal perturbation rates within fewer queries compared to existing methods in the hard-label setting.}, archive = {J_NEUCOM}, author = {Shilin Qiu and Qihe Liu and Shijie Zhou and Min Gou and Zhun Zhang and Zhewei Wu}, doi = {10.1016/j.neucom.2025.129839}, journal = {Neurocomputing}, month = {6}, pages = {129839}, shortjournal = {Neurocomputing}, title = {Hard-label adversarial attack with dual-granularity optimization on texts}, volume = {634}, year = {2025}, }
• (2025). Momentum gradient-based untargeted poisoning attack on hypergraph neural networks. NEUCOM, 634, 129835. (www)

  @article{ , abstract = {Hypergraph Neural Networks (HGNNs) have been successfully applied in various hypergraph-related tasks due to their excellent higher-order representation capabilities. Unfortunately, recent works have shown deep learning models vulnerable to diverse attacks. Most studies of attacks on graphs have focused on Graph Neural Networks (GNNs), and the study of attacks on HGNNs remains largely unexplored. In this paper, we try to bridge this gap. We design a new untargeted poisoning attack for HGNNs, MGHGA, which focuses on modifying node features. We consider the process of HGNNs training and use a surrogate model to implement the attack before hypergraph modeling. Precisely, MGHGA consists of two parts: feature selection and feature modification. We use a momentum gradient mechanism to choose the attack node features in the feature selection module. In the feature modification module, we use two feature generation approaches (direct modification and sign gradient) to enable MGHGA to be employed on discrete and continuous datasets. We conduct extensive experiments on seven benchmark datasets to validate the attack performance of MGHGA in the node and the visual object classification tasks. The results show that MGHGA improves performance by an average of 2% compared to the baselines.}, archive = {J_NEUCOM}, author = {Yang Chen and Stjepan Picek and Zhonglin Ye and Zhaoyang Wang and Haixing Zhao}, doi = {10.1016/j.neucom.2025.129835}, journal = {Neurocomputing}, month = {6}, pages = {129835}, shortjournal = {Neurocomputing}, title = {Momentum gradient-based untargeted poisoning attack on hypergraph neural networks}, volume = {634}, year = {2025}, }
• (2025). Multiple-level enhanced graph convolutional network for aspect sentiment triplet extraction. NEUCOM, 634, 129834. (www)

  @article{ , abstract = {Aspect Sentiment Triplet Extraction (ASTE) is a method for extracting aspect terms, opinion terms, and their corresponding sentiment polarities from a given sentence. Most of the existing studies use joint extraction methods to extract the triplets directly in a unified framework. However, most joint extraction methods only consider the semantic and syntactic dependency information of the sentence. Due to a lack of sentiment information and positional information, they are unable to accurately and completely express the aspect and opinion in the sentence. In order to solve the above problems, we introduce a Multiple-level Enhanced Graph Convolutional Network (MEGCN) for ASTE, which utilizes sentiment scores and sentiment polarity nodes alongside syntactic dependency information. This approach not only enriches contextual understanding by integrating sentiment data but also improves positional analysis of aspect and opinion terms through polarity nodes. Moreover, our dual-aware fusion module, combining semantic with sentiment-enhanced syntactic features through a biaffine attention mechanism and matrix construction, enables a deeper representation of aspect sentiment triplets. Our model demonstrates superior performance over existing methods on two widely recognized ASTE datasets.}, archive = {J_NEUCOM}, author = {Haowen Xu and Mingwei Tang and Tao Cai and Jie Hu and Zhongyuan Jiang and Deng Bian and Shixuan Lv}, doi = {10.1016/j.neucom.2025.129834}, journal = {Neurocomputing}, month = {6}, pages = {129834}, shortjournal = {Neurocomputing}, title = {Multiple-level enhanced graph convolutional network for aspect sentiment triplet extraction}, volume = {634}, year = {2025}, }
• (2025). Dense attention networks for texture classification. NEUCOM, 634, 129833. (www)

  @article{ , abstract = {Texture plays a fundamental role in visual perception and is a key feature for classifying materials and surfaces in images. However, due to the limited number of training samples for each texture class, current deep learning-based methods struggle to capture diverse texture patterns effectively. To address this challenge, we propose a Dense Attention Networks (DANets), a novel architecture for texture classification that leverages both attention mechanisms and dense connectivity. DANet aims to enhance the ability to capture various texture patterns by constructing a mixed-scale, densely connected squeeze-and-excitation attention network. Additionally, we introduce a new Texture Loss function that incorporates dynamic regularization to improve robustness and generalization. To further boost performance, we propose Trident Feature Fusion (TFF), a simple yet effective method for fusing feature vectors from shallow, middle, and deep layers, enabling more comprehensive texture representation. Experimental results on six benchmark texture datasets demonstrate that DANets outperform 26 representative methods, showing their effectiveness in improving texture classification accuracy.}, archive = {J_NEUCOM}, author = {Yongsheng Dong and Naixin Lu and Xuelong Li}, doi = {10.1016/j.neucom.2025.129833}, journal = {Neurocomputing}, month = {6}, pages = {129833}, shortjournal = {Neurocomputing}, title = {Dense attention networks for texture classification}, volume = {634}, year = {2025}, }
• (2025). Practical incremental learning: Striving for better performance-efficiency trade-off. NEUCOM, 634, 129831. (www)

  @article{ , abstract = {We revisit the storage of historical data and buffer management in replay-based Incremental Learning (IL), where the buffer is usually a complete copy of the storage. Storage or buffer size constraints present significant challenges in laboratory settings but may not be practical in real-world applications. Since storing historical data is generally feasible and affordable in most realistic scenarios, we propose a generalization to IL by decoupling the concepts of storage and buffer, dubbed Practical Incremental Learning (PIL). We argue that, in practice, storage size should be determined by cost considerations, while buffer size should be limited for efficiency. Under PIL, we establish a strong baseline by resampling the buffer for each training iteration, ensuring better historical data coverage. We then evaluate popular strategies within the IL community, revealing their ineffectiveness against this baseline. We identify key limitations through in-depth analysis and derive insightful design principles for PIL methods. Based on these principles, Cross-Task Distance Calibration with Distribution Annealing (DCDA) is proposed. It achieves significant improvements over state-of-the-art methods across multiple benchmarks and settings, with performance gains of up to 15%, demonstrating its industrial-level performance and usability.}, archive = {J_NEUCOM}, author = {Shixiong Xu and Bolin Ni and Xing Nie and Fei Zhu and Yin Li and Jianlong Chang and Gaofeng Meng}, doi = {10.1016/j.neucom.2025.129831}, journal = {Neurocomputing}, month = {6}, pages = {129831}, shortjournal = {Neurocomputing}, title = {Practical incremental learning: Striving for better performance-efficiency trade-off}, volume = {634}, year = {2025}, }
• (2025). Enhancing generalization in camera trap image recognition: Fine-tuning visual language models. NEUCOM, 634, 129826. (www)

  @article{ , abstract = {This study introduces a novel fine-tuning approach for enhancing the generalization capabilities of visual language models in the context of wildlife monitoring, particularly for camera trap image recognition. In this paper, we introduce Ecological Visual Language Models (Eco-VLMs), a model fine-tuned using an ecological subset of the ImageNet1K dataset (ImageNet1K-E), aimed at reducing the reliance on spurious correlations that affect the performance of models like CLIP when applied to specialized domains. By employing text augmentation techniques and expanding species names with rich descriptors, Eco-VLM is optimized to extract more distinctive features from images, thereby improving its discriminative capabilities for wildlife features. Meanwhile, random contrastive loss is proposed to improve the diversity of training data and the generalization of Eco-VLMs. The proposed Eco-CLIP and Eco-SigLIP model are rigorously evaluated against various camera trap datasets and demonstrates superior performance, with average F1 scores improved by 4.44 % and 3.79 % compared to the standard CLIP and SigLIP model. Intrinsic evaluations further confirm that Eco-VLMs have acquired a broader ecological knowledge base, highlighting its enhanced generalization abilities. This research contributes to the field by addressing the limitations of current visual language models in specialized ecological applications and underscores the potential of Eco-VLMs for improving wildlife monitoring efforts.}, archive = {J_NEUCOM}, author = {Zihe Yang and Ye Tian and Lifeng Wang and Junguo Zhang}, doi = {10.1016/j.neucom.2025.129826}, journal = {Neurocomputing}, month = {6}, pages = {129826}, shortjournal = {Neurocomputing}, title = {Enhancing generalization in camera trap image recognition: Fine-tuning visual language models}, volume = {634}, year = {2025}, }
• (2025). Training multi-bit spiking neural network with virtual neurons. NEUCOM, 634, 129825. (www)

  @article{ , abstract = {Back-Propagation Through Time (BPTT) with surrogate gradient has been widely applied for directly training the Spiking Neural Network (SNN), a biologically realistic type of Artificial Neural Network. The Multi-Level Firing model (MLF) improves SNN performance by introducing more sub-neurons in a unit and extending the output of spiking neurons from single-bit binary impulse to multiple-bit integers. However, it also worsens the excessive computational resources demanded by the BPTT process. This paper presents a novel neuron model called Virtual Neurons(VN) and its training method. VN can output multi-bit spikes as MLF does but avoid its usually heavy training cost. The VN has been proven numerically equivalent to a special case of several single-bit sub-neurons combined in the MLF way, making it theoretically compatible with 1-bit neuromorphic chips. Experimental results on various vision tasks demonstrate that VN-based SNNs not only have much lower computation costs compared to the MLF model during off-chip training but also achieve advanced performance in fewer time steps.}, archive = {J_NEUCOM}, author = {Haoran Xu and Zonghua Gu and Ruimin Sun and De Ma}, doi = {10.1016/j.neucom.2025.129825}, journal = {Neurocomputing}, month = {6}, pages = {129825}, shortjournal = {Neurocomputing}, title = {Training multi-bit spiking neural network with virtual neurons}, volume = {634}, year = {2025}, }
• (2025). Enhancing sentiment analysis with distributional emotion embeddings. NEUCOM, 634, 129822. (www)

  @article{ , abstract = {Sentiment classification tasks, such as emotion detection and sentiment analysis, are essential in modern natural language processing (NLP). Moreover, vector representation frameworks modeling semantic content underlie each state-of-the-art NLP algorithmic scheme. In sentiment classification, traditional methods often rely on such embedding vectors for semantic representation, yet they typically overlook the dynamic and sequential nature of emotions within textual data. In this work, we present a novel methodology that leverages the distributional patterns of emotions. An embedding framework that captures the inherent serial structure of emotional occurrences in text is introduced, modeling the interdependencies between emotion states as they unfold within a document. Our approach treats each sentence as an observation in a multivariate series of emotions, transforming the emotional flow of a text into a sequence of emotion strings. By applying distributional logic, emotion-based embeddings that represent both emotional and semantic information are derived. Through a comprehensive experimental framework, we demonstrate the effectiveness of these embeddings across various sentiment-related tasks, including emotion detection, irony identification, and hate speech classification, evaluated on multiple datasets. The results show that our distributional emotion embeddings significantly enhance the performance of sentiment classification models, offering improved generalization across diverse domains such as financial news and climate change discourse. Hence, this work highlights the potential of using distributional emotion embeddings to advance sentiment analysis, offering a more nuanced understanding of emotional language and its structured, context-dependent manifestations.}, archive = {J_NEUCOM}, author = {Charalampos M. Liapis and Aikaterini Karanikola and Sotiris Kotsiantis}, doi = {10.1016/j.neucom.2025.129822}, journal = {Neurocomputing}, month = {6}, pages = {129822}, shortjournal = {Neurocomputing}, title = {Enhancing sentiment analysis with distributional emotion embeddings}, volume = {634}, year = {2025}, }
• (2025). A novel multi-scale salient object detection framework utilizing nonlinear spiking neural p systems. NEUCOM, 634, 129821. (www)

  @article{ , abstract = {Salient object detection (SOD) is fundamental to computer vision applications ranging from autonomous driving and surveillance to medical imaging. Despite significant progress, existing methods struggle to effectively model multi-scale features and their complex interdependencies, particularly in challenging real-world scenarios with complex backgrounds and varying scales. To address these limitations, this paper proposes a novel detection framework that leverages the hierarchical processing capabilities of nonlinear spiking neural P (NSNP) systems. The proposed framework introduces three key innovations: a bio-inspired convolution mechanism that captures fine-grained local features with neural dynamics; a semantic learning module enhanced by Contextual Transformer Attention for comprehensive global context understanding; and an adaptive mixed attention-based fusion strategy that optimizes cross-scale feature integration. The experimental results on four challenging benchmark datasets demonstrate that the proposed method outperforms fourteen other state-of-the-art methods, achieving average improvements of 1.02%, 1.3%, 2.3%, and 0.1% on the four evaluation metrics ( S m , E ξ m , F β w , and M A E ), respectively. These advances validate the potential of spiking neural P systems in salient object detection, while opening new possibilities for bio-inspired approaches in visual computing.}, archive = {J_NEUCOM}, author = {Nan Zhou and Minglong He and Hong Peng and Zhicai Liu}, doi = {10.1016/j.neucom.2025.129821}, journal = {Neurocomputing}, month = {6}, pages = {129821}, shortjournal = {Neurocomputing}, title = {A novel multi-scale salient object detection framework utilizing nonlinear spiking neural p systems}, volume = {634}, year = {2025}, }
• (2025). Local and global Spatial–Temporal transformer for skeleton-based action recognition. NEUCOM, 634, 129820. (www)

  @article{ , abstract = {Skeleton-based action recognition represents a dynamic and expanding research domain in computer vision. Currently, GCN-based methods primarily rely on the graph topology to capture dependencies between joints, however, they are limited in the ability to capture long-distance dependencies. On the other hand, transformer-based methods have also been applied to skeleton-based action recognition, as transformers prove effective in capturing long-distance dependencies. Nevertheless, most transformer-based methods directly calculate pairwise global self-attention of all nodes in both spatial and temporal dimensions, making it challenging to distinguish the correlation between short-distance joints and underestimate the impact of short-term temporal dynamics. Additionally, some existing methods often utilize multi-stream fusion to combine features from different modalities, neglecting the fusion of low-level information from these modalities. In this work, we propose a novel L ocal and G lobal S patial– T emporal Trans former network (LG-STFormer) containing two key components: (1) LGA-module: local and global attention module. The LGA-module enables the model to capture richer temporal and spatial information. It consists of two parts: skeleton topology constraint spatial transformer (STC-SFormer) and attention-enhanced multiscale TCN (AM-TCN). The STC-SFormer focuses on the correlation between local joints and distant joints in the spatial dimension, while the AM-TCN integrates the global self-attention mechanism into multi-scale temporal convolution to capture local and global temporal motion patterns of joints effectively. (2) JBC-Fusion-module: The module consists of joint-bone cross fusion transformer (JBC-Former) and AM-TCN. Utilizing dynamic generation method for complementary features, the JBC-Former facilitates the fusion of low-level information between complementary features. Finally, we make extensive experiments on the NTU-RGB+D, NTU-RGB+D 120, and NW-UCLA datasets to show the competitive performance of the proposed LG-STFormer in the field of skeleton-based action recognition.}, archive = {J_NEUCOM}, author = {Ruyi Liu and Yu Chen and Feiyu Gai and Yi Liu and Qiguang Miao and Shuai Wu}, doi = {10.1016/j.neucom.2025.129820}, journal = {Neurocomputing}, month = {6}, pages = {129820}, shortjournal = {Neurocomputing}, title = {Local and global Spatial–Temporal transformer for skeleton-based action recognition}, volume = {634}, year = {2025}, }
• (2025). GKA: Graph-guided knowledge association for fine-grained visual categorization. NEUCOM, 634, 129819. (www)

  @article{ , abstract = {Fine-grained visual categorization aims to distinguish highly similar subclasses by exploiting subtle differences. However, existing methods predominantly emphasize the extraction of visual cues from individual images, overlooking the exploration of semantic relationships within and across classes. To this end, we introduce a novel approach termed Graph-based Knowledge Association (GKA). Specifically, we employ a positional embedding to model the relationship between instances in the feature space, and adaptively mine the connections between features of different instances via a graph neural network. The framework effectively aggregates features from neighboring nodes to enhance the understanding of discriminative features by exploiting complementary information between instances. Furthermore, a plain knowledge-guided module embeds this relational knowledge into the training of the backbone network for discriminative feature extraction, thus improving fine-grained classification performance. Empirical evaluations on four benchmark datasets for Fine-grained Visual Categorization (FGVC) demonstrate that our method achieves state-of-the-art (SOTA) performance.}, archive = {J_NEUCOM}, author = {Yuetian Wang and Shuo Ye and Wenjin Hou and Duanquan Xu and Xinge You}, doi = {10.1016/j.neucom.2025.129819}, journal = {Neurocomputing}, month = {6}, pages = {129819}, shortjournal = {Neurocomputing}, title = {GKA: Graph-guided knowledge association for fine-grained visual categorization}, volume = {634}, year = {2025}, }
• (2025). STAD-AI: Spatio-temporal anomaly detection in videos with attentive dual-stage integration. NEUCOM, 634, 129817. (www)

  @article{ , abstract = {This paper presents a novel and comprehensive framework for video anomaly detection, distinguished by its specialized spatio-temporal feature extraction and precise anomaly prediction capabilities. The proposed system employs an advanced spatio-temporal attention-based framework designed for effective video frame reconstruction. By isolating and amplifying critical feature regions within the frames, it enables the extraction of fine-grained spatial and temporal representations, which are crucial for detecting subtle anomalies. Complementing this, an attentive U-Net architecture is employed to predict anomalies with high precision, incorporating motion features to enhance temporal coherence and anomaly localization. The attention mechanism in both components is strategically designed to focus on critical areas within each frame and sequence, where abnormal activities are likely to occur, improving detection accuracy and reducing false positives. The two components are seamlessly integrated using a fusion strategy, combining their complementary strengths to enhance the system’s overall robustness and effectiveness. Extensive evaluations on benchmark datasets, including UCSD Peds1, UCSD Peds2, CUHK Avenue, and ShanghaiTech, demonstrate that STAD-AI achieves superior performance with AUC scores of 86.6%, 99.1%, 91.4%, and 77.7%, respectively. These results highlight the framework’s ability to effectively leverage spatial and temporal features for detecting anomalies with high precision, advancing the state-of-the-art in video anomaly detection.}, archive = {J_NEUCOM}, author = {Rangachary Kommanduri and Mrinmoy Ghorai}, doi = {10.1016/j.neucom.2025.129817}, journal = {Neurocomputing}, month = {6}, pages = {129817}, shortjournal = {Neurocomputing}, title = {STAD-AI: Spatio-temporal anomaly detection in videos with attentive dual-stage integration}, volume = {634}, year = {2025}, }
• (2025). Fractional gannet humming optimization enabled deep convolutional neural network for detection and segmentation of skin cancer. NEUCOM, 634, 129816. (www)

  @article{ , abstract = {Skin cancer is a dangerous disorder that is caused by an unchecked proliferation of aberrant skin cells that produce genetic mutations on the skin. When ultraviolet (UV) radiation from sunshine or tanning beds damages the skin cells in a way that leads to their rapid multiplication and formation of malignant tumours. Certain forms of skin cancer metastasize along nerves. This results in tingling, pain, itching, numbness, or a sensation like ants crawling under the skin. Skin cancer spreads to deeper tissues, including cartilage, muscle, and bone. The earlier prediction of skin lesions increases the possibility of survival rate. However, during diagnosis, an anomalous finding is made, and the condition is diagnosed as cancer. To overcome these challenges, a novel deep learning (DL) technique is developed in this research article for categorizing skin cancer by employing the proposed Fractional Gannet Humming Optimization_Deep Convolutional Neural Network (FGHO_DeepCNN). Initially, the input skin cancer image is subjected to the image pre-processing phase. The image pre-processing is done by the bilateral filter. Afterwards, skin lesion segmentation is carried out using an encoder-decoder with Dense-Residual block (DRB), which is trained by the Fractional Gannet optimization algorithm (FGOA). Here, the FGOA is formed by the integration of the Fractional Calculus (FC) concept with the Gannet Optimization Algorithm (GOA). Thereafter, image augmentation is done to enlarge the segmented image using geometric and colour space transformation. After that, a feature extraction process is conducted to obtain the significant features, like Completed Local Binary Pattern (CLBP), Gray Level Co-occurrence Matrix (GLCM), Local Vector Pattern (LVP), Significant Local Binary Pattern (SLBP) and CNN features. Finally, skin cancer detection is done using DeepCNN, which is tuned by the proposed FGHO. Here, the proposed FGHO is formed by the combination of FGOA with the artificial hummingbird algorithm (AHA). The experimental outcomes of the proposed FGHO_DeepCNN approach attained a better Positive Predictive Value (PPV), Negative Predictive Value (NPV), True Positive Rate (TPR), and True Negative Rate (TNR), and accuracy with values of 89.80 %, 89.40 %, 94.50 %, 94.00 % and 93.40 % respectively. The employed FGHO_DeepCNN has acquired excellent performance, thus achieving a PPV of 91.68 %, NPV of 88.46 %, TPR of 91.68 %, TNR of 91.23 % and accuracy of 90.67 % for dataset 2. In dataset 3, the FGHO_DeepCNN obtained superior performance than other techniques with a PPV of 90.56 %, NPV of 90.36 %, TPR of 90.95 %, TNR of 90.87 %, and accuracy of 90.15 %. The practical significance of the proposed FGHO_DeepCNN approach is that it is widely used in dermatology clinics and hospitals to detect skin cancer.}, archive = {J_NEUCOM}, author = {Aravapalli Rama Satish and Balajee Maram and Varaprasada Rao Perumalla and Mallikharjuna Rao K}, doi = {10.1016/j.neucom.2025.129816}, journal = {Neurocomputing}, month = {6}, pages = {129816}, shortjournal = {Neurocomputing}, title = {Fractional gannet humming optimization enabled deep convolutional neural network for detection and segmentation of skin cancer}, volume = {634}, year = {2025}, }
• (2025). Fairness in constrained spectral clustering. NEUCOM, 634, 129815. (www)

  @article{ , abstract = {Semi-supervised clustering methods have gained significant attention in both theoretical research and real-world applications, including economics, finance, marketing, and healthcare. Among these methods, constrained spectral clustering enhances clustering quality by incorporating pairwise constraints, namely, must-link and cannot-link constraints, which guide the clustering process by specifying whether certain data points should or should not belong to the same cluster. However, traditional constrained spectral clustering methods may inadvertently propagate biases present in the data or constraints, leading to unequal representation of sensitive groups, such as different genders or racial groups, across clusters. This imbalance raises concerns about fairness, an issue that remains largely unexplored in constrained spectral clustering. To address this gap, this paper proposes a novel method named fair-constrained Spectral Clustering (fair-cSC). The proposed method integrates fairness into the must-link and cannot-link constraints by defining a fair constraint matrix, ensuring that pairwise relationships do not introduce bias against any particular group. Additionally, a balance constraint is incorporated to enforce fairness across input data points, promoting equal representation of sensitive groups within clusters. Comprehensive experiments on six benchmarked datasets, including ablation studies, demonstrate that the proposed fair-cSC method effectively enhances fairness while preserving clustering quality. Furthermore, the ablation study provides insights into the method’s performance under different settings, reinforcing its robustness and applicability in real-world scenarios.}, archive = {J_NEUCOM}, author = {Laxita Agrawal and V. Vijaya Saradhi and Teena Sharma}, doi = {10.1016/j.neucom.2025.129815}, journal = {Neurocomputing}, month = {6}, pages = {129815}, shortjournal = {Neurocomputing}, title = {Fairness in constrained spectral clustering}, volume = {634}, year = {2025}, }
• (2025). Text-driven human image generation with texture and pose control. NEUCOM, 634, 129813. (www)

  @article{ , abstract = {Generating high-quality and diverse human images solely based on textual prompts is a challenging task in computer vision and computer graphics. Existing models do not have stable generation capabilities, and their control over image generation is insufficient, requiring an additional densepose map to provide pose information. In this work, we present a text-driven human image generation model, which enables control over texture and pose through textual input. The framework consists of three stages: (1) generating a pose map from the pose description which achieves control over the pose information of the generated human image. (2) converting the pose map to a parsing map based on the shape description, and (3) generating the final image by using textual descriptions of clothes texture along with a quality enhancement network. For the texture learning stage, we construct a texture-aware codebook using FSQ, learning vector representations for different clothes fabrics and patterns. Furthermore, we fine-tune the diffusion model to equip it with the capability of image denoising and image refining, thus improving the quality of the generated images. Extensive experimental results demonstrate our model’s ability of generating realistic human body images that align with the semantics of the input text.}, archive = {J_NEUCOM}, author = {Zhedong Jin and Guiyu Xia and Paike Yang and Mengxiang Wang and Yubao Sun and Qingshan Liu}, doi = {10.1016/j.neucom.2025.129813}, journal = {Neurocomputing}, month = {6}, pages = {129813}, shortjournal = {Neurocomputing}, title = {Text-driven human image generation with texture and pose control}, volume = {634}, year = {2025}, }
• (2025). Advancing DoA assessment through federated learning: A one-shot pseudo data approach. NEUCOM, 634, 129812. (www)

  @article{ , abstract = {Accurately measuring the Depth of Anaesthesia (DoA) during surgical procedures is crucial for patient safety. A significant challenge in developing effective machine learning models for DoA assessment is the lack of data from single organisations and preserving data privacy between institutions. Federated learning offers a solution by enabling multiple parties to collaboratively train models without exchanging data. However, traditional federated learning algorithms perform poorly in data heterogeneous, non-identically distributed data distribution scenarios. To address these challenges, we propose a one-shot federated learning framework, DoAFedP-NN, which facilitates federated learning with heterogeneous model development. The framework is tested in a range of model and data heterogeneity environments. This method enables the training of a global DoA prediction model across different medical facilities without sharing local data. The DoAFedP-NN model, utilising neural network design with entropy and spectral feature extraction, is compared to benchmark federated learning architectures, demonstrating its advantage in handling heterogeneous medical data. Experimental results show that DoAFedP-NN achieves robust DoA estimation when compared to the Bispectral (BIS) index, with high correlation coefficients of 0.8472 and 0.8542 across independent databases. The proposed model outperforms locally developed models, showing significant improvements when validated against external datasets from different medical facilities. This paper makes the key contributions: (1) introduces a one-shot pseudo-data method for federated learning; (2) demonstrates the effectiveness of this approach for EEG-based DoA using real-world databases; (3) showcases the model’s ability to achieve high correlation with the BIS index while preserving patient privacy in a range of client distribution scenarios and under cross-validation.}, archive = {J_NEUCOM}, author = {Thomas Schmierer and Tianning Li and Di Wu and Yan Li}, doi = {10.1016/j.neucom.2025.129812}, journal = {Neurocomputing}, month = {6}, pages = {129812}, shortjournal = {Neurocomputing}, title = {Advancing DoA assessment through federated learning: A one-shot pseudo data approach}, volume = {634}, year = {2025}, }
• (2025). HCKGL: Hyperbolic collaborative knowledge graph learning for recommendation. NEUCOM, 634, 129808. (www)

  @article{ , abstract = {Recently, the integration of knowledge graph and recommendation system has become a hot topic. Its popular solution is firstly combining the knowledge graph and user–item interaction graph to generate a unified Collaborative Knowledge Graph (CKG), and then learn the representations of users and items by applying graph convolutional networks to aggregate high-order neighbor information between entities in CKG. However, existing related methods mainly focus on learning representations in the Euclidean space, posing challenges in capturing the hierarchical structure and intricate relational logic between users and items. In view of this, we propose a novel hyperbolic CKG learning model HCKGL for recommendation, which leverages relation-specific curvature and attention-based geometric transformations to preserve the inherent features of CKG. Additionally, we address two significant challenges that existing methods have often overlooked. Firstly, in order to capture the relationship dependencies between neighbors and accurately calculate the contribution of neighbor information, we propose a hyperbolic graph attention network (HGAT), which combines the curvature of the relationship to assign weights. Secondly, we present a new graph contrastive learning technique (HMCL) that utilizes the hyperbolic embedding propagation and multi-level contrastive learning to improve the representations of users and items. Comprehensive experimental results on two widely used datasets demonstrate that HCKGL outperforms state-of-the-art baselines. The source code for our model is publicly available at: https://github.com/GDM-SCNU/HCKGL .}, archive = {J_NEUCOM}, author = {Huijuan Hu and Chaobo He and Xinran Chen and Quanlong Guan}, doi = {10.1016/j.neucom.2025.129808}, journal = {Neurocomputing}, month = {6}, pages = {129808}, shortjournal = {Neurocomputing}, title = {HCKGL: Hyperbolic collaborative knowledge graph learning for recommendation}, volume = {634}, year = {2025}, }
• (2025). Feature expansion and enhanced compression for class incremental learning. NEUCOM, 634, 129807. (www)

  @article{ , abstract = {Class incremental learning consists in training discriminative models to classify an increasing number of classes over time. However, doing so using only the newly added class data leads to the known problem of catastrophic forgetting of the previous classes. Recently, dynamic deep learning architectures have been shown to exhibit a better stability-plasticity trade-off by dynamically adding new feature extractors to the model in order to learn new classes followed by a compression step to scale the model back to its original size, thus avoiding a growing number of parameters. In this context, we propose a new algorithm that enhances the compression of previous class knowledge by cutting and mixing patches of previous class samples with the new images during compression using our Rehearsal-CutMix method. We show that this new data augmentation reduces catastrophic forgetting by specifically targeting past class information and improving its compression. Extensive experiments performed on the CIFAR and ImageNet datasets under diverse incremental learning evaluation protocols demonstrate that our approach consistently outperforms the state-of-the-art . The code will be made available upon publication of our work 1 .}, archive = {J_NEUCOM}, author = {Quentin Ferdinand and Benoit Clement and Panagiotis Papadakis and Quentin Oliveau and Gilles Le Chenadec}, doi = {10.1016/j.neucom.2025.129807}, journal = {Neurocomputing}, month = {6}, pages = {129807}, shortjournal = {Neurocomputing}, title = {Feature expansion and enhanced compression for class incremental learning}, volume = {634}, year = {2025}, }
• (2025). A bioinspired model of decision making guided by reward dimensions and a motivational state. NEUCOM, 634, 129806. (www)

  @article{ , abstract = {The decision-making process is a critical component of computational systems, whose processing involves the evaluation of various alternatives presented as possible solutions to a given problem, depending on the current context. This paper seeks to show how a neuroscience-based decision-making mechanism (DMM) integrating decision criteria, knowledge of reward stimuli, and motivational information helps to contribute to producing human-like adaptive behavior. To fulfill this objective, a computational model on DMM is proposed. The alternatives in this proposed model are constructed based on preferences, and the selection of the best alternative is guided by a goal-directed control scheme influenced by a motivational state (MS). The formation of preferences considers some dimensions of the reward, e.g., magnitude, probability of receiving the reward, incentive salience, and affective value. To validate the model exhibits a behavior considering parameters human being uses to compute its behavior, a case study was proposed. The case study’s objective is to gain the maximum reward (food) from the choice of a 4-choice card (a variation of Iowa Gambling Test), each card has a reward and a contingency probability associated with it. The analysis of the results of the case study shows that the model presents a short exploitation stage to find the contingency rule and choose the best option frequently according to some studies, also observed that the utility value of the card influenced the MS of hunger and other factors play a critical role in the DMM.}, archive = {J_NEUCOM}, author = {Diana G. Gómez-Martínez and Alison Muñoz-Capote and Oscar Hernández and Francisco Robles and Félix Ramos}, doi = {10.1016/j.neucom.2025.129806}, journal = {Neurocomputing}, month = {6}, pages = {129806}, shortjournal = {Neurocomputing}, title = {A bioinspired model of decision making guided by reward dimensions and a motivational state}, volume = {634}, year = {2025}, }
• (2025). SSDM: Generated image interaction method based on spatial sparsity for diffusion models. NEUCOM, 634, 129805. (www)

  @article{ , abstract = {Image interaction methods based on diffusion models are significantly superior to traditional methods. However, due to its slow sampling speed and high computational complexity, it could be more conducive to image editing applications. To address these issues, we propose a generated image interaction method based on spatial sparsity diffusion models (SSDM), which utilizes the spatial sparsity of the differences between the edited and original images to reduce computational complexity and accelerate image generation. It takes sparse block data as a constraint, uses difference masks, converts it into an index, and learns the spatial sparse features of differences to describe the image, thereby reducing the network parameters and computational complexity during the training process. In addition, the “overlap-add” and “overlap-save” mechanisms are used to ensure coherence and consistency between different boundaries. It also compensates and trains the sampled feature maps by reusing low-frequency information and introduces L p -norm to replace Euclidean distance to calculate the loss function, thereby enhancing the reconstruction effect of high-frequency image features. Compared to the original methods, experiments on the LSUN and CelebA-HQ datasets show that the proposed method has achieved better performance of PSNR by improving 1.2 dB, reduced computational complexity by 4.1 × , and increased processing speed by 3.6 × .}, archive = {J_NEUCOM}, author = {Zhuochao Yang and Jingjing Liu and Haozhe Zhu and Jianhua Zhang and Wanquan Liu}, doi = {10.1016/j.neucom.2025.129805}, journal = {Neurocomputing}, month = {6}, pages = {129805}, shortjournal = {Neurocomputing}, title = {SSDM: Generated image interaction method based on spatial sparsity for diffusion models}, volume = {634}, year = {2025}, }
• (2025). Backpropagation-based learning with local derivative approximation and memory replay in biologically plausible neural systems. NEUCOM, 634, 129804. (www)

  @article{ , abstract = {When learning, the brain modifies individual synaptic connections to reach a desired behavior. Animal and human brains have been shown to be incredibly capable of learning complex and varied functions across a wide variety of tasks. In recent years, artificial neural networks, inspired by human and animal brains, have shown great capabilities in learning a wide variety of difficult tasks. However, artificial neural networks primarily teach themselves through the use of backpropagation, a learning method which has no clear analogue within the brain. Additionally, Artificial Neural Networks primarily use continuous activation functions, which differ significantly from the spiking neuronal behavior present in the brain. In this paper, we discuss and demonstrate a biologically plausible learning method that approximates backpropagation through two techniques on Spiking Neural Networks. First, we show that the local temporal derivatives that are necessary for backpropagation can be approximately recovered through reconstruction using spike timings. Second, we show that through learning during a sleep phase, inspired by neuroscience research into memory replay, the localized parallel feedback path can learn to approximate the derivative through the forward path weight matrix, thus solving the weight transport problem. Finally, we demonstrate that the combination of these two methods can approach or exceed the accuracy of backpropagation-based methods for a variety of neuromorphic vision tasks while maintaining biological plausibility.}, archive = {J_NEUCOM}, author = {Richard Boone and Peng Li}, doi = {10.1016/j.neucom.2025.129804}, journal = {Neurocomputing}, month = {6}, pages = {129804}, shortjournal = {Neurocomputing}, title = {Backpropagation-based learning with local derivative approximation and memory replay in biologically plausible neural systems}, volume = {634}, year = {2025}, }
• (2025). Data-driven stabilization for linear sampled-data systems with unknown parameters: A pure data analytics perspective. NEUCOM, 634, 129798. (www)

  @article{ , abstract = {This paper investigates the sampled-data stabilization problem for a class of linear systems with unknown parameters, incorporating a data analytics perspective. Traditional sampled-data controller design methods are rendered ineffective due to the lack of knowledge about the system parameters. To address this challenge, a data-driven approach is proposed for designing a sampled-data stabilization controller using an offline-collected data set, eliminating the need for prior knowledge of the plant parameters. Specifically, a novel sampled-data control strategy with a matrix exponential gain is introduced to stabilize the system. The controller gain is designed based on data insights derived from offline information, without requiring access to the system matrices. Leveraging data-driven techniques and Lyapunov stability theory, sufficient conditions are established to guarantee the exponential stability of the system. The proposed sampled-data controller with matrix exponential gain significantly extends the maximum allowable sampling intervals compared to conventional methods. To optimize performance, a convex optimization approach is utilized to maximize the bounds of these sampling intervals. The efficacy of the proposed approach is validated through a practical example involving an operational amplifier circuit, demonstrating the power of integrating data analytics into control design for unknown-parameter systems.}, archive = {J_NEUCOM}, author = {Luyang Yu and Jiayi Ding and Ying Cui and Yurong Liu and Yamin Wang}, doi = {10.1016/j.neucom.2025.129798}, journal = {Neurocomputing}, month = {6}, pages = {129798}, shortjournal = {Neurocomputing}, title = {Data-driven stabilization for linear sampled-data systems with unknown parameters: A pure data analytics perspective}, volume = {634}, year = {2025}, }
• (2025). A new minimalist time series reduction technique. NEUCOM, 634, 129772. (www)

  @article{ , abstract = {We propose in this paper a minimalist time series reduction technique that comes with a new distance. The proposed technique relies on the extraction of a minimal number of salient features of a time series, which are leveraged by the new distance to perform time series classification. We prove in this paper that the new proposed distance lower bounds the Euclidean distance and has a tighter lower bound than Piecewise Aggregation Approximation. Furthermore, we conduct experiments using standard UCR univariate time series datasets and include a comparative study of the new proposed technique with related SAX and deep learning based time series reduction and classification techniques. The experiments show that the new distance based time series classification enjoys better accuracy results than classifications based on well-known distances and is competitive to deep learning based time series classification techniques.}, archive = {J_NEUCOM}, author = {Hamdi Yahyaoui and Hosam AboElFotoh and Yanjun Shu and Tianrun Gao}, doi = {10.1016/j.neucom.2025.129772}, journal = {Neurocomputing}, month = {6}, pages = {129772}, shortjournal = {Neurocomputing}, title = {A new minimalist time series reduction technique}, volume = {634}, year = {2025}, }
• (2025). Deformable symmetry attention for nuclear medicine image segmentation. NEUCOM, 634, 129757. (www)

  @article{ , abstract = {Prior knowledge of the medical domain has consistently enriched medical image analysis, yet its full potential remains to be explored. Our Deformable Symmetry Attention (DSA) aims to leverage anatomical symmetry prior, commonly used by clinicians, to enhance the automated analysis of medical images, especially in inherently grainy nuclear medicine images. DSA processes an explicit feature alignment attention. It starts with a prior alignment by reflecting the input image to leverage symmetry, followed by a fine-grained alignment using deformable image registration. Our novel Rotational Spatial Transformation Function (RSTF) enhances the registration by estimating local displacement with rotation in relative coordinates, enforcing local smoothness under anatomical variations. Additionally, the Attention Redistribution (AR) module addresses pixel-level mismatch in grainy nuclear medicine images by a set of learnable Gaussian kernels. By translating the anatomical symmetry prior into a locality prior, DSA seamlessly integrates with and enhances sophisticated convolutional neural networks (CNNs) and transformer-based segmentation models. In our study, the DSA framework, when combined with robust baseline models, demonstrated significant performance enhancements across multiple challenging medical imaging tasks. Specifically, it achieved a 3.8% to 8.48% improvement in segmenting lesions in adolescent Whole-Body Bone Scans (WBS) and a 2.88% improvement in the extended 3D MRI brain tumor segmentation experiment. Additionally, the method was validated in asymmetric scenarios, showing an improvement of 0.63% to 10.57% in chest X-ray anatomical segmentation. These results underscore DSA’s effectiveness and versatility across diverse experimental settings. It presents a novel and promising perspective for explicitly leveraging prior knowledge in medical image analysis.}, archive = {J_NEUCOM}, author = {Zeao Zhang and Pei Yang and Ruomeng Liu and Huawei Cai and Zhen Zhao and Quan Guo and Zhang Yi}, doi = {10.1016/j.neucom.2025.129757}, journal = {Neurocomputing}, month = {6}, pages = {129757}, shortjournal = {Neurocomputing}, title = {Deformable symmetry attention for nuclear medicine image segmentation}, volume = {634}, year = {2025}, }
• (2025). Audio–visual self-supervised representation learning: A survey. NEUCOM, 634, 129750. (www)

  @article{ , abstract = {Artificial intelligence developers leverage the inherent relationships among video, text, and audio to create enhanced representations of the world, mirroring the way humans use multiple senses to understand their environment. As such, multimodal learning, which integrates various data input modalities to augment the learning of intrinsic features, has been gaining traction. While applications in multimodal understanding have made strides with deep learning, they often rely heavily on supervised learning and extensive human annotation. This paper provides a comprehensive review of audio–visual self-supervised learning, a promising alternative that uses vast amounts of unlabeled data. It holds the potential to reshape areas like computer vision, and speech recognition. We begin by explaining the concept of audio–visual modalities in machine learning and then move into their role within self-supervised learning by discussing terminology, general pipelines, and underlying motivations. This is followed by an exploration of common pretext tasks in audio–visual self-supervised learning, along with the evaluation methods, datasets, and downstream tasks associated with it. We then highlight prevailing challenges in both audio–visual and self-supervised learning realms. The paper concludes by presenting open challenges, suggesting avenues for future research in this dynamic domain.}, archive = {J_NEUCOM}, author = {Manal AlSuwat and Sarah Al-Shareef and Manal AlGhamdi}, doi = {10.1016/j.neucom.2025.129750}, journal = {Neurocomputing}, month = {6}, pages = {129750}, shortjournal = {Neurocomputing}, title = {Audio–visual self-supervised representation learning: A survey}, volume = {634}, year = {2025}, }
• (2025). Research on self-adaptive grid point cloud down-sampling method based on plane fitting and mahalanobis distance gaussian weighting. NEUCOM, 634, 129746. (www)

  @article{ , abstract = {In this manuscript, a self-adaptive grid point cloud down-sampling method based on plane fitting was proposed, which could effectively reduce redundant data while better preserving the geometric features of the original model and maintaining high accuracy. This method first constructs initial voxel grids and divides the grids into large density and small density ones according to the point cloud density. After that, for small density grids, the boundary points are extracted first, and the rest areas are uniformly sampled, while for large density grids, a method based on Mahalanobis distance Gaussian weighting is proposed and adopted to estimate the normal vector of points, and feature points are determined and retained by calculating the information entropy. Then, three models in the public dataset, the Cat model, Bed_0355 model and Fandisk model, were employed as test subjects to compare the proposed method with two commonly used down-sampling methods: uniform sampling and voxel grid sampling methods. The results indicated that this new method was able to better retain the geometric features of the original models, especially high curvature and sharp parts, with smaller errors and fewer holes. Finally, this method was applied to the down-sampling of 3D scanning point clouds of two typical metal machine parts, threaded joint and sheet metal part, and the measured results demonstrated that this method not only effectively preserved the model features, but also guaranteed accuracy of key geometric dimensions after high reduction ratio down-sampling, such as the relative errors of thread tooth angles and hole inner diameters being less than 1 %.}, archive = {J_NEUCOM}, author = {Hongfei Zu and Jing Zhu and Xinfeng Wang and Xiang Zhang and Ning Chen and Gangxiang Guo and Zhangwei Chen}, doi = {10.1016/j.neucom.2025.129746}, journal = {Neurocomputing}, month = {6}, pages = {129746}, shortjournal = {Neurocomputing}, title = {Research on self-adaptive grid point cloud down-sampling method based on plane fitting and mahalanobis distance gaussian weighting}, volume = {634}, year = {2025}, }
• (2025). LTB-solver: Long-tailed bias solver for image synthesis of diffusion models. NEUCOM, 634, 129651. (www)

  @article{ , abstract = {Though diffusion models have shown the merits of generating high-quality visual data while preserving better diversity in recent studies, they do not generalize well on long-tailed datasets due to the minority classes lacking of diversity and semantic information. To overcome the aforementioned challenges, we first take a closer look at the collapse of tail category patterns under long-tail distributed data and propose an alternative but easy-to-use and effective solution, a L ong- T ailed B ias Solver in diffusion model image synthesis ( LTB-Solver ), which thereby enhances the overall diversity and quality of synthetic samples building upon the properties of the long-tailed distribution training data. Especially, we extract rich generative distribution knowledge of ‘head’ categories within proxy model and transfer the head-tail consistency distance to ‘tail’ categories, enabling the target diffusion model to learn diverse generation preserving inter-sample variation during the diffusion training process. Moreover, we incorporate the minority guidance loss function that better aligns training objectives with sampling behaviors and adjust the loss values for different classes by multiplying them with different weights. Extensive experiments are conducted on various datasets and several state-of-the-art diffusion model frameworks to verify the effectiveness of the proposed method. The results show that our method significantly improves the performance of diffusion models on long-tailed datasets by a large margin.}, archive = {J_NEUCOM}, author = {Siming Fu and Xiaoxuan He and Haoji Hu}, doi = {10.1016/j.neucom.2025.129651}, journal = {Neurocomputing}, month = {6}, pages = {129651}, shortjournal = {Neurocomputing}, title = {LTB-solver: Long-tailed bias solver for image synthesis of diffusion models}, volume = {634}, year = {2025}, }
• (2025). Accurate binary neural network based on rich information flow. NEUCOM, 633, 129837. (www)

  @article{ , abstract = {The binary neural network (BNN) compresses network parameters significantly, greatly reducing storage consumption and resource usage. This also means extremely low inference power consumption, making it a primary method for deploying machine learning on edge devices. However, due to this, there is a significant decrease in accuracy compared to full-precision networks, which restricts its potential for further advancement. To address this challenge, this study introduces the following novel contributions: (1) Introducing an innovative BNN structure named Rich Information Flow Network (RIF-Net) for amplitude correction, which effectively alleviates the information bottleneck in BNN and thus enhances its accuracy. (2) Incorporating additional shortcut connections and Batch Normalization (BN) layers into the architecture to further mitigate the impact of binary convolution on the output feature map's numerical distribution. (3) To demonstrate RIF-Net's performance on edge devices, the study implements the RIF-Net architecture on an FPGA development board. The experimental results show that RIF-Net can achieve the most advanced performance. For example, we achieved a Top-1 accuracy of 63.3 % on ImageNet using the ResNet-18 framework and a 1.0 % absolute accuracy improvement on the CIFAR-10 dataset using the ResNet-20 framework compared to other state-of-the-art methods. The RIF-Net accelerator implemented on the FPGA showcases precise real-time image classification capabilities.}, archive = {J_NEUCOM}, author = {Shilong Zhang and Xiaobo Zhou and Siyuan Chen}, doi = {10.1016/j.neucom.2025.129837}, journal = {Neurocomputing}, month = {6}, pages = {129837}, shortjournal = {Neurocomputing}, title = {Accurate binary neural network based on rich information flow}, volume = {633}, year = {2025}, }
• (2025). A dynamic-static feature fusion learning network for speech emotion recognition. NEUCOM, 633, 129836. (www)

  @article{ , abstract = {Speech is a paramount mode of human communication, and enhancing the quality and fluency of Human-Computer Interaction (HCI) greatly benefits from the significant contribution of Speech Emotion Recognition (SER). Feature representation poses a persistent challenge in SER. A single feature is difficult to adequately represent speech emotion, while directly concatenating multiple features may overlook the complementary nature and introduce interference due to redundant information. Towards these difficulties, this paper proposes a Multi-feature Learning network based on Dynamic-Static feature Fusion (ML-DSF) to obtain an effective hybrid feature representation for SER. Firstly, a Time-Frequency domain Self-Calibration Module (TFSC) is proposed to help the traditional convolutional neural networks in extracting static image features from the Log-Mel spectrograms. Then, a Lightweight Temporal Convolutional Network (L-TCNet) is used to acquire multi-scale dynamic temporal causal knowledge from the Mel Frequency Cepstrum Coefficients (MFCC). At last, both extracted features groups are fed into a connection attention module, optimized by Principal Component Analysis (PCA), facilitating emotion classification by reducing redundant information and enhancing the complementary information between features. For ensuring the independence of feature extraction, this paper adopts the training separation strategy. Evaluating the proposed model on two public datasets yielded a Weighted Accuracy (WA) of 93.33 % and an Unweighted Accuracy (UA) of 93.12 % on the RAVDESS dataset, and 94.95 % WA and 94.56 % UA on the EmoDB dataset. The obtained results outperformed the State-Of-The-Art (SOTA) findings. Meanwhile, the effectiveness of each module is validated by ablation experiments, and the generalization analysis is carried out on the cross-corpus SER tasks.}, archive = {J_NEUCOM}, author = {Peiyun Xue and Xiang Gao and Jing Bai and Zhenan Dong and Zhiyu Wang and Jiangshuai Xu}, doi = {10.1016/j.neucom.2025.129836}, journal = {Neurocomputing}, month = {6}, pages = {129836}, shortjournal = {Neurocomputing}, title = {A dynamic-static feature fusion learning network for speech emotion recognition}, volume = {633}, year = {2025}, }
• (2025). Bilinear-experts network with self-adaptive sampler for long-tailed visual recognition. NEUCOM, 633, 129832. (www)

  @article{ , abstract = {Long-tail distributed data hinders the practical application of state-of-the-art deep models in computer vision. Consequently, exclusive methodologies for handling the long-tailed problem are proposed, focusing on different hierarchies. For embedding hierarchy, existing works manually augment the diversity of tail-class features for specific datasets. However, prior knowledge about datasets is not always available for practical use, which brings unsatisfactory generalization ability in human fine-turned augmentation under such circumstances. To figure out this problem, we introduce a novel model named Bilinear-Experts Network (BENet) with Self-Adaptive Sampler (SAS). This model leverages model-driven perturbations to tail-class embeddings while preserving generalization capability on head classes through a designed bilinear experts system. The designed perturbations adaptively augment tail-class space and shift the class boundary away from the tail-class centers. Moreover, we find that SAS automatically assigns more significant perturbations to specific tail classes with relatively fewer training samples, which indicates SAS is capable of filtering tail classes with lower quality and enhancing them. Also, experiments conducted across various long-tailed benchmarks validate the comparable performance of the proposed BENet.}, archive = {J_NEUCOM}, author = {Qin Wang and Sam Kwong and Xizhao Wang}, doi = {10.1016/j.neucom.2025.129832}, journal = {Neurocomputing}, month = {6}, pages = {129832}, shortjournal = {Neurocomputing}, title = {Bilinear-experts network with self-adaptive sampler for long-tailed visual recognition}, volume = {633}, year = {2025}, }
• (2025). A review of SCADA-based condition monitoring for wind turbines via artificial neural networks. NEUCOM, 633, 129830. (www)

  @article{ , abstract = {In recent years, the supervisory control and data acquisition (SCADA) data has gained increasing research attention in the field of wind turbine condition monitoring. Artificial intelligence (AI) techniques have been widely applied to address condition monitoring challenges, and artificial neural networks (ANNs), recognized as a foundational component of modern AI, have proven to be particularly effective tools. Wind turbine condition monitoring focuses on analyzing the operational parameters of turbines to realize early fault detection, precise diagnostics, and accurate prognostics, thereby mitigating the risk of catastrophic faults, enhancing system reliability, and improving wind farm operational efficiency. Due to inherent issues in raw SCADA data, including missing values and abnormal data, preprocessing steps such as data cleaning are critical before feeding the data into ANN models. Additionally, the choice of ANN architecture typically depends on the specific requirements of condition monitoring tasks (e.g., fault detection, diagnosis, or prediction/prognosis) and the characteristics of SCADA datasets such as imbalance problem of fault samples. Hence, current research with respect to wind turbine condition monitoring generally follows two approaches: (1) utilizing classification models to identify fault types at specific time points, and (2) employing regression models to construct normal behavior models (NBMs) or track and predict continuous key performance indicators. This survey systematically reviews SCADA-based wind turbine condition monitoring methods within five years, emphasizing neural networks as key approaches, and structures the discussion around three core aspects: data preprocessing, classification models, and regression models. Moreover, the comparative strengths, capabilities, and limitations of various ANNs in each link are discussed. By providing an in-depth analysis, this paper aims to offer theoretical and practical insights to support the further development of condition monitoring technologies for wind turbines.}, archive = {J_NEUCOM}, author = {Li Sheng and Chunyu Li and Ming Gao and Xiaopeng Xi and Donghua Zhou}, doi = {10.1016/j.neucom.2025.129830}, journal = {Neurocomputing}, month = {6}, pages = {129830}, shortjournal = {Neurocomputing}, title = {A review of SCADA-based condition monitoring for wind turbines via artificial neural networks}, volume = {633}, year = {2025}, }
• (2025). Zero-shot low-dose CT denoising across variable schemes via strip-scanning diffusion models. NEUCOM, 633, 129828. (www)

  @article{ , abstract = {Artifacts and noise in low-dose CT (LDCT) may degrade image quality, potentially impacting subsequent diagnoses. In recent years, supervised image post-processing methods have been extensively studied for their effectiveness in noise reduction. However, clinical conditions often make it difficult to obtain paired normal-dose and low-dose CT images. Additionally, scanning protocols in clinical settings are diverse, necessitating different thickness or dose settings, which further complicates and increases the cost of low-dose data collection. These challenges limit the practical application and widespread adoption of supervised methods. This study introduces a novel end-to-end zero-shot strip-scanning diffusion model (SSDiff) that requires only a single model trained on normal-dose CT (NDCT) images to achieve LDCT image denoising across various scanning protocols with different slice thicknesses, doses, or devices. The sampling process employs a strip scanning strategy that combines overlapping strip information and input LDCT images to solve the maximum a posteriori problem to produce denoising results sequentially. We use only simple convolutional and attentional architectures and perform extensive experiments on three different datasets involving different doses, thicknesses, and devices; the results show that our method outperforms supervised methods in most cases, and visualization and blinded evaluations indicate that our method is very close to NDCT.}, archive = {J_NEUCOM}, author = {Bo Su and Jiabo Xu and Xiangyun Hu and Yunfei Zha and Jun Wan and Jiancheng Li}, doi = {10.1016/j.neucom.2025.129828}, journal = {Neurocomputing}, month = {6}, pages = {129828}, shortjournal = {Neurocomputing}, title = {Zero-shot low-dose CT denoising across variable schemes via strip-scanning diffusion models}, volume = {633}, year = {2025}, }
• (2025). LN-DETR: An efficient transformer architecture for lung nodule detection with multi-scale feature fusion. NEUCOM, 633, 129827. (www)

  @article{ , abstract = {Lung cancer remains one of the leading causes of mortality worldwide, and early detection is crucial for improving patient survival rates. Traditional lung nodule detection methods are inefficient and inaccurate, making them inadequate for clinical needs. Although deep learning methods have made progress in medical image analysis, existing approaches still perform poorly in detecting small, morphologically complex lung nodules, leading to missed detections and false positives. Additionally, the high computational complexity of previous models hinders real-time detection. To address these challenges, this study proposes a Transformer-based lung nodule detection model called LN-DETR. The model integrates a Partial Convolution-based Efficient Multi-scale Attention (PC-EMA) module, a Grouped and Shuffled Convolutional Cross-scale Feature Fusion (GS-CCFM) module, and introduces a Channel Transformer (CTrans) module. PC-EMA combines Efficient Multi-Scale Attention with partial convolution to enhance multi-scale feature extraction while optimizing computational efficiency. GS-CCFM uses Grouped and Shuffled Convolution (GSConv) to achieve efficient cross-scale feature fusion. The CTrans module employs a cross-channel attention mechanism to further strengthen feature fusion capabilities. Experimental results on the LUNA16 and Tianchi lung nodule datasets demonstrate that LN-DETR outperforms existing object detection models in detection accuracy, computational efficiency, and model complexity. On the LUNA16 dataset, LN-DETR achieved an F1 score of 91.5% and a mean Average Precision (mAP) of 93.1%; on the Tianchi dataset, the F1 score was 87.4% and the mAP was 86.4%, both significantly higher than baseline models. Furthermore, the reduced parameter count and computational overhead make the model more suitable for broader clinical applications.}, archive = {J_NEUCOM}, author = {Jiade Tang and Xiao Chen and Linyuan Fan and Zhenliang Zhu and Chen Huang}, doi = {10.1016/j.neucom.2025.129827}, journal = {Neurocomputing}, month = {6}, pages = {129827}, shortjournal = {Neurocomputing}, title = {LN-DETR: An efficient transformer architecture for lung nodule detection with multi-scale feature fusion}, volume = {633}, year = {2025}, }
• (2025). Observer-based adaptive neural network force tracking control for pneumatic polishing system end-effector with actuator saturation. NEUCOM, 633, 129824. (www)

  @article{ , abstract = {This paper focuses on the issue of addressing an adaptive neural network dynamic contact force tracking control for pneumatic polishing system end-effector with unmeasurable states and unknown environmental model. To tackle this problem, we utilize a neural network to approximate uncertain nonlinear functions in end-effector system model and design a neural network state observer to estimate unmeasurable states. Additionally, we combine the saturation characteristics of the pneumatic actuator and set the end-effector input voltage saturation condition. By employing the adaptive backstepping approach and the finite time convergence theory, we formulate an adaptive neural network force tracking control method. The proposed control scheme ensures the input and state stability of the end-effector of the pneumatic polishing system, and the ability to accurately track the desired contact force signal in a finite time is verified by stability analysis. Finally, the effectiveness and superiority of the proposed control scheme is verified through the tracking control experiments of static contact and dynamic contact in a variety of polishing scenarios, which provides a new idea for the application of pneumatic technology in polishing scenarios.}, archive = {J_NEUCOM}, author = {Zhiguo Yang and Jiange Kou and Zhanxin Li and Yushan Ma and Wenbo Zhao and Yixuan Wang and Yan Shi}, doi = {10.1016/j.neucom.2025.129824}, journal = {Neurocomputing}, month = {6}, pages = {129824}, shortjournal = {Neurocomputing}, title = {Observer-based adaptive neural network force tracking control for pneumatic polishing system end-effector with actuator saturation}, volume = {633}, year = {2025}, }
• (2025). Correlation-based switching mean teacher for semi-supervised medical image segmentation. NEUCOM, 633, 129818. (www)

  @article{ , abstract = {The mean teacher framework is one of the mainstream approaches in semi-supervised medical image segmentation. While training together in the traditional mean teacher framework, the teacher model and the student model share the same structure. An Exponential Moving Average (EMA) updating strategy is applied to optimize the teacher model. Although the EMA approach facilitates a smooth training process, it causes the model coupling and error accumulation problems. These issues constrain the model from precisely delineating the regions of pathological structures, especially for the low-contrast regions in medical images. In this paper, we propose a new semi-supervised segmentation model, namely Correlation-based Switching Mean Teacher (CS-MT), which comprises two teacher models and one student model to alleviate these problems. Particularly, two teacher models adopt a switching training strategy at every epoch to avoid the convergence and similarity between the teacher models and the student model. In addition, we introduce a feature correlation module in each model to leverage the similarity information in the feature maps to improve the model’s predictions. Furthermore, the stochastic process of CutMix operation destroys the structures of organs in medical images, generating adverse mixed results. We propose an adaptive CutMix manner to mitigate the negative effects of these mixed results in model training. Extensive experiments validate that CS-MT outperforms the state-of-the-art semi-supervised methods on the LA, Pancreas-NIH, ACDC and BraTS 2019 datasets.}, archive = {J_NEUCOM}, author = {Guiyuhan Deng and Hao Sun and Wei Xie}, doi = {10.1016/j.neucom.2025.129818}, journal = {Neurocomputing}, month = {6}, pages = {129818}, shortjournal = {Neurocomputing}, title = {Correlation-based switching mean teacher for semi-supervised medical image segmentation}, volume = {633}, year = {2025}, }
• (2025). DiffuseVAE++: Mitigating training-sampling mismatch based on additional noise for higher fidelity image generation. NEUCOM, 633, 129814. (www)

  @article{ , abstract = {Denoising Diffusion Probabilistic Models (DDPMs) have demonstrated remarkable results in image generation. However, there exist a mismatch between the training and sampling process in current diffusion models, in addition, the U-Net denoising network based on simple residual blocks cannot predict noise information accurately, which affects the generated quality. To address these limitations, we present a novel image generation method that achieves higher fidelity. First, by additionally adding the standard Gaussian noise in the diffusion forward process, which does not disrupt the forward process, our method alleviates the mismatch. Subsequently, an important efficient denoising network based on U-Net is presented, where our proposed Simple Squeeze-Excitation and Simple GLU, combined with Depthwise Separable Convolution, enhance the ability of the model to predict real noise using the Simplified Nonlinear No Activation (SNNA) block. Furthermore, considering the structural characteristics of the baseline model, we introduce an additional cross-attention mechanism to enable DDPM to focus on VAE stage characteristics. Allowing the model to more accurately capture and learn the noise information. Finally, it is shown after extensive experiments the proposed DiffuseVAE++ obtains significant gains in FID scores, improving from 3.84 to 2.41 on CIFAR-10 and from 3.94 to 2.30 on CelebA-64. In particular, the IS scores on CIFAR-10 reaches 10.10, which is comparable to the current state-of-the-art methods competitively ( e.g., U-ViT , StyleGAN2 ).}, archive = {J_NEUCOM}, author = {Xiaobao Yang and Wei Luo and Hailong Ning and Guorui Zhang and Wei Sun and Sugang Ma}, doi = {10.1016/j.neucom.2025.129814}, journal = {Neurocomputing}, month = {6}, pages = {129814}, shortjournal = {Neurocomputing}, title = {DiffuseVAE++: Mitigating training-sampling mismatch based on additional noise for higher fidelity image generation}, volume = {633}, year = {2025}, }
• (2025). CMGN: Text GNN and RWKV MLP-mixer combined with cross-feature fusion for fake news detection. NEUCOM, 633, 129811. (www)

  @article{ , abstract = {With the rapid development of social media, the influence and harm of fake news have gradually increased, making accurate detection of fake news particularly important. Current fake news detection methods primarily rely on the main text of the news, neglecting the interrelationships between additional texts. We propose a cross-feature fusion network with additional text graph construction to address this issue and improve fake news detection. Specifically, we utilize a text graph neural network (GNN) to model the graph relationships of additional texts to enhance the model’s perception capabilities. Additionally, we employ the RWKV MLP-mixer to process the news text and design a cross-feature fusion mechanism to achieve mutual fusion of different features, thereby improving fake news detection. Experiments on the LIAR, FA-KES, IFND, and CHEF datasets demonstrate that our proposed model outperforms existing methods in fake news detection.}, archive = {J_NEUCOM}, author = {ShaoDong Cui and Kaibo Duan and Wen Ma and Hiroyuki Shinnou}, doi = {10.1016/j.neucom.2025.129811}, journal = {Neurocomputing}, month = {6}, pages = {129811}, shortjournal = {Neurocomputing}, title = {CMGN: Text GNN and RWKV MLP-mixer combined with cross-feature fusion for fake news detection}, volume = {633}, year = {2025}, }
• (2025). H-SGANet: Hybrid sparse graph attention network for deformable medical image registration. NEUCOM, 633, 129810. (www)

  @article{ , abstract = {The integration of Convolutional Neural Networks (ConvNets) and Transformers has become a strong candidate for image registration, combining the strengths of both models and utilizing a large parameter space. However, this hybrid model, which treats brain MRI volumes as grid or sequence structures, struggles to accurately represent anatomical connectivity, diverse brain regions, and critical connections within the brain’s architecture. There are also concerns about the computational expense and GPU memory usage of this model. To address these issues, we propose a lightweight hybrid sparse graph attention network (H-SGANet). The network includes Sparse Graph Attention (SGA), a core mechanism based on Vision Graph Neural Networks (ViG) with predefined anatomical connections. The SGA module expands the model’s receptive field and integrates seamlessly into the network. To further enhance the hybrid network, Separable Self-Attention (SSA) is used as an advanced token mixer, combined with depth-wise convolution to form SSAFormer. This strategic integration is designed to more effectively extract long-range dependencies. As a hybrid ConvNet-ViG-Transformer model, H-SGANet offers three key benefits for volumetric medical image registration. It optimizes fixed and moving images simultaneously through a hybrid feature fusion layer and an end-to-end learning framework. Compared to VoxelMorph, a model with a similar parameter count, H-SGANet demonstrates significant performance enhancements of 3.5% and 1.5% in Dice score on the OASIS dataset and LPBA40 dataset, respectively. The code is publicly available at https://github.com/2250432015/H-SGANet/ .}, archive = {J_NEUCOM}, author = {Yufeng Zhou and Wenming Cao}, doi = {10.1016/j.neucom.2025.129810}, journal = {Neurocomputing}, month = {6}, pages = {129810}, shortjournal = {Neurocomputing}, title = {H-SGANet: Hybrid sparse graph attention network for deformable medical image registration}, volume = {633}, year = {2025}, }
• (2025). Decoupled contrastive learning for multilingual multimodal medical pre-trained model. NEUCOM, 633, 129809. (www)

  @article{ , abstract = {Multilingual multimodal pre-training aims to facilitate the integration of conceptual representations across diverse languages and modalities within a shared, high-dimensional semantic space. This endeavor in healthcare faces challenges related to language diversity, suboptimal multimodal interactions, and an absence of coherent multilingual multimodal representations. In response to these challenges, we introduce a novel multilingual multimodal medical pre-training model. Initially, we employ a strategic augmentation of the medical corpus by expanding the MIMIC-CXR report dataset to 20 distinct languages using machine translation techniques. Subsequently, we develop a targeted label disambiguation technique to address the labeling noise within decoupled contrastive learning. In particular, it categorizes and refines uncertain phrases within the clinical reports based on disease type, promoting finer-grained semantic similarity and improving inter-modality interactions. Building on these proposals, we present a refined multilingual multimodal medical pre-trained model, significantly enhancing the understanding of medical multimodal data and adapting the model to multilingual medical contexts. Experiments reveal that our model outperforms other baselines in medical image classification and multilingual medical image–text retrieval by up to 13.78% and 12.6%, respectively.}, archive = {J_NEUCOM}, author = {Qiyuan Li and Chen Qiu and Haijiang Liu and Jinguang Gu and Dan Luo}, doi = {10.1016/j.neucom.2025.129809}, journal = {Neurocomputing}, month = {6}, pages = {129809}, shortjournal = {Neurocomputing}, title = {Decoupled contrastive learning for multilingual multimodal medical pre-trained model}, volume = {633}, year = {2025}, }
• (2025). EPAM-net: An efficient pose-driven attention-guided multimodal network for video action recognition. NEUCOM, 633, 129781. (www)

  @article{ , abstract = {Existing multimodal-based human action recognition approaches are computationally intensive, limiting their deployment in real-time applications. In this work, we present a novel and efficient pose-driven attention-guided multimodal network (EPAM-Net) for action recognition in videos. Specifically, we propose eXpand temporal Shift (X-ShiftNet) convolutional architectures for RGB and pose streams to capture spatio-temporal features from RGB videos and their skeleton sequences. The X-ShiftNet tackles the high computational cost of the 3D CNNs by integrating the Temporal Shift Module (TSM) into an efficient 2D CNN, enabling efficient spatiotemporal learning. Then skeleton features are utilized to guide the visual network stream, focusing on keyframes and their salient spatial regions using the proposed spatial–temporal attention block. Finally, the predictions of the two streams are fused for final classification. The experimental results show that our method, with a significant reduction in floating-point operations (FLOPs), outperforms and competes with the state-of-the-art methods on NTU RGB-D 60, NTU RGB-D 120, PKU-MMD, and Toyota SmartHome datasets. The proposed EPAM-Net provides up to a 72.8x reduction in FLOPs and up to a 48.6x reduction in the number of network parameters. The code will be available at https://github.com/ahmed-nady/Multimodal-Action-Recognition .}, archive = {J_NEUCOM}, author = {Ahmed Abdelkawy and Asem Ali and Aly Farag}, doi = {10.1016/j.neucom.2025.129781}, journal = {Neurocomputing}, month = {6}, pages = {129781}, shortjournal = {Neurocomputing}, title = {EPAM-net: An efficient pose-driven attention-guided multimodal network for video action recognition}, volume = {633}, year = {2025}, }
• (2025). Cross-modal feature symbiosis for personalized meta-path generation in heterogeneous networks. NEUCOM, 633, 129780. (www)

  @article{ , abstract = {In heterogeneous graph neural networks (HGNNs), the capture of intricate relationships among various types of entities is essential to achieve advanced machine learning applications. Heterogeneous Information Networks (HINs), composed of interconnected multi-type nodes and edges, face significant challenges in managing semantic diversity and inherent heterogeneity. Traditional methods, which rely on manually designed meta-paths, struggle to adapt dynamically to personalized needs and often neglect the integration of structural and attribute features. To address these limitations, this paper introduces the Cross-Modal Symbiotic Meta-Path Generator (CSMPG) framework. CSMPG integrates two key modules: a Cross-Modal State Generation Module that encodes node structure and attribute information into task-aware state vectors and a Personalized Meta-Path Generation Module that dynamically generates and refines meta-paths using reinforcement learning. By leveraging downstream task feedback, CSMPG optimizes path selection to maximize performance. The framework effectively balances cross-modal feature integration and semantic diversity, uncovering impactful meta-paths that are often overlooked by traditional approaches. Experimental results demonstrate that CSMPG consistently enhances recommendation quality and significantly outperforms structure-only and predefined-path-based models.}, archive = {J_NEUCOM}, author = {Xiaotong Wu and Liqing Qiu and Weidong Zhao}, doi = {10.1016/j.neucom.2025.129780}, journal = {Neurocomputing}, month = {6}, pages = {129780}, shortjournal = {Neurocomputing}, title = {Cross-modal feature symbiosis for personalized meta-path generation in heterogeneous networks}, volume = {633}, year = {2025}, }
• (2025). Zero-shot text-to-parameter realtime translation for game character auto-creation and identity consistency editing. NEUCOM, 633, 129779. (www)

  @article{ , abstract = {In modern video games, character customization often requires adjusting numerous complex parameters, which can be inefficient. Consequently, there is growing interest in using text-based methods to generate 3D characters. The text-based method enhances players’ immersion in the game while improving development efficiency by simplifying the creation of diverse characters. This paper introduces a Text-to-Parameter Editable model (T2PE) that enables players to rapidly create high-quality game characters from textual descriptions. T2PE leverages supervision of CLIP to optimize both continuous and discrete facial parameters simultaneously. This differs from existing Text-to-Parameter (T2P) models, which rely on evolutionary search algorithm that interacts with the game engine to optimize discrete parameters. We also propose a game character editing method that allows modifications using new textual descriptions while preserving other features. Our experimental results show that the proposed model generates higher-quality 3D game characters than T2P models while significantly reducing latency. We observe that T2PE achieves 26x faster inference times and 3.3% increase in CLIP score on lower TFLOPS devices. To the best of our knowledge, T2PE is the first algorithm to support both text-based character generation and text-based character editing. Building upon this, we develop Fast-T2PE, an end-to-end translator trained using text–parameter pairs obtained from T2PE, which enables even faster generation of game characters. Experimental results show that Fast-T2PE further reduces response latency by 62.5% compared to T2PE.}, archive = {J_NEUCOM}, author = {Tao Wu and Xu Lu and Jia Yu and Weijun Cao and Wei Chen and Haidi Fan and Zhiqiang Zhang and Gen Dong and Siwei Zhou}, doi = {10.1016/j.neucom.2025.129779}, journal = {Neurocomputing}, month = {6}, pages = {129779}, shortjournal = {Neurocomputing}, title = {Zero-shot text-to-parameter realtime translation for game character auto-creation and identity consistency editing}, volume = {633}, year = {2025}, }
• (2025). A variable-gain fixed-time convergent neurodynamic network for time-variant quadratic programming under unknown noises. NEUCOM, 633, 129778. (www)

  @article{ , abstract = {This article proposes a variable-gain fixed-time convergent and noise-tolerant error-dynamics based neurodynamic network (VGFxTNT-EDNN) to solve time-varying quadratic programming problems, while being robust to unknown noises. Unlike existing finite-time convergent EDNNs, the newly designed VGFxTNT-EDNN guarantees fixed-time convergence by dynamically adjusting its variable parameters. Moreover, the VGFxTNT-EDNN effectively handles unknown noise, addressing a limitation of existing fixed-time or predefined-time convergent models, which typically assume that the noise is known. Theoretical analysis utilizing Lyapunov theory proves that the VGFxTNT-EDNN possesses fixed-time convergence and robustness properties. Numerical validations demonstrate superior noise tolerance and fixed-time convergence of the VGFxTNT-EDNN, as compared with the existing models. Finally, a path-tracking experiment is conducted by utilizing a Franka Emika Panda robot to verify the practicality of the VGFxTNT-EDNN.}, archive = {J_NEUCOM}, author = {Biao Song and Tinghe Hong and Weibing Li and Gang Chen and Yongping Pan and Kai Huang}, doi = {10.1016/j.neucom.2025.129778}, journal = {Neurocomputing}, month = {6}, pages = {129778}, shortjournal = {Neurocomputing}, title = {A variable-gain fixed-time convergent neurodynamic network for time-variant quadratic programming under unknown noises}, volume = {633}, year = {2025}, }
• (2025). LAR-pose: Lightweight human pose estimation with adaptive regression loss. NEUCOM, 633, 129777. (www)

  @article{ , abstract = {In this paper, LAR-Pose, a lightweight, high-resolution network for human pose estimation driven by adaptive regression loss is proposed and experimentally demonstrated based on MS COCO and MPII. The architecture of the LAR-Pose comprises two main components. One is a lightweight high-resolution backbone network, which utilizes a parallel high-resolution architecture with conditional channel weighting block to reduce the model size and computational complexity. The other is a dynamic residual refinement network, which calculates residuals from pseudo-heatmaps and scaling factors, improving training concentration for consistent distribution estimation, rather than predicting coordinates or heatmaps directly. Specific coordinates are derived through integral heatmap regression, effectively minimizing quantization errors. Our adaptive regression loss, which uses a flow model to fit the distribution of residuals in real-time, provides more sensitive parameter feedback than conventional heatmap loss, ensuring differentiability and continuity during backpropagation while enhancing performance. With a relatively small parameter scale, LAR-Pose achieves an AP of 73.5 on MS COCO and a PCKh of 90.9 on MPII, while the results outperform most advanced small networks and approach the performance of large networks.}, archive = {J_NEUCOM}, author = {Xudong Lou and Xin Lin and Henan Zeng and Xiangxian Zhu}, doi = {10.1016/j.neucom.2025.129777}, journal = {Neurocomputing}, month = {6}, pages = {129777}, shortjournal = {Neurocomputing}, title = {LAR-pose: Lightweight human pose estimation with adaptive regression loss}, volume = {633}, year = {2025}, }
• (2025). Passivity for undirected and directed fractional-order complex networks with adaptive output coupling. NEUCOM, 633, 129774. (www)

  @article{ , abstract = {In this paper, the passivity for adaptive output coupled fractional-order complex networks (CNs) under undirected and directed topologies is studied by selecting appropriate coupling weight adjustment strategies, based on which the output synchronization for the presented networks is discussed by leveraging the properties of Mittag-Leffler functions and the Laplace transform technique. Utilizing the developed distributed adaptive schemes, several passivity criteria for the fractional-order CNs with output coupling are derived. Moreover, several output synchronization conditions for the output coupled fractional-order CNs are given by employing the acquired passivity results. Ultimately, simulations from numerical examples are utilized to judge the effectiveness for the adaptive laws and the presented criteria.}, archive = {J_NEUCOM}, author = {Jin-Liang Wang and Chen-Guang Liu and Shun-Yan Ren and Tingwen Huang}, doi = {10.1016/j.neucom.2025.129774}, journal = {Neurocomputing}, month = {6}, pages = {129774}, shortjournal = {Neurocomputing}, title = {Passivity for undirected and directed fractional-order complex networks with adaptive output coupling}, volume = {633}, year = {2025}, }
• (2025). Automatic diagnosis of early pregnancy fetal nasal bone development based on complex mid-sagittal section ultrasound imaging. NEUCOM, 633, 129773. (www)

  @article{ , abstract = {Early prenatal screening of fetal nasal bone (FNB) development is crucial for detecting chromosomal abnormalities. Existing deep learning approaches primarily focus on detection rather than diagnosis of FNB. This paper introduces an early prenatal FNB development automated diagnostic system (FNB-ADS), which employs a cascaded hierarchical filtering method to reduce noise interference in mid-sagittal plane ultrasound images. Specifically, the system employs YOLOv8 for precise FNB localization, segments the nasal bone, tip, and prenasal skin using a specially designed lightweight segmentation network, and diagnoses developmental abnormalities using Resnet34 classification methods. Furthermore, this paper has collected and publicly released the FNB-UDV dataset, which includes a detection subset and a video subset. The detection subset contains 1,007 two-dimensional ultrasound images, while the video subset comprises 12 ultrasound videos. Upon a comprehensive evaluation, the diagnostic accuracy of FNB-ADS reaches 92.37% with a processing time of 0.14 s per image, and the video diagnostic accuracy is 98.69% with a per-frame inference speed of 0.37 s in the FNB-UDV dataset. Representing the first deep-learning approach tailored specifically for early pregnancy FNB ultrasound video diagnosis, FNB-ADS significantly enhances the standardization of diagnostic procedures and reduces the dependence on subjective clinical assessments. The dataset and code are available at https://github.com/SIGMACX/FNB-AD/tree/FNB-ADS .}, archive = {J_NEUCOM}, author = {Xi Chen and Xiaoyu Xu and Lyuyang Tong and Huangxuan Zhao and Bo Du}, doi = {10.1016/j.neucom.2025.129773}, journal = {Neurocomputing}, month = {6}, pages = {129773}, shortjournal = {Neurocomputing}, title = {Automatic diagnosis of early pregnancy fetal nasal bone development based on complex mid-sagittal section ultrasound imaging}, volume = {633}, year = {2025}, }
• (2025). MCU-net: A multi-prior collaborative deep unfolding network with gates-controlled spatial attention for accelerated MRI reconstruction. NEUCOM, 633, 129771. (www)

  @article{ , abstract = {Deep unfolding networks (DUNs) have shown promise in accelerating magnetic resonance imaging (MRI); however, they often face challenges such as high computational costs, slow convergence, and difficulty in fully exploiting complementarity in multiple priors. In the present study, we proposed a novel DUN, termed Multi-prior collaborative deep unfolding network (MCU-Net), to address these limitations. Our method features a parallel structure comprising different optimization-inspired subnetworks based on low-rank and sparsity, respectively. We designed a gates-controlled spatial attention module (GSAM), evaluating the relative confidence (RC) and overall confidence (OC) maps for intermediate reconstructions produced by different subnetworks. RC allocates greater weights to the image regions where each subnetwork excels, enabling precise element-wise collaboration. We designed correction modules to enhance the effectiveness in regions where both subnetworks exhibit limited performance, as indicated by low OC values, thereby obviating the need for additional branches. The gate units within GSAMs are designed to filter necessary information across multiple iterations, thus improving the accuracy of the learned confidence maps and enhancing robustness against accumulated errors. Through evaluating four datasets, MCU-Net outperforms state-of-the-art methods, achieving average PSNR improvements of 0.22, 0.24, 0.41 and 0.12 dB, respectively. Furthermore, applying the proposed strategy to various DUNs consistently results in accelerated convergence and improved reconstruction performance without extra FLOPs. Experimental results demonstrated that the proposed collaborative strategy can improve both reconstruction quality and computational efficiency compared to existing MRI reconstruction methods.}, archive = {J_NEUCOM}, author = {Xiaoyu Qiao and Weisheng Li and Guofen Wang and Yuping Huang}, doi = {10.1016/j.neucom.2025.129771}, journal = {Neurocomputing}, month = {6}, pages = {129771}, shortjournal = {Neurocomputing}, title = {MCU-net: A multi-prior collaborative deep unfolding network with gates-controlled spatial attention for accelerated MRI reconstruction}, volume = {633}, year = {2025}, }
• (2025). MEAI-net: Multiview embedding and attention interaction for multivariate time series forecasting. NEUCOM, 633, 129769. (www)

  @article{ , abstract = {Multivariate time series (MTS) forecasting plays a critical role in diverse societal applications, including stock market analysis and climate change research. While many existing deep learning models have been proven to be effective in MTS forecasting through complex neural network structures and self-attention mechanisms, several challenges remain: (1) insufficient modeling of complex temporal dependencies, (2) limited ability to handle information redundancy and noise, and (3) inadequate capture of periodic characteristics of time series. To address these problems, we propose a Multiview time-dependent feature Embedding and downsampled subsequences Attention Interaction Network (MEAI-Net) for MTS forecasting. First, MEAI-Net adopts a multiview time-dependent feature embedding mechanism to extract various temporal dependency features from the sequences. Second, it reduces redundancy in the temporal sequence features through downsampling. Third, a subsequences cross-attention module is introduced to enhance information exchange between subsequences. Furthermore, we propose the period consistency loss designed to more effectively capture periodic patterns in time series data. Comprehensive experiments conducted on 12 widely used time series datasets demonstrate that MEAI-Net displays promising performance, providing a competitive alternative to current state-of-the-art approaches in MTS forecasting.}, archive = {J_NEUCOM}, author = {Chunru Dong and Wenqing Xu and Feng Zhang and Qiang Hua and Yong Zhang}, doi = {10.1016/j.neucom.2025.129769}, journal = {Neurocomputing}, month = {6}, pages = {129769}, shortjournal = {Neurocomputing}, title = {MEAI-net: Multiview embedding and attention interaction for multivariate time series forecasting}, volume = {633}, year = {2025}, }
• (2025). Multi-view clustering via view-specific consensus kernelized graph learning. NEUCOM, 633, 129766. (www)

  @article{ , abstract = {Multi-view clustering has received extensive and in-depth research attention in recent years owing to its ability to reflect the nature of the real world from multiple perspectives. Kernel-based methods and subspace learning-based methods are two important categories of multi-view clustering. Compared with subspace-based algorithms, kernel-based algorithms can better address nonlinear relationships in feature spaces. However, the current kernel-based algorithms focus mainly on the diversity of different kernels, and obtaining the optimal kernel via linear combinations of multiple kernels, ignoring the cross-view information and space information in the original feature spaces. To address this issue, our paper proposes a novel algorithm named MC-VCKGL. Specifically, we first obtain view-specific consensus kernelized graphs of each view through kernel-based self-representation learning and by using the kernel trick. Moreover, Laplacian constraints are applied to maintain smoothness in the raw feature space of each view. We stack these kernelized graphs together to obtain a tensor, and then rotate this tensor and apply tensor nuclear norm constraints. As a result, the cross-view complementary information can be explored. We apply our algorithm to seven open datasets, including both text and image datasets. Experiments show that our method outperforms most state-of-the-art multi-view clustering algorithms.}, archive = {J_NEUCOM}, author = {Bing Hu and Tong Wu and Lixin Han and Shu Li and Yi Xu and Gui-fu Lu}, doi = {10.1016/j.neucom.2025.129766}, journal = {Neurocomputing}, month = {6}, pages = {129766}, shortjournal = {Neurocomputing}, title = {Multi-view clustering via view-specific consensus kernelized graph learning}, volume = {633}, year = {2025}, }
• (2025). Brain-like contour detector following retinex theory and gestalt perception grouping principles. NEUCOM, 633, 129765. (www)

  @article{ , abstract = {Numerous contour detection algorithms draw inspiration from biological vision systems. These algorithms imitate the way simple cells extract edges using Gabor filters. They also suppress edges generated by image textures simulating the non-classical receptive fields (NCRFs), thereby popping up the object contours within the image edges. However, these algorithms are not flawless and may yield imperfect results due to noise pollution, unsatisfactory lighting, limitations in image processing algorithms, and likewise. Weak strengths and pixel loss in contour segments are two common issues. In this paper, we provide two strategies to address these challenges. First, we separate the illumination component from the image following Retinex theory, extract the illumination contour using bio-inspired filters, and boost contour strengths by superimposing the illumination contour. Second, we complete object contours by filling small gaps in contours, using a proposed linking likelihood function that is a joint probability of element distance and orientation difference, following Gestalt perceptual grouping principles. Although not performance-oriented, the experimental results show that our endeavors improve the performance of bio-inspired contour detectors. More importantly, we demonstrate the significance of visual computation theories such as the Retinex theory and the Gestalt perception grouping principle for contour detection.}, archive = {J_NEUCOM}, author = {Rongtai Cai and Helin Que}, doi = {10.1016/j.neucom.2025.129765}, journal = {Neurocomputing}, month = {6}, pages = {129765}, shortjournal = {Neurocomputing}, title = {Brain-like contour detector following retinex theory and gestalt perception grouping principles}, volume = {633}, year = {2025}, }
• (2025). Attribute graph anomaly detection utilizing memory networks enhanced by multi-embedding comparison. NEUCOM, 633, 129762. (www)

  @article{ , abstract = {In complex attribute networks, accurately pinpointing anomalous nodes is vital for grasping network behavior and safeguarding network security. Traditional anomaly detection methods often struggle to fully harness the intricate relationships that underpin attributes and structures, thus curbing their practical effectiveness. To transcend this limitation, we introduce a novel graph anomaly detection model that harmoniously integrates node attributes and structural information. Our model employs multi-embedding contrast modules, coupled with memory network enhancements, to pinpoint anomalous nodes. Precisely, we crafted a multi-embedding contrast module to encode the attributes and structures inherent within nodes, generating a multitude of embedding representations. By scrutinizing the discrepancies between these representations, our model adeptly identifies nodes that deviate from attribute and structural consistency, indicating anomalies. Furthermore, we incorporate a memory network to reconstruct node attributes, thereby enhancing the attribute decoding process while preserving the straightforwardness of structural decoding. To validate our method, we conducted extensive experiments on five authoritative public graph datasets, comparing various graph anomaly detection methods using rigorous metrics such as AUC, precision, and recall. The experimental results unequivocally demonstrate that our proposed method surpasses current state-of-the-art techniques in detecting anomalous nodes within graphs, solidly validating its efficacy.}, archive = {J_NEUCOM}, author = {Lianming Zhang and Baolin Wu and Pingping Dong}, doi = {10.1016/j.neucom.2025.129762}, journal = {Neurocomputing}, month = {6}, pages = {129762}, shortjournal = {Neurocomputing}, title = {Attribute graph anomaly detection utilizing memory networks enhanced by multi-embedding comparison}, volume = {633}, year = {2025}, }
• (2025). Memory-based walk-enhanced dynamic graph neural network for temporal graph representation learning. NEUCOM, 633, 129759. (www)

  @article{ , abstract = {Depending on the ability of obtaining low-dimensional representations of nodes that preserve valuable structural information, graph representation learning has a wide range of applications in graph analysis and inference. However, real-world complex systems are naturally heterogeneous and time-varying, which makes it difficult to learn high-quality node representations. We propose a M emory-based W alk-enhanced D ynamic G raph neural N etwork (denoted as MWDGN) to fully exploit the dependencies and structural features in temporal graph. To capture long-term dependencies, we use a memory module to store and evolve dynamic node representations. MWDGN captures network structural information by constructing time-constrained walk sequences for each interaction node. The walk sequence features are creatively integrated into the update process of memory module, so as to capture the useful information of neighborhood structure features for the interaction node while preserving the long-term dependency of the temporal graph. In addition, we focus on the enlightenment of non-negligible temporal information for sensing key historical interaction nodes of the target node, and design a new aggregation method of historical interaction nodes information. It exploits the temporal attenuation effect of event impact to model short-term dependencies. We further exploit causal convolutional network to mine the potential associations of historical interaction node features of the target node. Comparison experiments on six datasets with mainstream baseline models demonstrate that MWDGN is capable of jointly extracting the heterogeneity and evolutionary patterns of nodes in the graph, improving the node representation quality, and enhancing the performance of the temporal link prediction and dynamic node classification tasks. The effectiveness of the proposed model is further proved by time complexity analysis, ablation study and parameter sensitivity analysis.}, archive = {J_NEUCOM}, author = {Zhigang Jin and Renjun Su and Hao Zhang and Xiaofang Zhao}, doi = {10.1016/j.neucom.2025.129759}, journal = {Neurocomputing}, month = {6}, pages = {129759}, shortjournal = {Neurocomputing}, title = {Memory-based walk-enhanced dynamic graph neural network for temporal graph representation learning}, volume = {633}, year = {2025}, }
• (2025). Inverse reinforcement learning by expert imitation for the stochastic linear–quadratic optimal control problem. NEUCOM, 633, 129758. (www)

  @article{ , abstract = {This article studies inverse reinforcement learning (IRL) for the linear–quadratic stochastic optimal control problem, where two agents are considered. A learner agent lacks knowledge of the expert agent’s cost function, but it reconstructs an underlying cost function by observing the expert agent’s states and controls, thereby imitating the expert agent’s optimal feedback control. We initially present a model-based IRL method, which consists of a policy correction and a policy update from the policy iteration in reinforcement learning, as well as a cost function weight reconstruction informed by the inverse optimal control. Afterward, under this scheme, we propose a model-free off-policy IRL method, which requires no system identification, only collecting behavior data from the learner agent and expert agent once during the iteration process. Moreover, the proofs of the method’s convergence, stability, and non-unique solutions are given. Finally, a numerical example and an inverse mean–variance portfolio optimization example are provided to validate the effectiveness of the presented method.}, archive = {J_NEUCOM}, author = {Zhongshi Sun and Guangyan Jia}, doi = {10.1016/j.neucom.2025.129758}, journal = {Neurocomputing}, month = {6}, pages = {129758}, shortjournal = {Neurocomputing}, title = {Inverse reinforcement learning by expert imitation for the stochastic linear–quadratic optimal control problem}, volume = {633}, year = {2025}, }
• (2025). Decoding before aligning: Scale-adaptive early-decoding transformer for visual grounding. NEUCOM, 633, 129756. (www)

  @article{ , abstract = {Visual grounding, the task of precisely localizing objects within an image as described by natural language expressions, has recently seen significant advancements through the integration of transformer decoders and early-fusion mechanisms. These approaches aim to enhance the decoding of crucial grounding-related semantics and the generation of discriminative representations. However, they display significant limitations, including reduced generalizability due to the need for custom plug-ins, a gap in effectively bridging decoded visual and linguistic features, and difficulties in processing single-scale features within complex, hierarchical scenes. To address these issues, this study introduces the Scale-Adaptive Early-Decoding Transformer (SA-EDTR), a novel approach that combines decoding and early-fusion mechanisms within a single transformer layer. By positioning the decoding stage before the alignment stage, the decoder adaptively generates richer, context-aware representations that guide the aligning module. This design allows SA-EDTR to effectively overcome previous limitations in feature alignment. Additionally, we integrate a Scale-Adaptive Fusion network (SAF) to fuse multi-scale and heterogeneous representations from parallel Early-Decoding Transformer layers, enhancing the capture of spatial, coarse, and fine-grained semantic details. Extensive experiments demonstrate that SA-EDTR outperforms current state-of-the-art methods in both REC and RES tasks across various mainstream visual grounding datasets, including RefCOCO, RefCOCO+, RefCOCOg-umd, and RefCOCOg-google. The results highlight the architecture’s remarkable capability in accurately localizing objects and interpreting intricate scenes.}, archive = {J_NEUCOM}, author = {Liuwu Li and Yi Cai and Jiexin Wang and Cantao Wu and Qingbao Huang and Qing Li}, doi = {10.1016/j.neucom.2025.129756}, journal = {Neurocomputing}, month = {6}, pages = {129756}, shortjournal = {Neurocomputing}, title = {Decoding before aligning: Scale-adaptive early-decoding transformer for visual grounding}, volume = {633}, year = {2025}, }
• (2025). Second-order consensus of matrix-weighted switched multiagent systems. NEUCOM, 633, 129755. (www)

  @article{ , abstract = {Due to the switching characteristics of many practical multi-agent systems, such as automatic speed control systems and hybrid quadcopters, for multidimensional individuals, combined with practical complexity, matrix-weighted switching dynamics are needed to model them. This paper considers the consensus issues of second-order switched multi-agent systems on matrix-weighted undirected and directed networks. A new matrix-weighted control algorithm suitable for both CT and DT subsystems is proposed. Under the proposed algorithms, based on variable transformation, matrix theory and stability theory, the consensus criteria are established for undirected and directed switched multi-agent networks that rely on the eigenvalues of the network and coupling gains, respectively. This also indicates that the matrix-weights and coupling gains have a significant impact on switched matrix-weighted consensus. Finally, through simulations, the validity of the obtained results of this essay are verified.}, archive = {J_NEUCOM}, author = {Suoxia Miao and Housheng Su}, doi = {10.1016/j.neucom.2025.129755}, journal = {Neurocomputing}, month = {6}, pages = {129755}, shortjournal = {Neurocomputing}, title = {Second-order consensus of matrix-weighted switched multiagent systems}, volume = {633}, year = {2025}, }
• (2025). Self-weighted subspace clustering with adaptive neighbors. NEUCOM, 633, 129754. (www)

  @article{ , abstract = {Subspace clustering has attracted increasing attention in recent years owing to its ability to process high-dimensional data effectively. However, existing subspace clustering methods often assume that different features are equally important, and on this basis, a similarity matrix is constructed to generate the clustering structure. However, this practice may significantly affect the clustering performance in cases where the importance of different features significantly differs or where many noisy features exist in the original data. To address these challenges, we propose a novel self-weighted subspace clustering method with adaptive neighbors (SWSCAN). A feature weighting scheme is introduced to assign appropriate weights to different features. Then, we use the self-expressive property and adaptive neighbors approach to capture both the global and local structures within the weighted data space. Moreover, we employ the alternating direction method of multipliers (ADMM) to effectively solve the optimization problem of SWSCAN. Empirical results on both synthetic and practical datasets validate that our proposed method outperforms other comparative clustering techniques and can learn appropriate weights for features.}, archive = {J_NEUCOM}, author = {Zhengyan Liu and Huiwen Wang and Lihong Wang and Qing Zhao}, doi = {10.1016/j.neucom.2025.129754}, journal = {Neurocomputing}, month = {6}, pages = {129754}, shortjournal = {Neurocomputing}, title = {Self-weighted subspace clustering with adaptive neighbors}, volume = {633}, year = {2025}, }
• (2025). Routeformer: Transformer utilizing routing mechanism for traffic flow forecasting. NEUCOM, 633, 129753. (www)

  @article{ , abstract = {Traffic flow prediction is vital for the development of intelligent transportation systems. The challenge lies in accurately capturing the complex and dynamic spatiotemporal dependencies influenced by real road network fluctuations. These dependencies can be simplified into three categories: (i) spatial dependencies among sensors at the same timestamp, (ii) temporal dependencies of the same sensor at different timestamps, and (iii) cross dimensional dependencies between different sensors at different timestamps. The third type of cross dimensional dependency requires considering the relationships between different sensors across multiple time points, which is not only complex but also difficult to capture accurately. Existing methods often describe it indirectly by merging spatiotemporal dependencies, but this approach is frequently insufficiently accurate. We aim to characterize this relationship more precisely by capturing the sequential dependencies among sensors, referred to as inter-series dependencies. Capturing inter-series dependencies does not require directly modeling the relationships between different sensors across multiple time points; rather, it focuses on the dependencies between the temporal patterns of different sensors. Our designed Temporal Routing Transformer captures temporal dependencies along the temporal axis while implicitly modeling the inter-series dependencies between sensors. At the same time, we capture spatial dependencies through the Spatial Routing Transformer and multi-scale temporal dependencies by using the Context-Aware Transformer. A series of evaluations were conducted on seven real world datasets, and Routeformer achieved state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Jun Qi and Hong Fan}, doi = {10.1016/j.neucom.2025.129753}, journal = {Neurocomputing}, month = {6}, pages = {129753}, shortjournal = {Neurocomputing}, title = {Routeformer: Transformer utilizing routing mechanism for traffic flow forecasting}, volume = {633}, year = {2025}, }
• (2025). Few-shot medical relation extraction via prompt tuning enhanced pre-trained language model. NEUCOM, 633, 129752. (www)

  @article{ , abstract = {Medical relation extraction is crucial for developing structured information to support intelligent healthcare systems. However, acquiring large volumes of labeled medical data is challenging due to the specialized nature of medical knowledge and privacy constraints. To address this, we propose a prompt-enhanced few-shot relation extraction (FSRE) model that leverages few-shot and prompt learning techniques to improve performance with minimal data. Our approach introduces a hard prompt concatenated to the original input, enabling contextually enriched learning. We calculate prototype representations by averaging the intermediate states of each relation class in the support set, and classify relations by finding the shortest distance between the query instance and class prototypes. We evaluate our model against existing deep learning based FSRE models using three biomedical datasets: the 2010 i2b2/VA challenge dataset, the CHEMPROT corpus, and the BioRED dataset, focusing on few-shot scenarios with limited training data. Our model demonstrates exceptional performance, achieving the highest accuracy across all datasets in most training configurations under a 3-way-5-shot condition and significantly surpassing the current state-of-the-art. Particularly, it achieves improvements ranging from 1.25% to 11.25% on the 2010 i2b2/VA challenge dataset, 3.4% to 20.2% on the CHEMPROT dataset, and 2.73% to 10.98% on the BioRED dataset compared to existing models. These substantial gains highlight the model’s robust generalization ability, enabling it to effectively handle previously unseen relations during testing. The demonstrated effectiveness of this approach underscores its potential for diverse medical applications, particularly in scenarios where acquiring extensive labeled data is challenging.}, archive = {J_NEUCOM}, author = {Guoxiu He and Chen Huang}, doi = {10.1016/j.neucom.2025.129752}, journal = {Neurocomputing}, month = {6}, pages = {129752}, shortjournal = {Neurocomputing}, title = {Few-shot medical relation extraction via prompt tuning enhanced pre-trained language model}, volume = {633}, year = {2025}, }
• (2025). Adaptive multi-model fusion learning for sparse-reward reinforcement learning. NEUCOM, 633, 129748. (www)

  @article{ , abstract = {In this paper, we address intrinsic reward generation for sparse-reward reinforcement learning, where the agent receives limited extrinsic feedback from the environment. Traditional approaches to intrinsic reward generation often rely on prediction errors from a single model, where the intrinsic reward is derived from the discrepancy between the model’s predicted outputs and the actual targets. This approach exploits the observation that less-visited state–action pairs typically yield higher prediction errors. We extend this framework by incorporating multiple prediction models and propose an adaptive fusion technique specifically designed for the multi-model setting. We establish and mathematically justify key axiomatic conditions that any viable fusion method must satisfy. Our adaptive fusion approach dynamically learns the best way to combine prediction errors during training, leading to improved learning performance. Numerical experiments validate the effectiveness of our method, showing significant performance gains across various tasks compared to existing approaches.}, archive = {J_NEUCOM}, author = {Giseung Park and Whiyoung Jung and Seungyul Han and Sungho Choi and Youngchul Sung}, doi = {10.1016/j.neucom.2025.129748}, journal = {Neurocomputing}, month = {6}, pages = {129748}, shortjournal = {Neurocomputing}, title = {Adaptive multi-model fusion learning for sparse-reward reinforcement learning}, volume = {633}, year = {2025}, }
• (2025). The SIFT based two-stage STC decoupled learning method for long-tailed SAR target recognition. NEUCOM, 633, 129747. (www)

  @article{ , abstract = {In remote sensing (RS) research, Deep Learning (DL) based methods have been widely applied to Synthetic Aperture Radar (SAR) auto target recognition with extraordinary success. However, as the real-world SAR images often exhibit imbalanced data distribution, numerous existing DL approaches fail to deliver optimal performance when the dataset conforms to a long-tailed distribution. The long-tailed characteristics lead in learning bias towards the majority classes, resulting in poor performance on the minority classes. In response to this issue, this paper proposes a Scale Invariant Feature Transform (SIFT) based two-stage decoupled learning method for long-tailed SAR target recognition, named STC decoupled learning. The proposed STC decoupled learning method is consist of two stages: the feature extraction learning stage and the classifier fine-tuning stage. Specifically, a supervised contrastive learning module is incorporated to obtain feature representation on long-tailed SAR samples at the first stage. Then, classed-balanced samples are trained for fine-tuning a linear-layer-based classifier with cross-entropy loss at the second stage. Besides, SIFT feature extraction module is employed on these two stages, which substantially leverages physical scattering attribution of SAR images. Extensive simulations on long-tailed SAR-AIRcraft-1.0, FUSAR and MSTAR datasets validate that the proposed method achieves state-of-the-art classification accuracy on long-tailed SAR target recognition tasks, as well as superior generalization ability towards varying scenarios. Our code is now available on https://github.com/XYLGroup/STC .}, archive = {J_NEUCOM}, author = {Ye Li and Jingyuan Xia and Huaizhang Liao and Xu Lan and Weidong Jiang}, doi = {10.1016/j.neucom.2025.129747}, journal = {Neurocomputing}, month = {6}, pages = {129747}, shortjournal = {Neurocomputing}, title = {The SIFT based two-stage STC decoupled learning method for long-tailed SAR target recognition}, volume = {633}, year = {2025}, }
• (2025). AA-mDLAM: An accelerated ADMM-based framework for training deep neural networks. NEUCOM, 633, 129744. (www)

  @article{ , abstract = {Stochastic gradient descent (SGD) and its many variants are the widespread optimization algorithms for training deep neural networks. However, SGD suffers from inevitable drawbacks, including vanishing gradients, lack of theoretical guarantees, and substantial sensitivity to input. The Alternating Direction Method of Multipliers (ADMM) has been proposed to address these shortcomings as an effective alternative to the gradient-based methods. It has been successfully employed for training deep neural networks. However, ADMM-based optimizers have a slow convergence rate. This paper proposes an accelerated framework for training deep neural networks, termed AA-mDLAM, which integrates Anderson acceleration within an Alternating Minimization approach inspired by ADMM to tackle this drawback. The main intention of the AA-mDLAM algorithm is to employ Anderson acceleration to alternating minimization by considering it as a fixed-point iteration and attaining a nearly quadratic convergence rate. We verify the effectiveness and efficiency of the proposed AA-mDLAM algorithm by conducting extensive experiments on seven benchmark datasets contrary to other state-of-the-art optimizers.}, archive = {J_NEUCOM}, author = {Zeinab Ebrahimi and Gustavo Batista and Mohammad Deghat}, doi = {10.1016/j.neucom.2025.129744}, journal = {Neurocomputing}, month = {6}, pages = {129744}, shortjournal = {Neurocomputing}, title = {AA-mDLAM: An accelerated ADMM-based framework for training deep neural networks}, volume = {633}, year = {2025}, }
• (2025). Knowledge graph embedding based on embedding permutation and high-frequency feature fusion for link prediction. NEUCOM, 633, 129743. (www)

  @article{ , abstract = {Knowledge graph embedding has excellent performance in capturing intrinsic relations and semantics in a wealth of information for link prediction. Knowledge graph embedding methods have achieved impressive results in recent years, especially those using convolutional neural networks. However, many previous approaches focus on interactions between relations and entities, ignoring interactions of internal data elements and the crucial role of high-frequency features. In this paper, we propose a novel approach, a knowledge graph Embedding model using 2D convolution operations integrating Embedding permutation strategy and High-frequency features fusion mechanism, named EHE, for link prediction. First, we design the embedding permutation mechanism for the embedding vectors. This mechanism leverages internal element permutation, efficiently broadening the local interactions of internal elements, especially for far-flung data elements in the one-dimensional space. Subsequently, a high-frequency feature fusion module is proposed to capture the high-frequency feature representations by using Sobel and Laplacian operators. Additionally, the projection attention mechanism is utilized to emphasize the unique semantic regions of interest in entities and relations. We assess our approach on several benchmark link prediction datasets. Considering the important metrics, MRR and H@1, our method achieves the overall best performance compared with existing state-of-the-art methods on five public datasets, showcasing its superior capacity for link prediction.}, archive = {J_NEUCOM}, author = {Qien Yu and Danilo Vasconcellos Vargas}, doi = {10.1016/j.neucom.2025.129743}, journal = {Neurocomputing}, month = {6}, pages = {129743}, shortjournal = {Neurocomputing}, title = {Knowledge graph embedding based on embedding permutation and high-frequency feature fusion for link prediction}, volume = {633}, year = {2025}, }
• (2025). Rehearsal-free continual few-shot relation extraction via contrastive weighted prompts. NEUCOM, 633, 129741. (www)

  @article{ , abstract = {The primary challenge in continual few-shot relation extraction is mitigating catastrophic forgetting. Prevailing strategies involve saving a set of samples in memory and replaying them. However, these methods pose privacy and data security concerns. To address this, we propose a novel rehearsal-free approach called Contrastive Weighted Prompt (CWP). This approach categorizes learnable prompts into task-generic and task-specific prompts. Task-generic prompts are shared across all tasks and are injected into the higher layers of the BERT encoder to capture general task knowledge. Task-specific prompts are generated by weighting all the prompts in a task-specific prompt pool based on their relevance to individual samples. These task-specific prompts are injected into the lower layers of BERT to extract task-specific knowledge. Task-generic prompts retain knowledge from prior tasks, while task-specific prompts reduce mutual interference among tasks and improve the relevance between prompts and individual samples. To further enhance the discriminability of the prompt embeddings for samples belonging to different relations, we introduced a relation-aware contrastive learning strategy. Experimental results on two standard datasets indicate that the proposed method outperforms baseline methods and demonstrates superiority in mitigating catastrophic forgetting.}, archive = {J_NEUCOM}, author = {Fengqin Yang and Mengen Ren and Delu Kong and Shuhua Liu and Zhiguo Fu}, doi = {10.1016/j.neucom.2025.129741}, journal = {Neurocomputing}, month = {6}, pages = {129741}, shortjournal = {Neurocomputing}, title = {Rehearsal-free continual few-shot relation extraction via contrastive weighted prompts}, volume = {633}, year = {2025}, }
• (2025). Dynamic neural learning for obstacle avoidance of humanoid robot performing cooperative tasks. NEUCOM, 633, 129727. (www)

  @article{ , abstract = {How to avoid the collision between a humanoid robot’s two manipulators is an important problem when two arms cooperate to perform the end task. In this paper, a dynamic neural learning for obstacle avoidance (DNLOA) scheme of a humanoid robot performing cooperative tasks is proposed. In this scheme, the trajectory tracking and mutual collision avoidance of the dual robot arms are converted to a quadratic programming (QP) framework, and the end-effector’s trajectory tracking, mutual collision avoidance and joint angle range are described as equality constraints or inequality constraints. Then the QP framework is solved by a simplified recurrent neural network (S-RNN). The framework is applied to the trajectory tracking tasks of drawing double crossed circles and writing Chinese characters with the humanoid robot’s dual manipulators. With the proposed DNLOA scheme, the tasks are done precisely without dual-arm mutual collision. Compared the scheme we proposed with the minimize velocity norm scheme (MVN) existed, experiments show that the proposed DNLOA scheme for humanoid robot is effective, accurate and practical.}, archive = {J_NEUCOM}, author = {Yamei Luo and Mingyang Zhang and Yu Liu and Junjie Lin and Zhijun Zhang}, doi = {10.1016/j.neucom.2025.129727}, journal = {Neurocomputing}, month = {6}, pages = {129727}, shortjournal = {Neurocomputing}, title = {Dynamic neural learning for obstacle avoidance of humanoid robot performing cooperative tasks}, volume = {633}, year = {2025}, }
• (2025). State prediction for multiple diffusion targets based on point pattern physics-informed neural network. NEUCOM, 633, 129714. (www)

  @article{ , abstract = {State prediction for multiple diffuse targets focuses on locating diffuse sources and concentration distributions. However, in real scenarios, problems such as multi-target overlap, unknown diffusion parameters, and difficult-to-measure environmental disturbances arise, causing the existing algorithms to have poor prediction accuracy. To address the above problems, this article proposes a multi-diffusion target state prediction algorithm based on point pattern physics-informed neural network (PP-PINN). This article targets the overlapping problem of multiple diffusion targets and uses the point pattern model to realize the accurate clustering of different diffusion targets and estimate their concentration distributions. Then, the physics-informed neural network is trained to estimate the diffusion parameters and modify the initial concentration distribution to reduce the environmental interference. The experimental results show that the proposed algorithm can significantly improve the performance of multi-diffusion target state prediction, and can provide important data support for the assessment of hazardous material leakage accidents.}, archive = {J_NEUCOM}, author = {Qiankun Sun and Lei Cai and Xiaochen Qin}, doi = {10.1016/j.neucom.2025.129714}, journal = {Neurocomputing}, month = {6}, pages = {129714}, shortjournal = {Neurocomputing}, title = {State prediction for multiple diffusion targets based on point pattern physics-informed neural network}, volume = {633}, year = {2025}, }
• (2025). FMVP: Fine-grained meta visual prompt enabled domain-specific few-shot classification. NEUCOM, 633, 129688. (www)

  @article{ , abstract = {Few-shot learning is a challenging and promising fundamental research. Inspired by recent advances in large language models (LLMs), visual prompt tuning has achieved notable performance gains in few-shot tasks by introducing only limited trainable parameters in the input space. Though effective, prompt tuning in few-shot settings heavily relies on well-initialized soft prompts and often lacks generalizability. Additionally, in certain specific fields, particularly in agriculture, there is a lack of high-precision fine-grained few-shot classification models. To our knowledge, this study is the first to employ prompt tuning for fine-grained few-shot plant disease classification ( specific to disease severity ). Specifically, we propose a novel F ine-grained M eta V isual P rompt tuning (FMVP) framework to systematically explore how visual prompts can enhance the generalizability of fine-grained few-shot domain-specific models. Firstly, a S parsity-aware M eta V isual P rompt tuning (SMVP) sub-module is proposed to learn a universal visual prompt initialization. SMVP utilizes pixel-level optimizable visual prompts for input transformation, jointly with a novel sparsity-aware meta-learning paradigm for parameter updating, boosting generalizability to unseen classes. Secondly, a F ine-grained C ross- A lignment (FCA) module is introduced to explore intra- and inter-image relational patterns, enhancing fine-grained recognition by extracting object-level cross-image semantic discriminative features. Extensive experiments on datasets such as mini -ImageNet, CUB, and FPV have shown that our model outperforms state-of-the-art (SOTA) models. Our work constitutes a valuable addition to domain-specific models for practical applications.}, archive = {J_NEUCOM}, author = {Minghui Li and Hongxun Yao}, doi = {10.1016/j.neucom.2025.129688}, journal = {Neurocomputing}, month = {6}, pages = {129688}, shortjournal = {Neurocomputing}, title = {FMVP: Fine-grained meta visual prompt enabled domain-specific few-shot classification}, volume = {633}, year = {2025}, }
• (2025). Evolutionary multi-objective neural architecture search via depth equalization supernet. NEUCOM, 633, 129674. (www)

  @article{ , abstract = {To provide a diverse selection of models suitable for different application scenarios, neural architecture search (NAS) is constructed as a multi-objective optimization problem aiming to simultaneously optimize multiple metrics such as model size and accuracy. Evolutionary algorithms (EA) have been shown to be an effective multi-objective approach that can balance different metrics. However, EA require many evaluations, and the evaluation of architectures is expensive. Training a supernet to evaluate an architecture is considered a promising method to reduce the cost of EA. But there are still many challenges in applying supernet to multi-objective NAS: (1) Supernet tends to give higher scores to shallower architectures, causing potential deeper architectures to be ignored. (2) The receptive field of the architecture has a large gap between search and evaluation, causing a decrease in performance. (3) Larger models are gradually eliminated during evolution, leading to a diversity disaster. We proposed a framework called DESEvo to solve these problems in this paper. DESEvo trains a depth equalization supernet to improve bias of supernet via a frequency rejection sampling method. In addition, DESEvo adaptively constrainted receptive field of architecture to reduce the gap. Finally, DESEvo developed a diversity-preserving strategy to enhance the diversity. Experimental results validate the efficiency and effectiveness of the algorithm, DESEvo can search a set of architectures that are more competitive compared to other state-of-the-art algorithms within 0.2 days, becoming the most efficient multi-objective NAS method in the supernet-based methods.}, archive = {J_NEUCOM}, author = {Juan Zou and Yang Liu and Yuan Liu and Yizhang Xia}, doi = {10.1016/j.neucom.2025.129674}, journal = {Neurocomputing}, month = {6}, pages = {129674}, shortjournal = {Neurocomputing}, title = {Evolutionary multi-objective neural architecture search via depth equalization supernet}, volume = {633}, year = {2025}, }
• (2025). Evolving J-CO-QL+ with fuzzy evaluators for flexible queryisng of JSON data sets. NEUCOM, 633, 129621. (www)

  @article{ , abstract = {How to introduce soft querying based on fuzzy sets in the novel J-CO-QL + query language (specifically designed to query collections of JSON documents from NoSQL databases) has been investigated by the authors in their past work. Specifically, capabilities for defining fuzzy operators and fuzzy aggregators were introduced through two distinct concepts on which two different language constructs were based. This paper proposes the unified concept of “fuzzy evaluator”, by means of which it is possible to define complex methods for evaluating the membership degrees of JSON documents to fuzzy sets, so as to capture complex semantics while analyzing data in a soft way. The paper both provides a formal meta-model for fuzzy evaluators, and proposes a novel statement for the J-CO-QL + language, so as to further foster soft-querying capabilities.}, archive = {J_NEUCOM}, author = {Paolo Fosci and Giuseppe Psaila}, doi = {10.1016/j.neucom.2025.129621}, journal = {Neurocomputing}, month = {6}, pages = {129621}, shortjournal = {Neurocomputing}, title = {Evolving J-CO-QL+ with fuzzy evaluators for flexible queryisng of JSON data sets}, volume = {633}, year = {2025}, }
• (2025). Structural investigations of multi-reservoir echo state networks for EEG-based emotion classification. NEUCOM, 632, 129856. (www)

  @article{ , abstract = {Echo State Networks (ESNs) employ non-gradient-based learning mechanisms that bypass backpropagation and pre-training, significantly reducing computational costs and training times, thus facilitating their widespread use in EEG analysis. To date, existing ESN architectures employed for EEG signal processing predominantly utilize single-layer reservoir structures, with limited research exploring the application of multi-reservoir ESN configurations that possess enhanced feature representation capabilities. This study addresses these gaps by proposing and validating three novel multi-reservoir ESN architectures: Pyramid ESN, Inverse Pyramid ESN, and Hourglass ESN, specifically for EEG-based emotion classification. Furthermore, the study introduces a concatenation mechanism that integrates the original data with the outputs from multiple reservoirs, thereby improving the representation power of multi-reservoir ESNs. The results indicate that the Hourglass ESN, which incorporates this concatenation mechanism, achieves the highest performance, yielding classification accuracies of 94.2 % and 94.3 % for arousal and valence, respectively, in binary classification tasks on the DEAP dataset. In four-class and eight-class tasks, accuracies reached 83.6 % and 78.7 %, respectively. Notably, the model demonstrates a 35.52 % average reduction in parameter complexity compared to single-layer ESNs, while achieving performance levels that meet the current state-of-the-art (SOTA) benchmarks for binary and four-class tasks, and exceed SOTA for the eight-class task. This study also investigates the impact of reservoir size on EEG-based emotion classification performance, revealing a saturation effect that provides valuable insights for the parameter design of multi-reservoir ESN models, enabling an optimal balance between computational efficiency and classification performance.}, archive = {J_NEUCOM}, author = {Yang Liu and Ruiqi Liang and Shule Xu and Xiang Guo}, doi = {10.1016/j.neucom.2025.129856}, journal = {Neurocomputing}, month = {6}, pages = {129856}, shortjournal = {Neurocomputing}, title = {Structural investigations of multi-reservoir echo state networks for EEG-based emotion classification}, volume = {632}, year = {2025}, }
• (2025). High-security image steganography integrating multi-scale feature fusion with residual attention mechanism. NEUCOM, 632, 129838. (www)

  @article{ , abstract = {Constructing a good cost function is crucial for minimizing embedding distortion in image steganography. Recently, deep learning-based adaptive cost learning in image steganography has achieved significant advancements. For GAN-based image steganography, an encoder-decoder structure is typically employed by the generator. However, the continual encoding process often results in a lack of detailed information. Even if the image resolution is restored through skip connections, the generator will still be limited. To address the issue, this paper proposes a novel GAN structure named UMSA-GAN. Firstly, we design a residual attention mechanism, Res-CBAM, integrated into the generator network, which enables focusing on high-frequency regions in the cover image. Secondly, multi-scale feature information is also fused using skip connections, which enables the generator to learn more shallow features. Finally, unlike most of the previous works that only utilized Xu-Net as the discriminator, dual steganalyzers are also introduced as the discriminator to further enhance performance. Extensive comparative experiments demonstrate that UMSA-GAN effectively learns features from the cover images and generates better embedding probability maps. Compared to traditional and state-of-the-art GAN-based steganographic methods, UMSA-GAN exhibits superior security performance. In addition, the rationality and superiority of UMSA-GAN are further verified by a large number of ablation studies.}, archive = {J_NEUCOM}, author = {Jiaqi Liang and Wei Xie and Haotian Wu and Junfeng Zhao and Xianhua Song}, doi = {10.1016/j.neucom.2025.129838}, journal = {Neurocomputing}, month = {6}, pages = {129838}, shortjournal = {Neurocomputing}, title = {High-security image steganography integrating multi-scale feature fusion with residual attention mechanism}, volume = {632}, year = {2025}, }
• (2025). Semi-supervised medical image segmentation using spiking neural P-like convolutional model and pseudo label-guided cross-patch contrastive learning. NEUCOM, 632, 129782. (www)

  @article{ , abstract = {Deep learning-based methods have achieved significant improvements in medical image segmentation in recent years. Due to the large amount of labeled data required for neural network training and the difficulty of medical image labeling, semi-supervised learning (SSL) has received great attention in medical image segmentation. However, existing SSL methods often fail to make full use of unlabeled data. Therefore, this study proposes a novel semi-supervised medical image segmentation framework that aims to produce more accurate predictions from unlabeled data. First, a simple and efficient segmentation network is designed, called UNet-ReS. It uses ResNet as the encoder to extract features, and uses spiking neural P-like convolutional neuron model inspired by nonlinear spiking neural P systems to build the decoder. UNet-ReS can generate high-quality unlabeled data prediction in a semi-supervised framework while obtaining reliable pseudo-label. Second, a pseudo-label guided cross-patch contrastive learning loss is proposed to improve the feature representation of deep semantic information from different classes in segmentation prediction. The intra-class aggregation and inter-class separability of deep semantic features from different classes are improved by minimizing the intra-class distance and maximizing the inter-class distance. This helps to improve the overall accuracy of segmentation predictions. The proposed method is validated on three different types of public datasets, including ACDC, Kvasir-SEG, and CRAG. The experimental results show that the proposed method outperforms other semi-supervised segmentation methods.}, archive = {J_NEUCOM}, author = {Chi Zhou and Lulin Ye and Hong Peng and Jun Wang and Zhicai Liu}, doi = {10.1016/j.neucom.2025.129782}, journal = {Neurocomputing}, month = {6}, pages = {129782}, shortjournal = {Neurocomputing}, title = {Semi-supervised medical image segmentation using spiking neural P-like convolutional model and pseudo label-guided cross-patch contrastive learning}, volume = {632}, year = {2025}, }
• (2025). Event-based video reconstruction via attention-based recurrent network. NEUCOM, 632, 129776. (www)

  @article{ , abstract = {Event cameras are novel sensors that capture brightness changes in the form of asynchronous events rather than intensity frames, offering unique advantages such as high dynamic range, high temporal resolution, and no motion blur. However, the sparse, asynchronous nature of event data poses significant challenges for visual perception, limiting compatibility with conventional computer vision algorithms that rely on dense, continuous frames. Event-based video reconstruction has emerged as a promising solution, though existing methods still face challenges in capturing fine-grained details and enhancing contrast. This paper presents a novel approach to video reconstruction from asynchronous event streams, leveraging the unique properties of event data to produce high-quality video. Our method integrates channel and pixel attention mechanisms to focus on essential features and incorporates deformable convolutions and adaptive mix-up operations to provide flexible receptive fields and dynamic fusion across down-sampling and up-sampling layers. Experimental results on multiple real-world event datasets demonstrate that our approach outperforms comparable methods trained on the same dataset, achieving superior video quality from pure event data. We also demonstrate the capability of our method for high dynamic range reconstruction and color video reconstruction using an event camera equipped with a Bayer filter.}, archive = {J_NEUCOM}, author = {Wenwen Ma and Shanxing Ma and Pieter Meiresone and Gianni Allebosch and Wilfried Philips and Jan Aelterman}, doi = {10.1016/j.neucom.2025.129776}, journal = {Neurocomputing}, month = {6}, pages = {129776}, shortjournal = {Neurocomputing}, title = {Event-based video reconstruction via attention-based recurrent network}, volume = {632}, year = {2025}, }
• (2025). Bilateral-aware and multi-scale region guided U-net for precise breast lesion segmentation in ultrasound images. NEUCOM, 632, 129775. (www)

  @article{ , abstract = {Breast cancer ranks among the leading health risks for women globally, with the significance of early diagnosis and intervention being paramount. Using computers to segment breast lesions from ultrasound images serves as a crucial auxiliary tool for diagnosing and studying this disease. However, the effectiveness of breast tumor segmentation is deeply impacted by the severe artifacts, the low contrast, and the diverse tumor shapes found in breast ultrasound images. The focus of this study is to devise an effective strategy to further alleviate the aforementioned issues and achieve more precise lesion segmentation. Specifically, we designed a Bilateral-Aware and Multi-Scale Region Guided U-Net (BA-MRGU-Net) with a bilateral perception strategy to segment breast tumors. Initially, we devised a Foreground and Background Aware Module (FBAM), primarily composed of an Adaptive Spatial Selection Unit (ASSU) and a Background Suppression Unit (BSU). The ASSU can help the network capture spatial context information that is more relevant to lesions. Concurrently, the BSU suppresses the feature responses of ultrasound artifacts and other tissues in the background. The FBAM can effectively distinguish between the foreground and background through these two independent branches. Subsequently, we developed a Multi-Scale Region Guided Module (MRGM) to utilize the feature maps across various scales to boost the network’s perception of the lesion region. We executed a range of experiments utilizing two widely used datasets with several state-of-the-art algorithms. The results reveal that our approach achieves improvements of varying degrees on multiple evaluation metrics and has superior segmentation performance.}, archive = {J_NEUCOM}, author = {Yangyang Li and Xintong Hou and Xuanting Hao and Ronghua Shang and Licheng Jiao}, doi = {10.1016/j.neucom.2025.129775}, journal = {Neurocomputing}, month = {6}, pages = {129775}, shortjournal = {Neurocomputing}, title = {Bilateral-aware and multi-scale region guided U-net for precise breast lesion segmentation in ultrasound images}, volume = {632}, year = {2025}, }
• (2025). An o(1/k) algorithm for multi-agent optimization with inequality constraints. NEUCOM, 632, 129770. (www)

  @article{ , abstract = {This paper presents a discrete-time solution algorithm for a constrained multi-agent optimization problem with inequality constraints. Its aim is to seek a solution to minimize the sum of all the agents’ objective functions while satisfy each agent’s local set constraint and nonlinear inequality constraints. Assume that agents’ local constraints are heterogeneous and all the objective functions are convex and continuous, but they may not be differentiable. Similar to the distributed alternating direction method of multipliers (ADMM) algorithm, the designed algorithm can solve the multi-agent optimization problem in a distributed manner and has a fast O ( 1 / k ) convergence rate. Moreover, it can deal with the nonlinear constraints, which cannot be handled by distributed ADMM algorithm. Finally, the proposed algorithm is applied to solve a robust linear regression problem, a lasso problem and a decentralized joint flow and power control problem with inequality constraints, respectively and thus the effectiveness of the proposed algorithm is verified.}, archive = {J_NEUCOM}, author = {Peng Li and Yiyi Zhao and Jiangping Hu and Jiangtao Ji}, doi = {10.1016/j.neucom.2025.129770}, journal = {Neurocomputing}, month = {6}, pages = {129770}, shortjournal = {Neurocomputing}, title = {An o(1/k) algorithm for multi-agent optimization with inequality constraints}, volume = {632}, year = {2025}, }
• (2025). Words shaping worlds: A comprehensive exploration of text-driven image and video generation with generative adversarial networks. NEUCOM, 632, 129767. (www)

  @article{ , abstract = {A growing interest in education, media, and entertainment has increased Artificial Intelligence-powered content generation, mainly via Generative Adversarial Networks (GANs), which shape images, videos, audio, and text. A generative adversarial network (GAN) is the combination of two deep neural networks: generator ( G ) and discriminator ( D ). These two components are trained competitively by pitting one against the other such that G generates new data while D authenticates the data. By leveraging powerful deep neural networks and competitive training, GANs can synthesize reasonable and realistic images and videos from the text description. This paper extensively reviews the recent state-of-the-art GAN models for text-to-image (T2I) and text-to-video (T2V) synthesis. In this regard, databases like ACM, IEEE Explore, Web of Science, and ScienceDirect were searched to find and analyze the relevant research articles conducted in this area in the last decade, specifically from 2014 to 2024. Secondly, T2I and T2V GAN methods were classified according to structure and functionality. Later, a comprehensive evaluation between T2I and T2V GAN-based methods was conducted, employing various qualitative and quantitative evaluation techniques. Finally, the paper concludes by discussing multiple applications, main challenges, and limitations of T2I and T2V GAN models for future consideration.}, archive = {J_NEUCOM}, author = {Anwar Ullah and Muhammad Numan and Mohd Nor Akmal Khalid and Abdul Majid}, doi = {10.1016/j.neucom.2025.129767}, journal = {Neurocomputing}, month = {6}, pages = {129767}, shortjournal = {Neurocomputing}, title = {Words shaping worlds: A comprehensive exploration of text-driven image and video generation with generative adversarial networks}, volume = {632}, year = {2025}, }
• (2025). TDCR: Transformer based decision conflict resolution model for collaborative scheduling. NEUCOM, 632, 129760. (www)

  @article{ , abstract = {With the strong adaptability and iterative capabilities of artificial intelligent algorithms, more and more enterprises are adopting human–machine collaboration to replace traditional enterprise management methods. So in recent years, the use of evolutionary algorithms and reinforcement learning methods to solve multi-objective optimization has made many breakthroughs, but in the intelligent manufacturing industry, such as production scheduling and inventory management and other aspects of decision optimization, because of the actual situation is too complicated, it is not conducive to determine the appropriate constraints, and then affect the performance of the model. This paper proposes a multi-decision optimization model for collaborative scheduling, which comprises two key components: a conflict resolution strategy module and a decision making module. This paper uses the attention mechanism to generate decision preference vectors in different manufacturing scenarios, so that the conflict resolution strategy can be dynamically changed and adds the keyword mask method close to downstream tasks during training to further improve the performance of the model. Finally, we evaluate the performance of our model in the conflict resolution task by selecting multiple data sets from multiple public data sets, and show satisfactory performance in this task, showing robustness in different scenarios. This study provides a valuable reference for conflict resolution between decision making.}, archive = {J_NEUCOM}, author = {Xiancheng Hu and Jian An and Xiaolin Gui and Long Chen and Siyu Tang and Xin He}, doi = {10.1016/j.neucom.2025.129760}, journal = {Neurocomputing}, month = {6}, pages = {129760}, shortjournal = {Neurocomputing}, title = {TDCR: Transformer based decision conflict resolution model for collaborative scheduling}, volume = {632}, year = {2025}, }
• (2025). Person verification and recognition by combining voice signal and online handwritten signature using hyperbolic function based transformer neural network. NEUCOM, 632, 129751. (www)

  @article{ , abstract = {With the increasing popularity of smart portable electronic gadgets, various voice based online person verification systems have been developed. However, these systems are susceptible to attacks where an illegitimate individual feeds a recorded voice of a legitimate person, resulting in false confirmations. To overcome this limitation of voice based person verification systems, this article proposes a hyperbolic function based encoded representation of transformer neural network (ERTNN) framework for person verification and recognition by combining online handwritten signature of the genuine person with his/her voice signal. The proposed hyperbolic function based ERTNN framework for person verification and recognition consists of one multi-headed attention layer with positional encoding, seven convolution layers with skip connections, two dense layers, and an output layer. The positional encoding scheme in the proposed hyperbolic function based ERTNN framework has been implemented using hyperbolic s i n e and hyperbolic c o s i n e functions. The mel-frequency cepstral coefficients (MFCC), MFCC-delta, and MFCC-delta-delta features of the voice signal have been combined with all the temporal points of online handwritten signature sample of a person to make a combined feature matrix. The combined feature matrix of voice signal and online handwritten signature has been fed as an input to the proposed framework to verify and recognize a person corresponding to the input feature matrix. The novelty of this work lies in proposing the hyperbolic function based positional encoding scheme in the ERTNN framework . The experiments have also been carried out using traditional ERTNN framework employing sinusoidal function based positional encoding scheme, learnable positional encoding based ERTNN framework, relative positional encoding based ERTNN framework as well as by removing the positional encoding scheme from the multi-headed attention layer to have a performance comparison with the proposed framework.}, archive = {J_NEUCOM}, author = {Rohitesh Kumar and Rajib Ghosh}, doi = {10.1016/j.neucom.2025.129751}, journal = {Neurocomputing}, month = {6}, pages = {129751}, shortjournal = {Neurocomputing}, title = {Person verification and recognition by combining voice signal and online handwritten signature using hyperbolic function based transformer neural network}, volume = {632}, year = {2025}, }
• (2025). Transref: Multi-scale reference embedding transformer for reference-guided image inpainting. NEUCOM, 632, 129749. (www)

  @article{ , abstract = {Image inpainting for completing complicated semantic environments and diverse hole patterns of corrupted images is challenging even for state-of-the-art learning-based inpainting methods trained on large-scale data. A reference image capturing the same scene of a corrupted image offers informative guidance for completing the corrupted image as it shares similar texture and structure priors to that of the holes of the corrupted image. In this work, we propose a Trans former-based encoder–decoder network for Ref erence-guided image inpainting, named TransRef . Specifically, the guidance is conducted progressively through a reference embedding procedure, in which the referencing features are subsequently aligned and fused with the features of the corrupted image. For precise utilization of the reference features for guidance, a reference-patch alignment (Ref-PA) module is proposed to align the patch features of the reference and corrupted images and harmonize their style differences, while a reference-patch transformer (Ref-PT) module is proposed to refine the embedded reference feature. Moreover, to facilitate the research of reference-guided image restoration tasks, we construct a publicly accessible benchmark dataset containing 50K pairs of input and reference images. Both quantitative and qualitative evaluations demonstrate the efficacy of the reference information and the proposed method over the state-of-the-art methods in completing complex holes. Code and dataset can be accessed at: https://github.com/Cameltr/TransRef .}, archive = {J_NEUCOM}, author = {Taorong Liu and Liang Liao and Delin Chen and Jing Xiao and Zheng Wang and Chia-Wen Lin and Shin’ichi Satoh}, doi = {10.1016/j.neucom.2025.129749}, journal = {Neurocomputing}, month = {6}, pages = {129749}, shortjournal = {Neurocomputing}, title = {Transref: Multi-scale reference embedding transformer for reference-guided image inpainting}, volume = {632}, year = {2025}, }
• (2025). Segmentation of 3D neuronal morphologies in microscopy images utilizing flexible open-curve snakes. NEUCOM, 632, 129742. (www)

  @article{ , abstract = {Active contour models, known as snakes, have effectively shown their capability to track and recreate neurites in three dimensions. Nevertheless, their efficacy is constrained when addressing attenuated filament signals tainted by noise. Moreover, tracking outcomes are contingent upon initial seeds, as they depend exclusively on forces generated from image gradients. To address these issues and enhance the traditional OCS tracker through the utilization of advanced algorithms, we present the Flexible Open Curve Snakes (FOCS). This unique framework synthesizes various forces via energy minimization and facilitates mutual reinforcement for localization, directional determination, and concurrent radii estimation to finalize the neurites’ structure. In FOCS, the open-curve tracking process undergoes dynamic deformation via iterative techniques that incorporate new deformation energy, modifications to strain orientations, and alterations in local radii. This method is perpetually enhanced by reducing a controllable energy function that encompasses the fitting forces and the curve’s length. FOCS effectively integrates snake architectures into a tracing framework, capturing the extensive information in volumetric neural data, hence addressing the issues of segmenting, tracing, and fully reconstructing neural structures in their natural environment. We illustrate the remarkable efficacy of FOCS by assessing many datasets, including BigNeuron (human and mouse) and Diadem. In these assessments, FOCS regularly surpassed current neuron tracking and tracing methodologies, achieving superior performance. Our approach significantly enhances performance measures, yielding gains of approximately 1.7% and 17% in the average overlap and distance scores on the BigNeuron dataset, respectively. Furthermore, it demonstrates a significant enhancement of roughly 4.1% in the average overlap score on the DIADEM dataset. The suggested approach is applicable to confocal and two-photon microscopy datasets, as demonstrated by the datasets employed.}, archive = {J_NEUCOM}, author = {Amir Vatani and Jie Song and Liang Xiao}, doi = {10.1016/j.neucom.2025.129742}, journal = {Neurocomputing}, month = {6}, pages = {129742}, shortjournal = {Neurocomputing}, title = {Segmentation of 3D neuronal morphologies in microscopy images utilizing flexible open-curve snakes}, volume = {632}, year = {2025}, }
• (2025). LSTM and GRU type recurrent neural networks in model predictive control: A review. NEUCOM, 632, 129712. (www)

  @article{ , abstract = {Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) neural networks are known for their capability of modeling numerous dynamical phenomena. Model Predictive Control (MPC) refers to a family of advanced control methods in which a dynamical model predicts online the future behavior of the controlled process, and an optimization procedure finds the best control policy. From the point of view of the control quality and the computational complexity of MPC, two issues are crucial: the model structure and the way the model is used in MPC. Both factors determine the resulting control quality and computational complexity of MPC. This article reviews possible methods of using LSTM and GRU type networks in MPC. First, we characterize several model variants that are utilized in MPC. Next, we review possible approaches to MPC based on LSTMs and GRUs, particularly the MPC methods leading to low computational complexity. Stability and robustness issues are also discussed. For a chemical pH reactor, the efficiency of LSTM and GRU models and a few neural network-based MPC algorithms are compared. Finally, we review numerous applications, including applications to real processes, hardware-in-the-loop solutions and example simulation studies.}, archive = {J_NEUCOM}, author = {Maciej Ławryńczuk and Krzysztof Zarzycki}, doi = {10.1016/j.neucom.2025.129712}, journal = {Neurocomputing}, month = {6}, pages = {129712}, shortjournal = {Neurocomputing}, title = {LSTM and GRU type recurrent neural networks in model predictive control: A review}, volume = {632}, year = {2025}, }
• (2025). SHoTGCN: Spatial high-order temporal GCN for skeleton-based action recognition. NEUCOM, 632, 129697. (www)

  @article{ , abstract = {Action recognition algorithms that leverage human skeleton motion data are highly attractive due to their robustness and high information density. Currently, the majority of algorithms in this domain employ graph convolutional neural networks (GCNs). However, these algorithms often neglect the extraction of high-order features. To address this limitation, we propose a novel approach called the Spatial High-Order Temporal Graph Convolution Network (SHoTGCN), designed to evaluate the impact of high-order features on human action recognition. Our method begins by deriving high-order features from human skeleton time series data through temporal interactions. Utilizing these high-order features significantly improves the algorithm’s ability to recognize human actions. Moreover, we found that the traditional feature extraction method, which employs Depthwise Convolution (DWConv) with a single 2D convolution, is suboptimal compared to a multibranch structure for feature extraction. To address this, we introduce a structure re-parameterization technique with DWConv, termed Rep-tDWConv, to enhance feature extraction. By integrating the Exponential Moving Average (EMA) model during the model fusion process, our proposed model achieves state-of-the-art (SOTA) performance, with accuracies of 90.4% and 92.0% on the XSub and XSet splits of the NTU RGB+D 120 dataset, respectively.}, archive = {J_NEUCOM}, author = {Qiyu Liu and Ying Wu and Bicheng Li and Yuxin Ma and Hanling Li and Yong Yu}, doi = {10.1016/j.neucom.2025.129697}, journal = {Neurocomputing}, month = {6}, pages = {129697}, shortjournal = {Neurocomputing}, title = {SHoTGCN: Spatial high-order temporal GCN for skeleton-based action recognition}, volume = {632}, year = {2025}, }
• (2025). MPDRM: A multi-scale personalized depression recognition model via facial movements. NEUCOM, 632, 129669. (www)

  @article{ , abstract = {Automatic depression recognition based on facial movements in videos has become a research hotspot. However, existing methods tend to confuse individual inherent facial behavioral habits with characteristics specific to depression, which leads to misjudgments. To address this, we propose a Multi-scale Personalized Depression Recognition Model (MPDRM) that mitigates the negative impact of individual differences, enabling the model to focus on general and robust facial depression cues. The proposed model consists of three main components: the Multi-scale Depression Feature Network (MDFN), the Multi-scale Personality Feature Network (MPFN), and the Relational Attention Recognition Module (RARM). The MDFN extracts depression-related information, while the contrastive learning-based MPFN extracts stable personalized information. In both MDFN and MPFN, we insert the Multi-scale Motion Pattern Extraction Module (MMP) to capture rich multi-scale spatiotemporal facial features. Finally, the RARM is designed to enhance the representation of depression and output the results. Cross-validation on a specifically constructed longitudinal dataset demonstrates that our model outperforms other models. Experimental results indicate that suppressing personalized information of facial movements can effectively improve the accuracy of depression recognition.}, archive = {J_NEUCOM}, author = {Zhenyu Liu and Bailin Chen and Shimao Zhang and Jiaqian Yuan and Yang Wu and Hanshu Cai and Xin Chen and Lin Liu and Yimiao Zhao and Huan Mei and Jiahui Deng and Yanping Bao and Bin Hu}, doi = {10.1016/j.neucom.2025.129669}, journal = {Neurocomputing}, month = {6}, pages = {129669}, shortjournal = {Neurocomputing}, title = {MPDRM: A multi-scale personalized depression recognition model via facial movements}, volume = {632}, year = {2025}, }
• (2025). A comprehensive overview of generative AI (GAI): Technologies, applications, and challenges. NEUCOM, 632, 129645. (www)

  @article{ , abstract = {Generative Artificial Intelligence (GAI) represents a forefront research domain and demonstrates the ability to generate innovative and creative content spanning text, images, audio, videos, and other technological forms. Recent breakthroughs in GAI, exemplified by remarkable products like ChatGPT and stable diffusion, have garnered significant attention and hold immense potential to shape the trajectory of societal development. This paper undertakes a comprehensive analysis of the current capabilities and limitations of GAI while exploring optimal strategies for its future application. Specifically, we provide an extensive overview of technical approaches employed in GAI, encompassing text, images, videos, audio, and multi-modal generation models. Furthermore, we summarize the commonly utilized training datasets and evaluation benchmarks of various modalities. These benchmarks serve as integral components for assessing the performance of GAI models. Subsequently, we delve into the current applications and future prospects of GAI across various fields. Finally, we discuss the challenges inherent in GAI and outline prospective directions for future advancements in the field, with the intention of offering valuable insights and inspiration to researchers.}, archive = {J_NEUCOM}, author = {Zhijun Zhang and Jian Zhang and Xiaodong Zhang and Weijian Mai}, doi = {10.1016/j.neucom.2025.129645}, journal = {Neurocomputing}, month = {6}, pages = {129645}, shortjournal = {Neurocomputing}, title = {A comprehensive overview of generative AI (GAI): Technologies, applications, and challenges}, volume = {632}, year = {2025}, }
• (2025). Model adaptive parameter fine-tuning based on contribution measure for image classification. NEUCOM, 632, 129634. (www)

  @article{ , abstract = {Fine-tuning is an important transfer learning technique that has achieved significant success in various image classification tasks lacking training data and requires only a small number of training epochs to achieve satisfactory results. However, with the increasing complexity of the model scale and structure, designing appropriate fine-tuning schemes for specific target tasks becomes increasingly difficult. In this paper, a contribution measure criterion is used to quantify the importance of the pre-trained model parameters to the target task, providing a basis for selecting fine-tuning parameters. In addition, we find that the fine-tuning ratio vary depends on the specific target task. Therefore, we propose an adaptive fine-tuning ratio search strategy to search the appropriate fine-tuning ratio for the given target task. Based on the above strategy, we propose an adaptive fine-tuning algorithm based on parameter contribution to customize the fine-tuning scheme for the target task. The experimental results show that the proposed algorithm can effectively quantify the contribution of model parameters, and our algorithm can adaptively adjust the fine-tuning ratio for the target task. Furthermore, our algorithm achieves state-of-the-art performance on seven publicly available image classification datasets widely used in transfer learning.}, archive = {J_NEUCOM}, author = {Le Feng and Fujian Feng and Yuan Yang and Mian Tan and Lin Wang}, doi = {10.1016/j.neucom.2025.129634}, journal = {Neurocomputing}, month = {6}, pages = {129634}, shortjournal = {Neurocomputing}, title = {Model adaptive parameter fine-tuning based on contribution measure for image classification}, volume = {632}, year = {2025}, }
• (2025). Zero-shot detection of LLM-generated text via text reorder. NEUCOM, 631, 129829. (www)

  @article{ , abstract = {With the rapid advancement of large language model (LLM) technology, particularly with the emergence of advanced models like ChatGPT, distinguishing between LLM-generated and human-written texts has become increasingly challenging. This phenomenon presents unprecedented challenges to academic integrity and authenticity, making the detection of LLM-generated content a pressing concern in scientific research. To effectively and accurately detect texts generated by LLMs, this study constructs a comprehensive dataset of medical paper introductions, encompassing both human-written and LLM-generated content. Based on this dataset, a simple and efficient black-box, zero-shot detection method is proposed. The method builds upon the hypothesis that fundamental differences exist in the linguistic logical ordering between human-written and LLM-generated texts. Specifically, this method reorders the original text using dependency parse trees, calculates the similarity score (Rscore) between the reordered text and the original, and integrates log-likelihood features as auxiliary metrics. The approach synthesizes the reordered similarity and log-likelihood scores to derive a composite metric, establishing an effective classification threshold for discriminating between human-written and LLM-generated texts. The experimental results show that our approach not only effectively detects LLM-generated texts but also identifies LLM-polished abstracts, outperforming current state-of-the-art zero-shot detection methods (SOTA).}, archive = {J_NEUCOM}, author = {Jingtao Sun and Zhanglong Lv}, doi = {10.1016/j.neucom.2025.129829}, journal = {Neurocomputing}, month = {5}, pages = {129829}, shortjournal = {Neurocomputing}, title = {Zero-shot detection of LLM-generated text via text reorder}, volume = {631}, year = {2025}, }
• (2025). Fine-grained text and image guided point cloud completion with CLIP model. NEUCOM, 631, 129768. (www)

  @article{ , abstract = {In this work, we propose a novel multimodal fusion network for point cloud completion, which can simultaneously combine visual and textual information to accurately predict the semantic and geometric characteristics of incomplete shapes in a more generalizable manner. Specifically, to overcome the limited prior information due to the small-scale dataset, we utilize a pre-trained vision-language model trained on a large dataset of image-text pairs. Consequently, the textual and visual encoders of this large-scale model exhibit a stronger generalization ability. Then, we introduce a multi-stage feature fusion strategy to progressively incorporate textual and visual features into the backbone network. To further explore the effectiveness of fine-grained text descriptions, we build a text corpus with fine-grained descriptions, providing richer geometric details for 3D shapes. These detailed text descriptions are used for training and evaluating our network. Extensive quantitative and qualitative experiments demonstrate the superior performance of our method compared to state-of-the-art point cloud completion networks. Code is available at https://github.com/songwei100110/FTPNet .}, archive = {J_NEUCOM}, author = {Jun Zhou and Wei Song and Mingjie Wang and Hongchen Tan and Nannan Li and Xiuping Liu}, doi = {10.1016/j.neucom.2025.129768}, journal = {Neurocomputing}, month = {5}, pages = {129768}, shortjournal = {Neurocomputing}, title = {Fine-grained text and image guided point cloud completion with CLIP model}, volume = {631}, year = {2025}, }
• (2025). Towards multi-fusion graph neural network for single-cell RNA sequence clustering. NEUCOM, 631, 129764. (www)

  @article{ , abstract = {Clustering analysis plays a crucial role in single-cell RNA sequencing (scRNA-seq) data analysis, in which the graph neural network (GNN)-based clustering methods have rapidly emerged as a promising technique. Despite considerable progress, the previous scRNA-seq clustering methods still suffer from two critical limitations. First, they mostly treat the node attributes and cell–cell topological information equally, neglecting their (probably) different reliability. Second, they usually only consider the learned representation of the last layer, lacking the ability to fuse multi-scale discriminative information embedded in different layers. In view of this, this paper presents a new single-cell multi-fusion graph neural network (scMFGNN) for scRNA-seq clustering. Particularly, we utilize a multi-fusion graph neural network (MFGNN) for learning discriminative representations while preserving the structural information latent in multi-scale network layers. To cope with the high-dispersion, high-heterogeneity, and high-dimensionality of scRNA-seq data, a zero-inflated negative binomial (ZINB) module is incorporated into the network structure. Furthermore, the consistency between node representations and graph topological information is constrained to guide the joint learning process. It is noteworthy that scMFGNN can dynamically fuse multi-scale representations from multiple layers and meanwhile adaptively combine node representations and topological structural information from the same layer for representation learning and clustering. Experiments on multiple scRNA-seq datasets demonstrate the superiority of scMFGNN over the state-of-the-art. Code available: https://github.com/youngcmm/scMFGNN .}, archive = {J_NEUCOM}, author = {Chen-Min Yang and Dong Huang and Yuan-Kun Xu and Xiuting He and Guang-Yu Zhang and Chang-Dong Wang}, doi = {10.1016/j.neucom.2025.129764}, journal = {Neurocomputing}, month = {5}, pages = {129764}, shortjournal = {Neurocomputing}, title = {Towards multi-fusion graph neural network for single-cell RNA sequence clustering}, volume = {631}, year = {2025}, }
• (2025). Deep attributed network representation learning via enhanced local attribute neighbor. NEUCOM, 631, 129763. (www)

  @article{ , abstract = {Network representation learning aims to transform nodes in a network into low-dimensional spatial vectors while preserving the topological structure information of the network and its fundamental attributes, which has a wide range of practical applications. However, most existing attributed network representation learning methods only preserve part of the attributes and local or global topology information of the network, and do not fully capture the full attribute information of the complex interactions and the full topology information of the deep potential in the network. In the process of learning node embedding, it is a difficult and challenging task to fully and comprehensively capture and fuse the attribute and topology information in the network. To this end, we propose a new attributed network representation learning framework via enhanced local attribute neighbor, aiming to more effectively capture the global and local attribute information as well as the full topology information more comprehensively from the entire network. Specifically, a global attribute autoencoder is designed to model the mutual influence relationship of long-distance node attribute information, capture the global attribute neighbors of nodes from the whole network, and get the global attribute information of the complex interactions in the network. Additionally, a new random walk guide index, i.e., comprehensive influence, is designed to efficiently obtain the potential local and global topological structure information in the network. While at the same time, an enhanced local attribute neighbor skip-gram model is designed to obtain the local attribute information of nodes, so as to achieve the purpose of obtaining the network information in a full-aspect and multi-dimensional manner. We conduct extensive experiments on five real-world datasets for three downstream network analysis tasks: node classification, link prediction, and node clustering. The experimental results show that the method can achieve superior performance on each network analysis task, with the highest improvement of 3.94 % and 4.19 % in Micro-F1 and Macro-F1 over the optimal baseline method in node classification, respectively; 5.86 % and 5.2 % in Accuracy and Area Under Curve (AUC) in link prediction, respectively; and 7.73 %, 9.86 %, and 14.41 % in Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), and Completeness (Comp) in clustering, respectively, which proves the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Lili Han and Hui Zhao}, doi = {10.1016/j.neucom.2025.129763}, journal = {Neurocomputing}, month = {5}, pages = {129763}, shortjournal = {Neurocomputing}, title = {Deep attributed network representation learning via enhanced local attribute neighbor}, volume = {631}, year = {2025}, }
• (2025). Fixed-time stabilization and synchronization of fuzzy inertial neural networks via aperiodically semi-intermittent control. NEUCOM, 631, 129720. (www)

  @article{ , abstract = {This research focuses fixed-time stabilization (FTST) and fixed-time synchronization (FTS) within fuzzy inertial neural networks (FINNs) that features discontinuous activation functions (DAF) as well as mixed delays. First, we embrace a more general model, which with fuzzy items and state-switched connection weights. Then, we extend the result of FTST to FTS. Next, by creating aperiodically semi-intermittent controller (ASIC) and utilizing the principles of non-smooth theory, we derive the stability criterions for FTST. Lastly, in the simulation portion, we demonstrate the effectiveness and practical application of our findings through two concrete examples.}, archive = {J_NEUCOM}, author = {Junhao Hu and Guodong Zhang and Yan Li and Tianqi Wang}, doi = {10.1016/j.neucom.2025.129720}, journal = {Neurocomputing}, month = {5}, pages = {129720}, shortjournal = {Neurocomputing}, title = {Fixed-time stabilization and synchronization of fuzzy inertial neural networks via aperiodically semi-intermittent control}, volume = {631}, year = {2025}, }
• (2025). MambaSOD: Dual mamba-driven cross-modal fusion network for RGB-D salient object detection. NEUCOM, 631, 129718. (www)

  @article{ , abstract = {The purpose of RGB-D Salient Object Detection (SOD) is to pinpoint the most visually conspicuous areas within images accurately. Numerous conventional models heavily rely on CNN and overlook the long-range contextual dependencies, subsequent transformer-based models have addressed the issue to some extent but introduce quadratic computational complexity. Moreover, incorporating spatial information from depth maps has been proven effective for this task and the primary challenge is how to effectively fuse the complementary information from RGB and depth. Recent advancements in Mamba, particularly its superior ability to perform long-range modeling within linear efficiency, have motivated our exploration of its potential in the RGB-D SOD task. In this paper, we propose a dual Mamba-driven cross-modal fusion network for RGB-D SOD, named MambaSOD, which effectively leverages Mamba’s long-range dependency modeling capability. Specifically, we employ a dual Mamba-driven feature extractor to process RGB and depth inputs to obtain features with global contextual information. Then, we design a cross-modal fusion Mamba to perform modality-specific feature enhancement and model the inter-modal correlation between the RGB and depth features. To the best of our knowledge, this work is an innovative attempt to explore the potential of the pure Mamba in the RGB-D SOD task, offering a novel perspective. Numerous experiments conducted on seven prevailing datasets demonstrate our method’s superiority over eighteen state-of-the-art RGB-D SOD models. The source code will be released at https://github.com/YueZhan721/MambaSOD .}, archive = {J_NEUCOM}, author = {Yue Zhan and Zhihong Zeng and Haijun Liu and Xiaoheng Tan and Yinli Tian}, doi = {10.1016/j.neucom.2025.129718}, journal = {Neurocomputing}, month = {5}, pages = {129718}, shortjournal = {Neurocomputing}, title = {MambaSOD: Dual mamba-driven cross-modal fusion network for RGB-D salient object detection}, volume = {631}, year = {2025}, }
• (2025). Revolutionizing graph classification generalization with an adaptive causality-enhanced framework. NEUCOM, 631, 129716. (www)

  @article{ , abstract = {Current graph classification tasks predominantly utilize subgraph-based and pooling-based Graph Neural Networks (GNNs), which exploit statistical correlations within graphs to attain state-of-the-art performance in graph representation learning. Nevertheless, these statistical correlations often exhibit significant variability outside the training distribution, resulting in suboptimal generalization in classification tasks. This limitation stems from the model’s reliance on statistical correlations without adequately distinguishing between causal effects and spurious factors. To enhance stable and robust generalization, particularly in the context of complex and large-scale graph data, we propose an innovative A daptive C ausal Module and Causality- E nhanced (ACE) framework. This framework is designed to impart superior generalization capabilities and interpretability to any GNNs employed in diverse graph learning tasks. Specifically, the ACE framework focuses on extracting and leveraging causal information from graph data. This framework comprises three core components: the adaptive causal module, controlled backdoor adjustment, the causality enhancement module. First, the adaptive causal module is utilized to extract both causal features and confounding factors from the graph data. Second, controlled backdoor adjustment is applied to weaken spurious correlations introduced by confounders. Finally, the causality enhancement module facilitates the model’s acquisition of crucial causal information without requiring additional training. In essence, the ACE framework addresses the generalization deficiencies of GNNs models from a causal perspective, obviating the need for retraining on out-of-distribution data. Comprehensive experiments conducted on both synthetic biased datasets and real-world datasets demonstrate that the ACE framework significantly bolsters GNNs generalization and interpretability under distributional shifts, outperforming existing state-of-the-art graph classification methodologies. The code for the ACE framework is accessible via the following link: https://github.com/12chen20/ACE .}, archive = {J_NEUCOM}, author = {Jiali Chen and TieFeng Ma and Liwen Xu}, doi = {10.1016/j.neucom.2025.129716}, journal = {Neurocomputing}, month = {5}, pages = {129716}, shortjournal = {Neurocomputing}, title = {Revolutionizing graph classification generalization with an adaptive causality-enhanced framework}, volume = {631}, year = {2025}, }
• (2025). Unsupervised domain-adaptive object detection: An efficient method based on UDA-DETR. NEUCOM, 631, 129711. (www)

  @article{ , abstract = {Object detection based on deep learning has a wide range of applications in everyday life. However, when a domain gap exists between the training data (source domain) and real-world data (target domain), the performance of many object detectors significantly deteriorates. To address this issue, numerous Unsupervised Domain Adaptation (UDA) detectors attempt to reduce domain discrepancies and align cross-domain features. While these methods have achieved some success, they often align global features in a class-agnostic manner, neglecting the differences in feature distributions across categories within each domain. Our approach emphasizes the importance of both global image features and local instance features across domains, as well as category-specific information. Specifically, we propose an Encoder Feature Alignment (EFA) module, which introduces domain queries to adversarially align encoder-generated features, enabling the detector to extract more domain-invariant features. Additionally, we design an Instance-level Feature Alignment (IFA) module that extracts class-specific central features from the decoder for category-aware cross-domain feature alignment. During each training iteration, local class features progressively converge to global class features, guided by contrastive learning and adversarial loss to achieve Global Feature Alignment (GFA). Our method achieves 46.0% mean accuracy (mAP) in the Weather Adaptation scenario. Compared to the baseline model, a 19.1% mAP gain is achieved (26.9% → 46.0%). Extensive experimental results show that our proposed model achieves excellent detection performance and strong generalization ability on multiple domain adaptation benchmark datasets.}, archive = {J_NEUCOM}, author = {Hanguang Xiao and Tingting Zhou and Shidong Xiong and Jinlan Li and Zhuhan Li and Xin Liu and Tianhao Deng}, doi = {10.1016/j.neucom.2025.129711}, journal = {Neurocomputing}, month = {5}, pages = {129711}, shortjournal = {Neurocomputing}, title = {Unsupervised domain-adaptive object detection: An efficient method based on UDA-DETR}, volume = {631}, year = {2025}, }
• (2025). Designing adaptive deep denoisers for plug-and-play FBS with stable iterations. NEUCOM, 631, 129710. (www)

  @article{ , abstract = {Plug-and-Play Forward–Backward Splitting (PnP-FBS) leverages one off-the-shelf denoiser for image reconstruction in an iterative fashion. With this paradigm, the same denoiser can be employed for various reconstruction tasks without requiring any modification. Due to the simplicity and effectiveness, PnP-FBS has attracted widespread attention in imaging community. As the iteration style in implementation, its stability analysis has become a fundamental topic. However, this issue is particularly challenging, especially for powerful deep convolutional neural network (DCNN) denoisers, since they usually lack closed-form expressions. Recently, opening a new path different from DCNNs, researchers have proposed a novel fixed deep denoiser by unfolding FBS. Encouragingly, for PnP-FBS with this denoiser, the outputs of iterations converge to a high-quality reconstruction, thereby ensuring the stability. Notably, it is well known that, benefiting from flexibility, replacing a fixed denoiser with its adaptive version generally yields superior performance in practice. Motivated by this viewpoint and the unfolding FBS strategy in existing work, in this paper, we propose a method for constructing adaptive deep denoisers. For the constructed denoisers, it is theoretically proven that, under a mild assumption, there exists an upper bound on the distance between the outputs of the PnP-FBS iterations and the ground-truth image. That is, the stability of PnP-FBS is guaranteed. To validate the theoretical results, we develop a specific instance of the denoisers following the proposed construction method and plug it into PnP-FBS for image reconstruction. Experimental results demonstrate the stability of the iterations. Additionally, as expected, with the developed adaptive instance, the PnP-FBS achieves more accurate reconstructions than the one with the fixed version.}, archive = {J_NEUCOM}, author = {Qingliang Guo and Zhonggui Sun}, doi = {10.1016/j.neucom.2025.129710}, journal = {Neurocomputing}, month = {5}, pages = {129710}, shortjournal = {Neurocomputing}, title = {Designing adaptive deep denoisers for plug-and-play FBS with stable iterations}, volume = {631}, year = {2025}, }
• (2025). Data-based discrete-time two-player zero-sum delayed game via policy iteration Q-learning method. NEUCOM, 631, 129709. (www)

  @article{ , abstract = {In the framework of Q-learning, the research in this paper investigates a data-based discrete-time two-player zero-sum delayed game system. Crucial issue of the delayed game system is to solve the discrete-time delayed game algebraic Riccati equation (DTDGARE). However, to solve this equation, it is typically necessary to have complete system dynamics information and to account for the impact of delay terms, which introduces certain challenges during the solution process. To address the aforementioned challenges, a sophisticated delay operator technique is employed to address the challenges posed by the delay terms. Subsequently, the Q-function incorporates delayed terms, which, unlike the classical Q-function from reinforcement learning, is specifically designed to circumvent the need for directly solving the DTDGARE. Consequently, we propose a novel data-based delayed policy iteration Q-learning (DPIQL) algorithm, designed to learn the optimal Q-function directly from real data, thereby overcoming the limitations associated with requiring complete knowledge of system dynamics. Furthermore, by constructing a Banach space, a rigorous and innovative mathematical proof leveraging Kantorovich’s theorem ensures the convergence of the newly proposed DPIQL algorithm. Finally, the effectiveness and feasibility of the developed algorithm are validated through numerical simulations based on a real drone flight dataset.}, archive = {J_NEUCOM}, author = {Zongyang Jiang and Haiying Zhang and Yu Xiao}, doi = {10.1016/j.neucom.2025.129709}, journal = {Neurocomputing}, month = {5}, pages = {129709}, shortjournal = {Neurocomputing}, title = {Data-based discrete-time two-player zero-sum delayed game via policy iteration Q-learning method}, volume = {631}, year = {2025}, }
• (2025). Biologically plausible unsupervised learning for self-organizing spiking neural networks with dendritic computation. NEUCOM, 631, 129707. (www)

  @article{ , abstract = {Spiking neural networks (SNNs) are event-driven networks that possess efficient information processing capabilities. Compared to artificial neural networks (ANNs), SNNs are closer to the behavior of the brain in terms of information processing. Improving the biological plausibility of SNNs can help to explore and understand the computational mechanisms of the brain more effectively. In this paper, we propose a self-organizing SNN model with dendritic computation based on biological plausibility. Considering the impact of dendrites on neuronal and circuit functions, we proposed a multi-compartment neuron model with dendrites, which can enrich the dynamics and representation capabilities of neurons. Inspired by the organization of the brain network, we propose an adaptive self-organizing inhibition strategy that can cluster neurons encoding similar features together to help the network learn richer organizational relationships. In addition, we propose a partially shared connection method, which can effectively reduce the number of neurons in the network and significantly accelerate the training speed of the network. The experimental results on the MNIST, FashionMNIST, N-MNIST, and DVSGesture datasets show that our biologically plausible model has excellent potential to explore and understand brain mechanisms.}, archive = {J_NEUCOM}, author = {Geng Zhang and Shuangming Yang and Xuetao Zhang and Badong Chen}, doi = {10.1016/j.neucom.2025.129707}, journal = {Neurocomputing}, month = {5}, pages = {129707}, shortjournal = {Neurocomputing}, title = {Biologically plausible unsupervised learning for self-organizing spiking neural networks with dendritic computation}, volume = {631}, year = {2025}, }
• (2025). Neural network optimal adaptive control for affine continuous-time nonlinear systems with unknown internal dynamics. NEUCOM, 631, 129685. (www)

  @article{ , abstract = {A neural network (NN) optimal adaptive control for affine nonlinear continuous-time system with unknown internal dynamics is presented. The optimal adaptive control is given by a nonlinear differential equation with variable structure. A block-diagram of the controlled system is designed and investigated. Hardware and software implementation possibilities of the network are discussed. The NN does not need any learning and has moderate complexity. The trajectories of optimal adaptive control and system state variable are globally stable and convergent to unique steady states. It is shown that these trajectories are convergent to the steady states in finite time. Sliding modes of the trajectories are investigated. An accuracy of the network operation in the presence of disturbances of its nonlinearities is analyzed. Using the network for partial case of optimal tracking control is investigated. Computer simulations of the network operation that confirm theoretical derivations and illustrate high performance of the network showing its practical applications are provided.}, archive = {J_NEUCOM}, author = {Pavlo V. Tymoshchuk}, doi = {10.1016/j.neucom.2025.129685}, journal = {Neurocomputing}, month = {5}, pages = {129685}, shortjournal = {Neurocomputing}, title = {Neural network optimal adaptive control for affine continuous-time nonlinear systems with unknown internal dynamics}, volume = {631}, year = {2025}, }
• (2025). Advanced deep learning model for direct phase-only hologram generation using complex-valued neural networks. NEUCOM, 631, 129672. (www)

  @article{ , abstract = {This study proposes a novel deep learning-based approach for generating high-quality computer-generated holography (CGH). Data-centric deep learning models, in such cases, often face challenges due to limited training datasets and high computational complexity. The primary objective is to develop a lightweight, model-driven framework that generates CGH directly from input scenes using a complex-valued deep neural network architecture. The proposed architecture integrates residual blocks and employs a convolutional neural network (CNN) with rectified linear units (ReLU) and exponential linear units (ELU) as activation functions. This combination enhances the model's fitting capability through a hybrid structure of CNNs and dual network blocks. By utilizing complex-valued CNNs, the model more accurately processes magnitude and phase information, improving performance. Residual blocks optimize feature extraction and minimize data loss, enabling the generation of highly precise holograms. Furthermore, the angular-spectrum method (ASM) is employed to simulate wave propagation, allowing the network to accurately learn phase information from the input data in an end-to-end process. Experimental validation using the high-resolution DIV2K dataset demonstrates that the proposed method outperforms previous techniques. The structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) averages exceed 0.9055 and 32 dB, respectively. In addition to these two metrics, our proposed method was also evaluated using a no-reference image quality metric, the naturalness image quality evaluator (NIQE), and by analyzing visual difference maps. Notably, the proposed method achieves more than a 7 % improvement in PSNR compared to existing approaches.}, archive = {J_NEUCOM}, author = {Shariar Md Imtiaz and Tuvshinjargal Amgalan and F.M. Fahmid Hossain and Rupali Kiran Shinde and Hoonjong Kang and Ganbat Baasantseren and Ki-Chul Kwon and Nam Kim}, doi = {10.1016/j.neucom.2025.129672}, journal = {Neurocomputing}, month = {5}, pages = {129672}, shortjournal = {Neurocomputing}, title = {Advanced deep learning model for direct phase-only hologram generation using complex-valued neural networks}, volume = {631}, year = {2025}, }
• (2025). Non-parametric identifier of systems with uncertain model using a lyapunov-based continuous form of back-propagation method. NEUCOM, 631, 129670. (www)

  @article{ , abstract = {Non-parametric identifiers are dynamic systems that produce approximate models based on input–output paired data. This work develops a non-parametric identification algorithm for continuous systems based on the linear combination of functions that form a closed subset of a basis in a functional space. The algorithm considers the design of a differential form for the parameter adjustment, which is obtained using the second stability method of Lyapunov. Inspired by the gradient descent method, the proposed algorithm formally adjusts parameters despite the nonlinear relationships between parameters that define the identifier structure. Formal stability studies prove the explicit construction of parameter adjustment laws, highlighting the design conditions. This study shows the explicit design of an identifier with one single layer of activation functions in the basis, a single nested layer of such functions (like in a two-layer neural network), and a general form based on N-nested arrangements of functions like deep learning structures. This algorithm’s approximation capacity is confirmed by comparing it with some methods based on Autoregressive and Hammerstein-Wiener Models, available in MATLAB®software. The comparison of the proposed approach gets at least similar results to the best obtained with the comparative strategies. The proposed technique introduces a general form of deep learning tools to identify the evolution of dynamic systems with uncertain mathematical models.}, archive = {J_NEUCOM}, author = {Dusthon Llorente-Vidrio and Isaac Chairez and Rita Q. Fuentes-Aguilar}, doi = {10.1016/j.neucom.2025.129670}, journal = {Neurocomputing}, month = {5}, pages = {129670}, shortjournal = {Neurocomputing}, title = {Non-parametric identifier of systems with uncertain model using a lyapunov-based continuous form of back-propagation method}, volume = {631}, year = {2025}, }
• (2025). ELAPSE: Expand latent action projection space for policy optimization in offline reinforcement learning. NEUCOM, 631, 129665. (www)

  @article{ , abstract = {Offline Reinforcement Learning (RL) is dedicated to optimizing policy based on static datasets. Most offline RL algorithms optimize the policy to be close to behavior policy to avoid extrapolation errors caused by out-of-distribution actions. Therefore, Conditional Variational AutoEncoders (CVAE), which enable the approximation of the underlying behavior policy distribution based on the dataset, have been integrated into recent offline RL methods. However, there is an inherent issue of posterior collapse in the CVAE model. In this paper, we first define the latent action projection space collapse issue in offline RL. Then, we theoretically explain how the issue of collapse can lead to the deterioration of policy optimization and experimentally visualize this phenomenon. Therefore, a novel and easy-to-implement algorithm is proposed, which expands the latent action projection space through Batch Normalization. We call this algorithm Expand Latent Action Projection SpacE (ELAPSE), which balances the trade-off between the CVAE constraint and policy optimization, and avoids unnecessary perturbations that were present in previous works. Experiments prove that ELAPSE can effectively improve offline RL performance and demonstrate state-of-the-art results on the D4RL benchmarks. Our implementation is available on https://github.com/Mr-XcHan/ELAPSE .}, archive = {J_NEUCOM}, author = {Xinchen Han and Hossam Afifi and Michel Marot}, doi = {10.1016/j.neucom.2025.129665}, journal = {Neurocomputing}, month = {5}, pages = {129665}, shortjournal = {Neurocomputing}, title = {ELAPSE: Expand latent action projection space for policy optimization in offline reinforcement learning}, volume = {631}, year = {2025}, }
• (2025). Multi-modal cognitive maps for language and vision based on neural successor representations. NEUCOM, 631, 129662. (www)

  @article{ , abstract = {Cognitive maps are a proposed concept on how the brain efficiently organizes memories and retrieves context out of them. The entorhinal-hippocampal complex is heavily involved in episodic and relational memory processing, as well as spatial navigation and is thought to built cognitive maps via place and grid cells. To make use of the promising properties of cognitive maps, we set up a multi-modal neural network using successor representations which are able to model place cell dynamics and cognitive map representations. Here, we use multi-modal inputs consisting of images and word embeddings. The network learns the similarities between novel inputs and the training database and therefore the representation of the cognitive map successfully. Subsequently, the prediction of the network can be used to infer from one modality to another with over 90% accuracy. The proposed method could therefore be a building block to improve current AI systems for better understanding of the environment and the different modalities in which objects appear. The association of specific modalities with certain encounters can therefore lead to context awareness in novel situations when similar encounters with less information occur and additional information can be inferred from the learned cognitive map. Cognitive maps, as represented by the entorhinal-hippocampal complex in the brain, organize and retrieve context from memories, suggesting that large language models (LLMs) like ChatGPT could harness similar architectures to function as a high-level processing center, akin to how the hippocampus operates within the cortex hierarchy. Finally, by utilizing multi-modal inputs, LLMs can potentially bridge the gap between different forms of data (like images and words), paving the way for context-awareness and grounding of abstract concepts through learned associations, addressing the symbol grounding problem in AI.}, archive = {J_NEUCOM}, author = {Paul Stoewer and Achim Schilling and Pegah Ramezani and Hassane Kissane and Andreas Maier and Patrick Krauss}, doi = {10.1016/j.neucom.2025.129662}, journal = {Neurocomputing}, month = {5}, pages = {129662}, shortjournal = {Neurocomputing}, title = {Multi-modal cognitive maps for language and vision based on neural successor representations}, volume = {631}, year = {2025}, }
• (2025). NN-based adaptive event-triggered predefined time control of flexible joint robot with full-state error constraints. NEUCOM, 631, 129658. (www)

  @article{ , abstract = {For flexible joint robot (FJR) system with full-state error constraints, this paper proposes an event-triggered adaptive predefined time control method. This study employs a radial basis function neural networks (RBFNN) to handle unknown nonlinear dynamic functions. In addition, it makes use of the universal barrier Lyapunov function to ensure that both the full-state errors and tracking errors converge to a user-specified range of control accuracy. Based on the event-triggered control strategy, a communication limitation control scheme is established to reduce the execution time of the controller and save communication resources. Combined with the DSC technique, the differential explosion problem is avoided by utilizing the constructed predefined time filters. By using the proposed predefined time stability criterion and Lyapunov function, it has been demonstrated that the controlled system achieves practical predefined time stable (PPTS) and all control signals bounded within the predefined time interval. Simulation results are utilized to verify the predefined time control performance.}, archive = {J_NEUCOM}, author = {Hongyu Guan and Shuai Sui and Yuchao Sui and C.L. Philip Chen}, doi = {10.1016/j.neucom.2025.129658}, journal = {Neurocomputing}, month = {5}, pages = {129658}, shortjournal = {Neurocomputing}, title = {NN-based adaptive event-triggered predefined time control of flexible joint robot with full-state error constraints}, volume = {631}, year = {2025}, }
• (2025). FCN+: Global receptive convolution makes FCN great again. NEUCOM, 631, 129655. (www)

  @article{ , abstract = {Fully convolutional network (FCN) is a seminal work for semantic segmentation. However, due to its limited receptive field, FCN cannot effectively capture global context information which is vital for semantic segmentation. As a result, it is beaten by state-of-the-art methods which leverage different filter sizes for larger receptive fields. In this paper, we propose a novel global receptive convolution (GRC) to effectively increase the receptive field of FCN for context information extraction, which results in an improved FCN termed FCN+. The GRC provides global receptive field for convolution without introducing any extra learnable parameters . Specifically, the GRC first divides the channels of the filter into two groups. The grid sampling locations of the first group are shifted to different spatial coordinates across the whole feature map, according to their channel indexes. This can help the convolutional filter capture the global context information. The grid sampling location of the second group remains unchanged to keep the original location information. Convolving using these two groups, the GRC can integrate the global context into the original location information of each pixel for better dense prediction results. With the GRC built in, FCN+ can achieve comparable performance to state-of-the-art methods for semantic segmentation tasks, as verified on PASCAL VOC 2012, Cityscapes and ADE20K. Our code will be released at https://github.com/Zhongying-Deng/FCN_Plus .}, archive = {J_NEUCOM}, author = {Xiaoyu Ren and Zhongying Deng and Jin Ye and Junjun He and Dongxu Yang}, doi = {10.1016/j.neucom.2025.129655}, journal = {Neurocomputing}, month = {5}, pages = {129655}, shortjournal = {Neurocomputing}, title = {FCN+: Global receptive convolution makes FCN great again}, volume = {631}, year = {2025}, }
• (2025). Morphology generalizable reinforcement learning via multi-level graph features. NEUCOM, 631, 129644. (www)

  @article{ , abstract = {Controlling a group of robots with diverse morphologies using a unified policy, known as morphology generalizable control, is a challenging problem in robotic control. Existing graph neural network-based (GNN-based) methods suffer from inefficient modular communication due to non-adjacent modules having to communicate across multiple hops, while transformer-based methods neglect morphology prior information which is crucial for morphology generalizable control. To overcome these limitations, in this work, we propose MG 2 ( M orphology G eneralizable Reinforcement Learning via M ulti-level G raph Features) which incorporates multi-level graph features derived from the morphology graph into the transformer architecture. To effectively incorporate morphology information while achieving efficient modular communication, MG 2 introduces graph features three-levels, local, global, and relative graph features, and incorporates them into the transformer architecture. By introducing morphology prior information, MG 2 improves multi-task training and generalization performance in morphology-generalizable reinforcement learning. The performance enhancements are evaluated primarily on the SMP benchmark and consolidated on several UNIMAL robots.}, archive = {J_NEUCOM}, author = {Yansong Pan and Rui Zhang and Jiaming Guo and Shaohui Peng and Fan Wu and Kaizhao Yuan and Yunkai Gao and Siming Lan and Ruizhi Chen and Ling Li and Xing Hu and Zidong Du and Zihao Zhang and Xin Zhang and Wei Li and Qi Guo and Yunji Chen}, doi = {10.1016/j.neucom.2025.129644}, journal = {Neurocomputing}, month = {5}, pages = {129644}, shortjournal = {Neurocomputing}, title = {Morphology generalizable reinforcement learning via multi-level graph features}, volume = {631}, year = {2025}, }
• (2025). Quadratic graph attention network (Q-GAT) for robust construction of gene regulatory network. NEUCOM, 631, 129635. (www)

  @article{ , abstract = {Gene regulatory relationships can be abstracted as a gene regulatory network (GRN), which plays a key role in characterizing complex cellular processes and pathways. Recently, graph neural networks (GNNs), as a class of deep learning models, have emerged as a useful tool to infer gene regulatory relationships from gene expression data. However, deep learning models have been found to be vulnerable to noise, which greatly hinders the adoption of deep learning in constructing GRNs, because high noise is often unavoidable in the process of gene expression measurement. Is it possible to develop a robust GNN for constructing GRNs? In this paper, we give a positive answer by proposing a Quadratic Graph Attention Network (Q-GAT) with a dual attention mechanism. We study the changes in the predictive accuracy of Q-GAT and 9 state-of-the-art baselines by introducing different levels of adversarial perturbations. Experiments in the E. coli and S. cerevisiae datasets suggest that Q-GAT outperforms the state-of-the-art models in robustness. Lastly, we dissect why Q-GAT is robust through the signal-to-noise ratio (SNR), interpretability analyses and explainability evaluation. The former informs that nonlinear aggregation of quadratic neurons can amplify useful signals and suppress unwanted noise, thereby facilitating robustness, while the latter reveals that Q-GAT can leverage more features in prediction thanks to the dual attention mechanism, which endows Q-GAT with the ability to confront adversarial perturbation. We have shared our code in https://github.com/Minorway/Q-GAT_for_Robust_Construction_of_GRN for readers’ evaluation.}, archive = {J_NEUCOM}, author = {Hui Zhang and Xuexin An and Qiang He and Yudong Yao and Yudong Zhang and Feng-Lei Fan and Yueyang Teng}, doi = {10.1016/j.neucom.2025.129635}, journal = {Neurocomputing}, month = {5}, pages = {129635}, shortjournal = {Neurocomputing}, title = {Quadratic graph attention network (Q-GAT) for robust construction of gene regulatory network}, volume = {631}, year = {2025}, }
• (2025). Causality-driven sequence segmentation assisted soft sensing for multiphase industrial processes. NEUCOM, 631, 129612. (www)

  @article{ , abstract = {The dynamic characteristics of multiphase industrial processes present significant challenges in the field of industrial big data modeling. Traditional soft sensing models frequently neglect the process dynamics and have difficulty in capturing transient phenomena like phase transitions. To address this issue, this paper introduces a causality-driven sequence segmentation (CDSS) method. The CDSS method segments the industrial time series into difference phases, based on the causal mechanism shifts. A novel distance-based dual similarity metric is designed to evaluate the temporal consistency of causal mechanisms, including both the causal similarity and the stable similarity. Furthermore, a soft sensing model, called phase-specific temporal-causal graph convolutional network (PS-TC-GCN), is established for each phase, by using the segmented phase data and the discovered temporal causal graphs (TCGs). The numerical examples are utilized to validate the proposed CDSS method, and the segmentation results demonstrate that the CDSS has excellent performance on segmenting both the stable and the unstable multiphase series. Especially, the CDSS has higher accuracy in separating non-stationary time series compared to other methods. The effectiveness of the proposed CDSS method and the PS-TC-GCN model is also verified through the penicillin fermentation process. The experimental results indicate that the breakpoints identified by the CDSS are accurately aligned with the ground truth. Furthermore, the PS-TC-GCN achieves higher predictive accuracy, compared with other soft sensing models.}, archive = {J_NEUCOM}, author = {Yimeng He and Xinmin Zhang and Xiangyin Kong and Le Yao and Zhihuan Song}, doi = {10.1016/j.neucom.2025.129612}, journal = {Neurocomputing}, month = {5}, pages = {129612}, shortjournal = {Neurocomputing}, title = {Causality-driven sequence segmentation assisted soft sensing for multiphase industrial processes}, volume = {631}, year = {2025}, }
• (2025). Efficiency-safety coordination optimization in drilling process under complex formations. NEUCOM, 630, 129732. (www)

  @article{ , abstract = {Efficient and safe drilling process under complex formation conditions has significant economic value for drilling engineering. Rate of penetration (ROP) and drill-string vibration are two widely used objectives for optimizing the drilling efficiency and safety, yet a simultaneous optimization of both is still a challenge due to their strong coupling relationship. To address this issue, an artificial intelligence-based coordinated optimization strategy is proposed for real-time dual-objective optimization of the two targets. First, by exploiting the multivariate distribution characteristics of drilling parameters, a Gaussian mixture model is applied for formation identification. Then, an information entropy-based coupling coordination degree (CCD) model is proposed to quantify the coordination relationship, and constrain optimization objectives. Finally, a dual-objective optimization model with ROP and vibration, served as the optimization objectives, is constructed, and adaptive model constraints are designed for different formations. A real-time optimization strategy combining non-dominated sorting genetic algorithm II and sliding window technique is developed to determine optimal operational parameters, and a CCD-based decision criterion is developed for the Pareto solution set. Experiments on industrial data from an actual drilling field indicate that the proposed strategy leads to significant gains in drilling performances, providing a practical solution for drilling engineering.}, archive = {J_NEUCOM}, author = {Xiao Wu and Xuzhi Lai and Jie Hu and Chengda Lu and Xiao Yang and Yang Zhou and Min Wu}, doi = {10.1016/j.neucom.2025.129732}, journal = {Neurocomputing}, month = {5}, pages = {129732}, shortjournal = {Neurocomputing}, title = {Efficiency-safety coordination optimization in drilling process under complex formations}, volume = {630}, year = {2025}, }
• (2025). Medical imaging-based artificial intelligence in pneumonia: A narrative review. NEUCOM, 630, 129731. (www)

  @article{ , abstract = {Pneumonia is a pulmonary inflammation caused by various pathogens and remains one of the leading causes of global mortality. In clinic, chest X-ray (CXR), computed tomography (CT), and lung ultrasound (LUS) have been the most commonly used imaging techniques for patients with pneumonia. Over the past decade, advances in artificial intelligence (AI) have sparked increasing interests in computer-aided diagnosis (CAD) system development aimed at improving clinical workflow efficiency and promoting a more comprehensive understanding of medical imaging information. The deep learning-based image analysis for pneumonia flourished during the pandemic of coronavirus disease 2019 (COVID-19), focusing on clinical problems include the development of rapid screening system for COVID-19 pneumonia, differential diagnosis, severity assessment, and predicting prognosis. In this narrative review, we firstly establish the general characteristics of imaging techniques used for pneumonia and outlined the typical manifestations used by clinicians in diagnosis and assessment. Then we provide a state-of-the-art review on the emerging applications in developing AI-based tools for pneumonia lesion detection and segmentation and auxiliary diagnosis. Efforts by researchers in combining multimodal imaging data and investigating correlations between different imaging modalities are also presented. Finally, we discussed the challenges in acquiring high-quality data, utilizing multimodal data, and promoting clinical adoption that must be overcome.}, archive = {J_NEUCOM}, author = {Yanping Yang and Wenyu Xing and Yiwen Liu and Yifang Li and Dean Ta and Yuanlin Song and Dongni Hou}, doi = {10.1016/j.neucom.2025.129731}, journal = {Neurocomputing}, month = {5}, pages = {129731}, shortjournal = {Neurocomputing}, title = {Medical imaging-based artificial intelligence in pneumonia: A narrative review}, volume = {630}, year = {2025}, }
• (2025). Dual control for autonomous airborne source search with nesterov accelerated gradient descent: Algorithm and performance analysis. NEUCOM, 630, 129729. (www)

  @article{ , abstract = {Dual Control for Exploitation and Exploration (DCEE) shows promising performance by realizing optimal trade-off between exploitation and exploration under an unknown environment. However, it is computationally intensive and lacks rigorously established properties such as stability and convergence. This paper addresses these two issues by developing the Nesterov Accelerated Gradient Descent (NAGD) based DCEE, i.e. DCEE-NAGD, where the NAGD is applied to both the source term estimation and the path planning in the DCEE framework. It shows that DCEE-NAGD significantly reduces the search time by driving the search agent moving towards the estimated airborne source location (exploitation) and actively searching new data to reduce the current estimation uncertainty (exploration) with the help of NAGD. The convergence of both the source term estimation and the path planning of the DCEE-NAGD algorithm is rigorously established by applying the mean value theorem and mathematical transformation. More specifically, the convergence boundaries and the convergence rates of the source term estimation and the whole DCEE-NAGD algorithm are rigorously established. Both theoretic analysis and simulations confirm the proposed DCEE-NAGD algorithm significantly improves the performance so reduces the autonomous search time.}, archive = {J_NEUCOM}, author = {Guoqiang Tan and Wen-Hua Chen and Jun Yang and Xuan-Toa Tran and Zhongguo Li}, doi = {10.1016/j.neucom.2025.129729}, journal = {Neurocomputing}, month = {5}, pages = {129729}, shortjournal = {Neurocomputing}, title = {Dual control for autonomous airborne source search with nesterov accelerated gradient descent: Algorithm and performance analysis}, volume = {630}, year = {2025}, }
• (2025). A lie group laplacian support vector machine for semi-supervised learning. NEUCOM, 630, 129728. (www)

  @article{ , abstract = {As a semi-supervised learning (SSL) method, the Laplacian Support Vector Machine (LapSVM) utilizes both labeled and unlabeled data to form a manifold regularization with graph theory and often receives a better performance than other SVM methods when the labeled data samples are insufficient. However, current LapSVMs usually struggle to handle high-dimensional data with transformation invariance such as images and videos, ignore the neighboring and distribution information of labeled data samples, and are sensitive to data samples around the decision boundary. Focusing on these problems, we propose a novel SVM method for SSL called LG-LapSVM by incorporating Lie group theory, the theory of local behavior similarity, and the RoBoSS loss function into the existing LapSVM framework. LG-LapSVM first represents data samples as transformation-invariant features on a Lie group manifold where a Laplacian graph is constructed according to geodesic distances between group elements, neighboring information, and class distributions of labeled data samples; then forms the Lie group manifold regulation by using the Laplacian matrix generated from the Laplacian graph and Lie group kernel metrics between each pair of data samples; finally trains an SVM classifier with an objective function including the RoBoSS loss function for labeled data samples and the Lie group manifold regulation. Comprehensive experiments on three public datasets validate the effectiveness of LG-LapSVM and the superiority of the method compared to nine typical SVM methods for SSL.}, archive = {J_NEUCOM}, author = {Yue Zhang and Li Liu and Qian Qiao and Fanzhang Li}, doi = {10.1016/j.neucom.2025.129728}, journal = {Neurocomputing}, month = {5}, pages = {129728}, shortjournal = {Neurocomputing}, title = {A lie group laplacian support vector machine for semi-supervised learning}, volume = {630}, year = {2025}, }
• (2025). Neural augmentation based panoramic high dynamic range stitching. NEUCOM, 630, 129726. (www)

  @article{ , abstract = {Due to saturated regions of inputting low dynamic range (LDR) images and large intensity changes among the LDR images caused by different exposures, it is challenging to produce an information enriched panoramic LDR image without visual artifacts for a high dynamic range (HDR) scene through stitching multiple geometrically synchronized LDR images with different exposures and pairwise overlapping fields of views (OFOVs). Fortunately, the stitching of such images is innately a perfect scenario for the fusion of a physics-driven approach and a data-driven approach due to their OFOVs. Based on this new insight, a novel neural augmentation based panoramic HDR stitching algorithm is proposed in this paper. The physics-driven approach is built up using the OFOVs. Different exposed images of each view are initially generated by using the physics-driven approach, are then refined by a data-driven approach, and are finally used to produce panoramic LDR images with different exposures. All the panoramic LDR images with different exposures are combined together via a multi-scale exposure fusion algorithm to produce the final panoramic LDR image. Experimental results demonstrate the proposed algorithm outperforms existing panoramic stitching algorithms.}, archive = {J_NEUCOM}, author = {Chaobing Zheng and Yilun Xu and Weihai Chen and Shiqian Wu and Sen Zhang and Zhengguo Li}, doi = {10.1016/j.neucom.2025.129726}, journal = {Neurocomputing}, month = {5}, pages = {129726}, shortjournal = {Neurocomputing}, title = {Neural augmentation based panoramic high dynamic range stitching}, volume = {630}, year = {2025}, }
• (2025). Intermittent sampled-data synchronization of delayed reaction–diffusion neural networks. NEUCOM, 630, 129724. (www)

  @article{ , abstract = {This article investigates the synchronization of delayed reaction–diffusion neural networks (RDNNs) based on intermittent sampled-data (SD) control under spatially local averaged measurements (SLAMs), where distributed and discrete delays are introduced. In order to save the control cost and reduce communications burden, we introduce intermittent SD control based on SLAMs to design the synchronization scheme, which requires only a limited number of actuators and sensors. Meanwhile, we construct the interval-dependent Lyapunov functional (LF) for the synchronization error system, which is continuous during the sampling period. Then, according to the designed LF and several inequality techniques, some criteria for intermittent SD synchronization are established to ensure that the error system is exponentially stable. In addition, we transform the controller design issue into a feasibility problem of liner matrix inequalities (LMIs). In the end, two numeric examples are offered to ensure the feasibility and effectiveness of the method.}, archive = {J_NEUCOM}, author = {Hong-Yu Chen and Zi-Peng Wang and Junfei Qiao and Jin-Liang Wang and Huai-Ning Wu and Tingwen Huang}, doi = {10.1016/j.neucom.2025.129724}, journal = {Neurocomputing}, month = {5}, pages = {129724}, shortjournal = {Neurocomputing}, title = {Intermittent sampled-data synchronization of delayed reaction–diffusion neural networks}, volume = {630}, year = {2025}, }
• (2025). ComNC: A unified framework for trends prediction integrating node and concept effects. NEUCOM, 630, 129721. (www)

  @article{ , abstract = {Recent studies have highlighted the potential of graph neural networks (GNNs) in embedding relational interactions within time-varying systems. These relationships predominantly encompass node-based pairwise interactions, concept-based synergistic relationships arising from common external shocks affecting multiple nodes, and latent, unobservable connections. While the effectiveness of embedding these relationships individually has been widely acknowledged, few studies have successfully integrated them into a unified framework. The primary challenge stems from the inherent ambiguity of these relationships and the dynamic nature of their interactions. In this study, we introduce ComNC, a novel framework designed to predict the trends of a time-varying system by unifying these three types of relationships. Specifically, ComNC estimates implicit node-based and concept-based relationships from observable time-series signals. It then captures the interactions between these diverse relationships through a cross-graph attention mechanism. The model is fed with input signals along with the three distinct relationships and their estimated dynamic interactions, which are subsequently fused by a unified feature fusion module to make final predictions. Experimental evaluations conducted on three real-world datasets demonstrate the superiority of ComNC over state-of-the-art methods, including AD-GAT, VGNN, HGNN, and HGTAN. Our method achieves at least a 2.10% improvement in Directional Accuracy (DA) and a 3.09% increase in Area Under the Curve (AUC). These results underscore the framework’s effectiveness in capturing multi-type relationships and its potential to advance predictive modeling in complex networks.}, archive = {J_NEUCOM}, author = {Sanchuan Xiao and Qing Li and Xiaoyue Gong and Jingmei Zhao and Lingyun Gu and Long Peng}, doi = {10.1016/j.neucom.2025.129721}, journal = {Neurocomputing}, month = {5}, pages = {129721}, shortjournal = {Neurocomputing}, title = {ComNC: A unified framework for trends prediction integrating node and concept effects}, volume = {630}, year = {2025}, }
• (2025). 3D path planning for AUVs under ocean currents by prioritized experience replay mechanism. NEUCOM, 630, 129719. (www)

  @article{ , abstract = {In this article, the 3D path planning issue for autonomous underwater vehicles (AUVs) with real ocean currents (OCs) is investigated by using the prioritized experience replay (PER) mechanism based reinforcement learning (RL) approach. To effectively use the real OC information and successfully avoid the obstacles for path planning in 3D case, a comprehensive reward function is presented. For the purpose of addressing the issue of low sampling utilization due to random sampling, the prioritized experience replay (PER) mechanism is implemented to adjust the weighting of samples. By combining the soft actor–critic (SAC) algorithm with PER strategy, the path planning algorithm proposed in this article effectively achieves obstacle avoidance in the complex ocean environments. Subsequently, the simulation experiments are given to compare this method with some traditional RL approaches, and the experiments demonstrate that the presented algorithm shows more efficient.}, archive = {J_NEUCOM}, author = {Haoran Huang and Kai Song and Yun Chen and Haixiang Jin and Yacun Guan}, doi = {10.1016/j.neucom.2025.129719}, journal = {Neurocomputing}, month = {5}, pages = {129719}, shortjournal = {Neurocomputing}, title = {3D path planning for AUVs under ocean currents by prioritized experience replay mechanism}, volume = {630}, year = {2025}, }
• (2025). Spore: Spatio-temporal collaborative perception and representation space disentanglement for remote heart rate measurement. NEUCOM, 630, 129717. (www)

  @article{ , abstract = {Remote Photoplethysmography (rPPG) leverages standard RGB cameras for contactless heart rate monitoring, overcoming the limitations of traditional PPG technology in telemedicine and offering a highly scalable, cost-effective health monitoring solution. Despite the advancements of current deep learning methods, which utilize spatiotemporal convolutional networks to capture subtle rPPG signals, these approaches often fail to fully exploit local similarities and global quasi-periodicity in both spatial and temporal dimensions. Additionally, non-physiological noise remains prevalent in the representation space, impeding the accurate estimation of physiological parameters across diverse representation domains. To address these measurement challenges, we propose Spore, a novel training strategy that integrates a Spatio-Temporal Cooperative Perception Network (STCPNet) and a Separable Network (SpNet). Spore effectively disentangles noise and extracts physiological signals through differential orthogonal disentanglement and parallel approximation techniques, ensuring precise measurement of heart rate. STCPNet meticulously aggregates semantic context across spatial and temporal dimensions, enhancing global-level and trend cross-correlations in a fine-grained manner. Meanwhile, the resource-efficient SpNet identifies and constructs target representation spaces by realigning the distribution of the source latent space, thereby adaptively capturing disentangled physiological signal patterns from the computationally intensive STCPNet. For validation, extensive experiments were conducted not only on multiple benchmark datasets but also through deployment testing in real-world scenarios. The results demonstrate that our proposed training strategy achieves state-of-the-art performance in heart rate measurement while maintaining resource efficiency. The code will be released at https://github.com/zacheryzhang/spore .}, archive = {J_NEUCOM}, author = {Zexing Zhang and Huimin Lu and Zhihai He and Adil Al-Azzawi and Songzhe Ma and Chenglin Lin}, doi = {10.1016/j.neucom.2025.129717}, journal = {Neurocomputing}, month = {5}, pages = {129717}, shortjournal = {Neurocomputing}, title = {Spore: Spatio-temporal collaborative perception and representation space disentanglement for remote heart rate measurement}, volume = {630}, year = {2025}, }
• (2025). Spatial and channel level feature redundancy reduction for differentiable neural architecture search. NEUCOM, 630, 129713. (www)

  @article{ , abstract = {Differentiable architecture search (DARTS), based on the continuous relaxation of the architectural representation and gradient descent, achieves effective results in Neural Architecture Search (NAS) field. Among the neural architectures, convolutional neural networks (CNNs) have achieved remarkable performance in various computer vision tasks. However, convolutional layers inevitably extract redundant features as the limitation of the weight-sharing property by convolutional kernels, thus slowing down the search efficiency of DARTS. In this paper, we propose a novel search approach named Slim-DARTS from the perspective of reducing feature redundancy, to further achieve high-speed and efficient neural architecture search. At the level of spatial redundancy, we design a spatial reconstruction module to eliminate spatial feature redundancy and facilitate representative feature learning. At the channel redundancy level, partial channel connection is applied to randomly sample a small subset of channels for operation selection to reduce unfair competition among candidate operations. And we introduce a group of channel parameters to automatically adjust the proportion of selected channels. The experimental results show that our research greatly improves search efficiency and memory utilization, achieving classification error rates of 2.39% and 16.78% on CIFAR-10 and CIFAR-100, respectively.}, archive = {J_NEUCOM}, author = {Shantong Yin and Ben Niu and Rui Wang and Xin Wang}, doi = {10.1016/j.neucom.2025.129713}, journal = {Neurocomputing}, month = {5}, pages = {129713}, shortjournal = {Neurocomputing}, title = {Spatial and channel level feature redundancy reduction for differentiable neural architecture search}, volume = {630}, year = {2025}, }
• (2025). ScCDAN: Constraint domain adaptation network for cell type annotation across single cell RNA sequencing data. NEUCOM, 630, 129708. (www)

  @article{ , abstract = {The development of single cell sequencing technology enables researchers to investigate tissue heterogeneity at the cellular level, with cell type annotation being the core of its analysis. Although various computational methods have been proposed, batch effects remain a major issue affecting annotation accuracy. In this study, a constraint domain adaptation network, named scCDAN, is proposed to address batch effects that lead to poor cell type annotation. Firstly, a domain alignment module is used to align the distributions of source and target domain data through adversarial learning strategies to reduce domain discrepancies. Furthermore, to achieve finer granularity in differentiating cell types, a category boundary constraint module is designed to regulate the positional relationships between cells of the same type and those of different types in the feature space. scCDAN combines domain alignment and category boundary constraint modules to directly and indirectly address batch effects, thereby improving the accuracy of cell type annotation. The effectiveness of scCDAN is validated on simulated, cross-platforms, and cross-species datasets. Experimental results demonstrate that scCDAN outperforms comparative methods in both cell type annotation and batch effects correction.}, archive = {J_NEUCOM}, author = {Junliang Shang and Yan Zhao and Baojuan Qin and Limin Zhang and Xianghan Meng and Daohui Ge and Jin-Xing Liu}, doi = {10.1016/j.neucom.2025.129708}, journal = {Neurocomputing}, month = {5}, pages = {129708}, shortjournal = {Neurocomputing}, title = {ScCDAN: Constraint domain adaptation network for cell type annotation across single cell RNA sequencing data}, volume = {630}, year = {2025}, }
• (2025). AniArtAvatar: Animatable 3D art avatar from a single image. NEUCOM, 630, 129706. (www)

  @article{ , abstract = {We present a novel approach for generating animatable 3D-aware art avatars from a single image, with controllable facial expressions, head poses, and shoulder movements. Unlike previous reenactment methods, our approach utilizes a pre-trained view-conditioned 2D diffusion model to synthesize multi-view images from a single art portrait with a neutral expression. With the generated colors and normals, we synthesize a static avatar using an SDF-based neural surface. For avatar animation, we extract control points, transfer the motion with these points, and deform the implicit canonical space. Firstly, we render the front image of the avatar, extract the 2D landmarks, and project them to the 3D space using a trained SDF network. We extract 3D driving landmarks using 3DMM and transfer the motion to the avatar landmarks. We deform the SDF-based neural surface utilizing neutral and expressive 3D avatar landmarks to animate the expression. To animate the avatar pose, we manually set the body height and bound the head and torso of an avatar with two cages. The head and torso can be animated by transforming the two cages. Our approach is a one-shot pipeline that can be applied to various styles. Experiments demonstrate that our method can generate high-quality 3D art avatars with desired control over different motions.}, archive = {J_NEUCOM}, author = {Shaoxu Li and Ye Pan}, doi = {10.1016/j.neucom.2025.129706}, journal = {Neurocomputing}, month = {5}, pages = {129706}, shortjournal = {Neurocomputing}, title = {AniArtAvatar: Animatable 3D art avatar from a single image}, volume = {630}, year = {2025}, }
• (2025). A clustering algorithm based on dominating (dominated) roughness. NEUCOM, 630, 129705. (www)

  @article{ , abstract = {As a pivotal technique within data mining, clustering is integral to a variety of downstream applications. Clustering methods that employ rough sets are often prone to issues of system instability. However, the dominance-based rough set approach (DRSA) enhances semantic interpretation and generalization capabilities in data object description. This paper integrates the generalization potential of DRSA with rough set approximation concepts to devise a measure of object similarity. Utilizing a heuristic framework, strategies for both initial configuration and partitioning are formulated. From a novel viewpoint, this study introduces a DRSA-based clustering algorithm. Subsequent analyses on real datasets evaluate the accuracy and robustness of the proposed algorithm. Visual assessments underscore the effectiveness of this approach in managing complex spatial structures. Furthermore, experiments concerning fuzzy boundary disturbances, normalization variations, and noise resistance affirm the robustness of the method.}, archive = {J_NEUCOM}, author = {Bin Yu and Ruihui Xu and Weiping Ding}, doi = {10.1016/j.neucom.2025.129705}, journal = {Neurocomputing}, month = {5}, pages = {129705}, shortjournal = {Neurocomputing}, title = {A clustering algorithm based on dominating (dominated) roughness}, volume = {630}, year = {2025}, }
• (2025). Reinforcement learning-based secure training for adversarial defense in graph neural networks. NEUCOM, 630, 129704. (www)

  @article{ , abstract = {The security of Graph Neural Networks (GNNs) is crucial for ensuring the reliability and protection of the systems they are integrated within real-world applications. However, current approaches lack the ability to prevent GNNs from learning high-risk information, including edges, nodes, convolutions, etc. In this paper, we propose a secure GNN learning framework called Reinforcement Learning-based Secure Training Algorithm. We first introduce a model conversion technique that transforms the training process of GNNs into a verifiable Markov Decision Process model. To maintain the security of model we employ Deep Q-Learning algorithm to prevent high-risk information messages. Additionally, to verify whether the strategy derived from Deep Q-Learning algorithm meets safety requirements, we design a model transformation algorithm that converts MDPs into probabilistic verification models, thereby ensuring our method’s security through formal verification tools. The effectiveness and feasibility of our proposed method are demonstrated by achieving a 6.4% improvement in average accuracy on open-source datasets under adversarial attack graphs.}, archive = {J_NEUCOM}, author = {Dongdong An and Yi Yang and Xin Gao and Hongda Qi and Yang Yang and Xin Ye and Maozhen Li and Qin Zhao}, doi = {10.1016/j.neucom.2025.129704}, journal = {Neurocomputing}, month = {5}, pages = {129704}, shortjournal = {Neurocomputing}, title = {Reinforcement learning-based secure training for adversarial defense in graph neural networks}, volume = {630}, year = {2025}, }
• (2025). QUAV flight control based on axially symmetric DRL. NEUCOM, 630, 129703. (www)

  @article{ , abstract = {Deep reinforcement learning (DRL) has emerged as a prominent technique for advancing flight control systems. However, existing research often neglects the inherent symmetry present in quadrotor unmanned aerial vehicle (QUAV) dynamics, resulting in the drawback of instability and inefficiency. To tackle these problems, we propose an axially symmetric network to enhance flight control performance. To be specific, a converting module is proposed to fuse the vertical state and horizontal state to realize the stable and axially symmetric control performance. Furthermore, the proposed method exhibits generality and it could be validated using various DRL algorithms. Through a series of comparative experiments, we validate the superiority of the proposed controller, demonstrating notable improvements in both the efficiency and robustness of flight control operations.}, archive = {J_NEUCOM}, author = {Yirui Zhang and Haoran Han and Jian Cheng}, doi = {10.1016/j.neucom.2025.129703}, journal = {Neurocomputing}, month = {5}, pages = {129703}, shortjournal = {Neurocomputing}, title = {QUAV flight control based on axially symmetric DRL}, volume = {630}, year = {2025}, }
• (2025). Image–text aggregation for open-vocabulary semantic segmentation. NEUCOM, 630, 129702. (www)

  @article{ , abstract = {Existing works on open-vocabulary semantic segmentation explore utilizing large-scale vision-language models. Recent methods have relied mostly on visual features while treating text features as supporting components. Our method explores the potential of image-related text information. This paper proposes a novel open-vocabulary semantic segmentation method based on image–text aggregation (ITA). We design a dominant category unearthing module to mine text features strongly correlated with the image, facilitating the aggregation of image–text information. Additionally, we employ a detail enhancement module to mitigate the problem of losing image details. Moreover, our ITA accomplishes single-stage semantic segmentation via the image–text aggregation module. It outperforms the two-stage methods, which have the inherent challenges of inaccurate cropped image recognition and multiple forwarding, and thus demonstrates better performance and efficiency. Experimental results on multiple widely used benchmark datasets demonstrate that our ITA achieves excellent segmentation performance compared with the state-of-the-art open-vocabulary semantic segmentation methods. The code is available at https://github.com/huanglab-research/ITA .}, archive = {J_NEUCOM}, author = {Shengyang Cheng and Jianyong Huang and Xiaodong Wang and Lei Huang and Zhiqiang Wei}, doi = {10.1016/j.neucom.2025.129702}, journal = {Neurocomputing}, month = {5}, pages = {129702}, shortjournal = {Neurocomputing}, title = {Image–text aggregation for open-vocabulary semantic segmentation}, volume = {630}, year = {2025}, }
• (2025). TIxAI: A trustworthiness index for eXplainable AI in skin lesions classification. NEUCOM, 630, 129701. (www)

  @article{ , abstract = {Skin cancer is one of the leading causes of mortality worldwide. Early diagnosis can ensure more effective patient treatment and outcomes, but, this is challenging due to the high similarity between different skin lesion types. There is a growing interest in developing Artificial Intelligence (AI)-based systems for automated skin lesion classification. However, current AI models are not transparent, leading to a lack of trust from clinicians who struggle to interpret and validate AI decisions. To this end, in this paper, a fine tuned EfficientNet-B0-based classifier is first developed to classify dermoscopic images of Melanoma (MEL), Nevus (NV) and Seborrheic Keratosis (SK) skin lesions gathered from the International Skin Imaging Collaboration (ISIC) dataset. Next, the explainability of the model is investigated. In particular, a new Trustworthiness Index for eXplainable AI, herein referred to as TIxAI , is proposed. The TIxAI is based on the difference between the relevance degree of the lesion and non-lesion areas, leading to the conclusion that the higher the TIxAI , the more trustworthy the classifier is expected to be. Experimental results support the use of the proposed TIxAI to assess and benchmark the reliability of classifiers also in other real-world applications.}, archive = {J_NEUCOM}, author = {Cosimo Ieracitano and Francesco Carlo Morabito and Amir Hussain and Muhammad Suffian and Nadia Mammone}, doi = {10.1016/j.neucom.2025.129701}, journal = {Neurocomputing}, month = {5}, pages = {129701}, shortjournal = {Neurocomputing}, title = {TIxAI: A trustworthiness index for eXplainable AI in skin lesions classification}, volume = {630}, year = {2025}, }
• (2025). Noise-robust semi-supervised clustering learning framework considering weighted consensus and pairwise similarities. NEUCOM, 630, 129700. (www)

  @article{ , abstract = {With the growing prevalence of unstructured and noisy data in real-world applications, developing noise-robust semi-supervised clustering methods has become increasingly critical. Existing approaches often face challenges such as sensitivity to noise and inefficiency in leveraging limited labeled information effectively. With this motivation, this paper proposes a Noise-Robust semi-supervised clustering learning framework considering weighted consensus and pairwise similarities (NRWCPS). Inspired by the hierarchical clustering-based weighted consensus learning, NRWCPS proposes a semi-supervised ensemble clustering framework that is highly computationally efficient. Here, the weighted consensus mechanism is applied to consensus agreements among samples to design a comprehensive and robust clustering model. The proposed framework evaluates the significance of each partition in the final consensus by assessing the strength of its clusters, thereby minimizing the influence of noise effectively. Meanwhile, the pairwise similarities are used for sample selection, in which training samples are divided into clean and noisy sets. Results demonstrate that the proposed framework achieves superior clustering accuracy and robustness compared to existing techniques, particularly in noisy environments. Specifically, NRWCPS provides 3.9 % better average accuracy compared to the best existing method.}, archive = {J_NEUCOM}, author = {Guangchuan Hu and Amin Rezaeipanah}, doi = {10.1016/j.neucom.2025.129700}, journal = {Neurocomputing}, month = {5}, pages = {129700}, shortjournal = {Neurocomputing}, title = {Noise-robust semi-supervised clustering learning framework considering weighted consensus and pairwise similarities}, volume = {630}, year = {2025}, }
• (2025). Safe tracking control for unmanned aerial helicopter under input saturation based on reinforcement learning and disturbance observer. NEUCOM, 630, 129699. (www)

  @article{ , abstract = {This paper investigates safe tracking control for unmanned aerial helicopter (UAH) systems under outside disturbances, model uncertainties, and input saturation by utilizing reinforcement learning (RL) and an improved coordinated disturbance observer (CDO). First, the nonlinear model of the UAH system is decomposed into a position loop and an attitude loop to facilitate controller design. Second, for the position loop, a continuously differentiable function is introduced to approximate input saturation, and an improved safety protection algorithm (ISPA) is developed to ensure that all states remain within a specified safe set. Additionally, by employing an actor-critic architecture, uncertainties can be approximated through online learning, and an improved CDO is proposed to estimate the lumped disturbance, which consists of outside disturbances and approximation errors. Third, a position tracking controller is presented that uses disturbance estimation and uncertainty approximation to accurately track the safe signals generated by the ISPA. Fourth, similar to the position loop, an attitude tracking controller is implemented. Subsequently, Lyapunov stability theory is employed to analyze the bounded stability of the derived closed-loop systems, and a co-design method for the CDOs, controllers, and neural network (NN) weights is established. Finally, simulation results and comparisons demonstrate the effectiveness and advantage of our proposed control scheme.}, archive = {J_NEUCOM}, author = {Chunyu Zhang and Tao Li and Yongqiang Ye and Zehui Mao}, doi = {10.1016/j.neucom.2025.129699}, journal = {Neurocomputing}, month = {5}, pages = {129699}, shortjournal = {Neurocomputing}, title = {Safe tracking control for unmanned aerial helicopter under input saturation based on reinforcement learning and disturbance observer}, volume = {630}, year = {2025}, }
• (2025). PRNet: A progressive refinement network for referring image segmentation. NEUCOM, 630, 129698. (www)

  @article{ , abstract = {The effective feature alignment between language and image is necessary for correctly inferring the location of reference instances in the referring image segmentation (RIS) task. Previous studies usually resort to assisting target localization with the help of external detectors or using a coarse-grained positional prior during multimodal feature fusion to implicitly enhance the modal alignment capability. However, these approaches are either limited by the performance of the external detector and the design of the matching algorithm, or ignore the fine-grained features in the reference information when using the coarse-grained prior processing, which may lead to inaccurate segmentation results. In this paper, we propose a new RIS network, Progressive Refinement Network (PRNet), which aims to gradually improve the alignment quality between language and image from coarse to fine. The core of the PRNet is the Progressive Refinement Localization Scheme (PRLS), which consists of a Coarse Positional Prior Module (CPPM) and a Refined Localization Module (RLM). The CPPM obtains rough prior positional information and corresponding semantic features by calculating the similarity matrix between sentence and image. The RLM fuses information from the visual and language modalities by densely aligning pixels with word features and utilizes the prior positional information generated by the CPPM to enhance the textual semantic understanding, thus guiding the model to perceive the position of the reference instance more accurately. Experimental results show that the proposed PRNet performs well on all three public datasets, RefCOCO, RefCOCO+, and RefCOCOg.}, archive = {J_NEUCOM}, author = {Jing Liu and Huajie Jiang and Yongli Hu and Baocai Yin}, doi = {10.1016/j.neucom.2025.129698}, journal = {Neurocomputing}, month = {5}, pages = {129698}, shortjournal = {Neurocomputing}, title = {PRNet: A progressive refinement network for referring image segmentation}, volume = {630}, year = {2025}, }
• (2025). Dynamic channel token vision transformer with linear computation complexity and multi-scale features. NEUCOM, 630, 129696. (www)

  @article{ , abstract = {Original self-attention has the problem of quadratical complexity. In this paper, we propose a novel paradigm for tokenization that decouples the token scope from the spatial dimension. This new approach introduces dynamic tokens, which reduce computational complexity to linear while capturing multi-scale features. This paradigm is implemented in the proposed Dynamic Channel Token Vision Transformer (DCT-ViT), combining Window Self-Attention (WSA) and Dynamic Channel Self-Attention (DCSA) to capture both fine-grained and coarse-grained features. Our hierarchical window settings in DCSA prioritizes small tokens. DCT-ViT-S/B achieves a 82 . 9 % / 84 . 3 % Top-1 accuracy on ImageNet-1k (Deng et al., 2009) and a 47 . 9 / 49 . 8 m A P b and a 43 . 4 / 44 . 6 m A P m on COCO 2017 (Lin et al., 2014) for Mask R-CNN (He et al., 2017) 3 × schedule. The visualization of features in DCSA shows that dynamic channel tokens recognize objects at very early stages.}, archive = {J_NEUCOM}, author = {Yijia Guan and Kundong Wang}, doi = {10.1016/j.neucom.2025.129696}, journal = {Neurocomputing}, month = {5}, pages = {129696}, shortjournal = {Neurocomputing}, title = {Dynamic channel token vision transformer with linear computation complexity and multi-scale features}, volume = {630}, year = {2025}, }
• (2025). Two-tier deep and machine learning approach optimized by adaptive multi-population firefly algorithm for software defects prediction. NEUCOM, 630, 129695. (www)

  @article{ , abstract = {Software plays a progressively crucial role, where automated software systems control essential operations. Since development needs also progressively expand, manual code reviews become increasingly difficult, frequently resulting in testing lasting longer than development itself. An encouraging option for enhancing defect identification within the source code involves combining artificial intelligence and natural language processing (NLP). Analyzing source code offers an efficient approach to enhance defect detection and prevent errors in the code. This study investigates source code analysis using NLP and machine learning, where traditional and contemporary techniques of error detection are evaluated. Metaheuristics algorithms are utilized to tune machine learning classifiers, and an altered variant of the well-known firefly algorithm is proposed as part of this research. A two-tier framework is suggested, consisting of a convolutional neural network (CNN), which handles complex feature spaces, while eXtreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), and categorical boosting (CatBoost) classifiers are employed within the second-tier for improving defect detection. Supplementary simulations employing custom term frequency inverse document frequency encoding are also executed to showcase the capabilities of the suggested framework. In total, seven experiments are carried out with publicly accessible datasets. The accuracy of CNN is 80.6% for the defect prediction task, which is enhanced with the second layer using XGBoost, AdaBoost, and CatBoost to nearly 81.5%. The experiments with the NLP approach exhibit superior outcomes, where XGBoost, AdaBoost, and CatBoost achieve accuracies of 99.6%, 99.7%, and 99.8%, indicating the large potential of the suggested approach in the software testing domain.}, archive = {J_NEUCOM}, author = {John Philipose Villoth and Miodrag Zivkovic and Tamara Zivkovic and Mahmoud Abdel-salam and Mohamed Hammad and Luka Jovanovic and Vladimir Simic and Nebojsa Bacanin}, doi = {10.1016/j.neucom.2025.129695}, journal = {Neurocomputing}, month = {5}, pages = {129695}, shortjournal = {Neurocomputing}, title = {Two-tier deep and machine learning approach optimized by adaptive multi-population firefly algorithm for software defects prediction}, volume = {630}, year = {2025}, }
• (2025). STCLR: Sparse temporal contrastive learning for video representation. NEUCOM, 630, 129694. (www)

  @article{ , abstract = {Temporal Contrastive Learning for Video Representation (TCLR) is the first contrastive framework that uses temporal losses to enforce the temporal distinctiveness of the features explicitly. However, its local–local temporal contrastive loss contrasts non-overlapping dense adjacent clips that cover small motion dynamics due to temporal coherence in videos. These adjacent local clips could represent the same action phase, and that makes the local–local loss enforce temporal distinctiveness of the features within the same action phase, whereas much better distinctiveness comes from contrasting different action phases. To overcome these limitations, Sparse TCLR framework (STCLR) is proposed, which uses random sparse sampling to sample local clips that cover different motion dynamics of the video. Using sparse local clips with the local–local loss overcomes the need for the second TCLR temporal loss, the global–local loss. In addition, a novel temporal pretext (Same Speed Localization) is proposed for intra/inter action temporal self-supervised learning and combined with the temporal loss. We prove that TCLR can be simplified to a framework that uses two losses only instead of three. STCLR achieves better feature transferability by a significant margin than TCLR on video retrieval. STCLR outperforms TCLR by 0.58%, 1.21%, 0.88%, and 0.98% and by 4.32%, 5.97%, 6.51%, and 2.78% on Top 1, Top 5, Top 10, and Top 20 retrieval measures on UCF-101 and HMDB-51, respectively. On action recognition, STCLR outperforms TCLR on Diving-48 by 9.68%, which is a significant improvement, and on HMDB-51 by 0.83%, while it achieves a lower accuracy by 0.72% on UCF-101.}, archive = {J_NEUCOM}, author = {Hussein Altabrawee and Mohd Halim Mohd Noor}, doi = {10.1016/j.neucom.2025.129694}, journal = {Neurocomputing}, month = {5}, pages = {129694}, shortjournal = {Neurocomputing}, title = {STCLR: Sparse temporal contrastive learning for video representation}, volume = {630}, year = {2025}, }
• (2025). ALPPA: An accuracy-lossless and privacy-preserving aggregation strategy for federated knowledge graph completion model. NEUCOM, 630, 129693. (www)

  @article{ , abstract = {Federated knowledge graph completion facilitates data sharing and collaborative modeling among multiple clients, providing an effective solution to complete missing knowledge. To achieve privacy-preserving while maintaining model accuracy, researchers have proposed some aggregation strategies. However, these strategies typically focus solely on protecting the privacy of local model parameters uploaded by clients to the server, while neglecting the privacy of the aggregated global model parameters on the server. This limitation may allow the server to steal the global model parameters, posing risks of privacy leakages and potential economic losses for participating clients. To address this issue, we propose an accuracy-lossless and privacy-preserving aggregation (ALPPA) strategy. In our approach, clients inject noise into their local model parameters before uploading them to the server and use a shift matrix to disrupt the noise sequence. This ensures that the server cannot recover the original noise sequence, thereby preventing the global model parameters from being exposed to the server and achieving dual privacy-preserving for both local and global model parameters. After receiving the aggregated global model parameters, clients perform a noise removal operation to restore accurate and noise-free global model parameters. Since the noise is completely removed before the local training phase, the final federated model achieves the same accuracy as a model trained without noise, thus ensuring lossless accuracy. Building on ALPPA, we further propose an optimized strategy (ALPPA+) that significantly improves communication efficiency. Optimizing the noise injection module by compressing the noise matrix into a vector, ALPPA+ reduces the uplink communication cost by 50% compared to ALPPA. This optimization maintains privacy-preserving and model accuracy while significantly improving communication efficiency, offering a more efficient and secure solution for practical federated knowledge graph completion applications.}, archive = {J_NEUCOM}, author = {Songsong Liu and Xiao Song and Yong Li and Kaiqi Gong and Yuchun Tu}, doi = {10.1016/j.neucom.2025.129693}, journal = {Neurocomputing}, month = {5}, pages = {129693}, shortjournal = {Neurocomputing}, title = {ALPPA: An accuracy-lossless and privacy-preserving aggregation strategy for federated knowledge graph completion model}, volume = {630}, year = {2025}, }
• (2025). An explainable recommendation algorithm based on content summarization and linear attention. NEUCOM, 630, 129692. (www)

  @article{ , abstract = {Recommendation algorithms can alleviate the problem of information explosion and cater to the needs of users to quickly lock in preferred items, promote business development, and have important theoretical significance and broad theoretical value. Explainable recommendation algorithms can not only complete recommendation tasks, but also generate recommendation explanations, so that users can more easily accept preferences. Research related to natural language text generation has promoted the progress of explainable text generation technology for recommendation systems. This paper proposes an explainable recommendation algorithm based on content summarization and linear attention mechanism. The model uses the keyword extraction algorithm to extract key information from user comment text as an important feature of subsequent text generation tasks, and further introduces linear Transformer to improve the training speed of the model and enhance its scalability. In addition, the model also uses the Vivaldi synthetic coordinate algorithm to deeply mine user and item features and uses the Kolmogorov–Arnold neural network model to reduce the error of predicted ratings. Compared with existing leading algorithms, the algorithm in this paper has achieved significant improvements in text generation and recommendation rating prediction. This paper reveals that applying the linear attention mechanism to the explainable recommendation algorithm can greatly reduce the training cost and improve scalability, and the fusion of synthetic coordinates and attention can further mine the hidden information of the recommendation system, effectively improving the performance of the recommendation algorithm.}, archive = {J_NEUCOM}, author = {Jun Ai and Haolin Li and Zhan Su and Fengyu Zhao}, doi = {10.1016/j.neucom.2025.129692}, journal = {Neurocomputing}, month = {5}, pages = {129692}, shortjournal = {Neurocomputing}, title = {An explainable recommendation algorithm based on content summarization and linear attention}, volume = {630}, year = {2025}, }
• (2025). GOSNet: RGB-T salient object detection network based on global omnidirectional scanning. NEUCOM, 630, 129691. (www)

  @article{ , abstract = {Salient Object Detection (SOD) is a key task in computer vision. With the development of thermal infrared technology, thermal modal images provide additional salient semantic information for SOD tasks, which has led to a wide interest in RGB-T SOD. However, thermal modal images are insensitive to edge and structural information, posing challenges for RGB-T SOD tasks. Clear edges and complete structures are crucial information for ensuring the quality of saliency target prediction. Recent research has leveraged large pre-trained backbones and complex modules for saliency prediction, leading to increased computational costs. We refocus the key challenge of SOD on maintaining edge and structural information under computational constraints. Our proposed lightweight Global Omni-directional Scanning Network (GOSNet) addresses this challenge using MobileNetV2 as backbone, incorporating an ultralight Implicit Global Filtering Module (IGFM) and Multi-scale Semantic Guidance Strategy (MSGS) based on linear serialized attention. IGFM preserves edge and structural features through local–global information fusion, while the Omni-directional Scanning Decoder Module (OSDM) integrates edge information using directional gradient convolutions without additional computation. We further enhance performance through transfer learning with an Omnidirectional Filtering Attention Mechanism (OFAM), optimizing the backbone for SOD tasks without extra modules. The average improvements of GOSNet over existing lightweight SOTA methods are 0.60%, 1.81%, 1.7% and 4.88%.}, archive = {J_NEUCOM}, author = {Bochang Jiang and Dan Luo and Zihan Shang and Sicheng Liu}, doi = {10.1016/j.neucom.2025.129691}, journal = {Neurocomputing}, month = {5}, pages = {129691}, shortjournal = {Neurocomputing}, title = {GOSNet: RGB-T salient object detection network based on global omnidirectional scanning}, volume = {630}, year = {2025}, }
• (2025). An active learning model based on image similarity for skin lesion segmentation. NEUCOM, 630, 129690. (www)

  @article{ , abstract = {Medical image segmentation requires experienced doctors to provide accurate segmentation results as ground truth, since it involves plentiful medical knowledge. The deep learning-based model requires a large number of medical images and ground truth, during the training stage, which brings a heavy workload to the doctor. Therefore, we design an active learning based on image similarity framework for selection strategy to select more representative images for doctors to obtain the ground truth, which saves their time in segmenting similar medical images. Eventually, these representative images will serve as training sets to train a more accurate segmentation model. In this paper, we propose the active learning-based segmentation (ALS) model for medical images, in which we use the mean structure similarity index measure to measure how similar the two images are. Then we select some highly similar images, since medical image segmentation is different from semantic segmentation, there is no interference from similar targets, so we do not need to cull interfering images when segmenting medical images. Finally, we reduce the amount of data to 75%, 50%, 25%, and 12.5% to validate our ALS model. Even a small amount of image through ALS model selection as a training set, and the final segmentation model of training also has a good performance. The active learning model can reduce the cost of labeling and obtain accurate segmentation results through experiments, and even better than the segmentation results with a large number of highly similar data.}, archive = {J_NEUCOM}, author = {Xiu Shu and Zhihui Li and Chunwei Tian and Xiaojun Chang and Di Yuan}, doi = {10.1016/j.neucom.2025.129690}, journal = {Neurocomputing}, month = {5}, pages = {129690}, shortjournal = {Neurocomputing}, title = {An active learning model based on image similarity for skin lesion segmentation}, volume = {630}, year = {2025}, }
• (2025). Multi-view clustering integrating anchor attribute and structural information. NEUCOM, 630, 129689. (www)

  @article{ , abstract = {Multisource data has driven the development of advanced clustering algorithms, such as multi-view clustering, which critically rely on the construction of similarity matrices. Traditional algorithms typically generate these matrices based solely on node attributes. However, for certain directed real-world networks, neglecting the asymmetric structural relationships between nodes may compromise the accuracy of clustering results. This paper introduces a novel multi-view clustering algorithm, AAS, which employs a two-step proximity approach using anchors in each view, effectively integrating both attribute and directed structural information. This method enhances the clarity of cluster features within the similarity matrices. The construction of the anchor structural similarity matrix utilizes strongly connected components of directed graphs. The entire process—from the construction of similarity matrices to clustering—is formulated within a unified optimization framework. Comparative experiments conducted on the modified Attribute SBM dataset, benchmarked against seven other algorithms, demonstrate the effectiveness and superiority of AAS.}, archive = {J_NEUCOM}, author = {Xuetong Li and Xiao-Dong Zhang}, doi = {10.1016/j.neucom.2025.129689}, journal = {Neurocomputing}, month = {5}, pages = {129689}, shortjournal = {Neurocomputing}, title = {Multi-view clustering integrating anchor attribute and structural information}, volume = {630}, year = {2025}, }
• (2025). How robust are ensemble machine learning explanations?. NEUCOM, 630, 129686. (www)

  @article{ , abstract = {To date, several explainable AI methods are available. The variability of the resulting explanations can be high, especially when many input features are considered. This lack of robustness may limit their usability. In this paper we try to fill this gap, by contributing a methodology that: i) is able to measure the robustness of a given set of explanations; ii) suggests how to improve robustness, by tuning the model parameters. Without loss of generality, we exemplify our proposal for ensemble tree models, which typically reach a high predictive performance in classification problems. We consider a toy case study with artificially generated data as well as two real case studies whose application domain is cybersecurity and more precisely the models used for detecting phishing websites.}, archive = {J_NEUCOM}, author = {Maria Carla Calzarossa and Paolo Giudici and Rasha Zieni}, doi = {10.1016/j.neucom.2025.129686}, journal = {Neurocomputing}, month = {5}, pages = {129686}, shortjournal = {Neurocomputing}, title = {How robust are ensemble machine learning explanations?}, volume = {630}, year = {2025}, }
• (2025). Type affinity network for distantly supervised relation extraction. NEUCOM, 630, 129684. (www)

  @article{ , abstract = {Distantly Supervised Relation Extraction (DSRE) focuses on mining relational information between entities from text, which is a critical step in constructing and expanding knowledge graphs. However, this task typically faces the problem of long-tail distribution of relationships. Furthermore, much previous research has been limited to specific relational features, overlooking the potential discriminative information in the dependencies among these features. To address these issues, we propose a Type Affinity Network (TAN) that explicitly models the dependencies among relational features. Using differentiated entity-type information and local context information, we extract high-quality feature representations. The dynamic integration of type dependency information via a type affinity matrix effectively leverages potential discriminative information, enhancing relation extraction accuracy. Furthermore, aggregating implicit features of relationships forms base points as references, improving the ability of the model to handle long-tail relations. These relationship base points differ from traditional hierarchical structures, providing a more general method. We conduct extensive experiments on two benchmark DSRE datasets, namely NYT-10 and WIKI-20 m. Compared with multiple state-of-the-art DSRE methods, the TAN not only shows significant performance improvements in handling long-tail relations, but also demonstrates overall performance improvement.}, archive = {J_NEUCOM}, author = {Wei Song and Juncheng Zhou and Xiao Liu}, doi = {10.1016/j.neucom.2025.129684}, journal = {Neurocomputing}, month = {5}, pages = {129684}, shortjournal = {Neurocomputing}, title = {Type affinity network for distantly supervised relation extraction}, volume = {630}, year = {2025}, }
• (2025). Diversity augmentation and multi-fuzzy label for semi-supervised semantic segmentation. NEUCOM, 630, 129681. (www)

  @article{ , abstract = {Semantic segmentation aims to provide pixel-wise accurate predictions for images. Semi-supervised semantic segmentation aims to learn a semantic segmentation model using a limited number of labeled images and a large fraction of unlabeled images. Existing methods primarily focus on introducing additional models or complex training procedures but overlook the model itself and such complex strategies tend to discard many usable pixels, exacerbating the class imbalance problem. In this paper, we propose DAM for semi-supervised semantic segmentation, a simple yet effective method that mainly focuses on the inputs and outputs of the model itself. For the input component, we posit that diverse data augmentations can provide more semantic information. Therefore, we propose a method called Random Diversity Augmentations. Given an unlabeled image, we apply different triple-level data augmentations to provide more semantic information. For the output component, our approach is inspired by the fact that many unreliable predictions are confused only among the top classes rather than all classes, so we contend that fuzzy pixels can still provide valuable guidance to the model. Specifically, we select fuzzy pixels based on confidence and assign multi-fuzzy labels to these pixels for training the model, which allows us to leverage the information more effectively. Our straightforward DAM achieves new state-of-the-art performance on SSS different benchmarks. Code is available at https://github.com/Wang-zhenyan/DAM .}, archive = {J_NEUCOM}, author = {Zhenyan Wang and Zhenxue Chen and Chengyun Liu and Yaping Zhao and Xinming Zhu and Q.M. Jonathan Wu}, doi = {10.1016/j.neucom.2025.129681}, journal = {Neurocomputing}, month = {5}, pages = {129681}, shortjournal = {Neurocomputing}, title = {Diversity augmentation and multi-fuzzy label for semi-supervised semantic segmentation}, volume = {630}, year = {2025}, }
• (2025). Eliciting analogical reasoning from language models in retrieval-augmented translation under low-resource scenarios. NEUCOM, 630, 129680. (www)

  @article{ , abstract = {Retrieval-Augmented Neural Machine Translation (RANMT), which augments translation models with relevant examples fetched by a similarity retriever, is proficient in well-resourced translations. However, its inherent advantages are not fully realized in low-resource contexts, where the sparsity of data can often result in less relevant or less useful information for translation. Recent literature indicates that nearest-neighbor examples from small training data have the unfortunate effect of impairing RANMT performance. Our examination of 16 low-resource tasks reveals a sharp deterioration in performance of a multilingual language model when conditioned on retrieved examples compared to direct translation. To address this problem, we explore a framework based on analogical reasoning, aiming to enhance the capacity of language models to infer translations from parallel examples in limited data settings. This framework mimics a cognitive process of human translation by structuring examples in analogy patterns. We propose a multi-objective learning strategy that augments vanilla training for conditional translation to learn latent knowledge from examples. We also investigate different retrieval methods for selecting translation examples based on lexical similarity, semantic relatedness, or a combination of both. The results show that our approach is effective in optimizing RANMT in low-resource settings, delivering notable improvements across all retrieval settings. In particular, augmented training akin to reasoning with analogies in two directions, contributes significantly to deriving benefits from examples, even when their relevance is limited. Moreover, our approach demonstrates superior performance in low-resource translation tasks compared to prompting large language models in few-shot contexts. It also proves to be competitive with models that have been extensively trained using substantial amounts of supervised data.}, archive = {J_NEUCOM}, author = {Liyan Wang and Bartholomäus Wloka and Yves Lepage}, doi = {10.1016/j.neucom.2025.129680}, journal = {Neurocomputing}, month = {5}, pages = {129680}, shortjournal = {Neurocomputing}, title = {Eliciting analogical reasoning from language models in retrieval-augmented translation under low-resource scenarios}, volume = {630}, year = {2025}, }
• (2025). NFI-SGAT: Node feature initialization-based streaming graph learning model with graph attention network for critical node identification in temporal networks. NEUCOM, 630, 129679. (www)

  @article{ , abstract = {Most real-world systems can be represented as temporal networks. Identifying critical nodes is crucial to better understanding and controlling the structure and evolution of the network, thereby optimizing and regulating the network effectively. However, it is difficult to identify critical nodes in temporal networks as there are new links emerging or removed and the structure of temporal network keeps changing. Recent years have witnessed increasing efforts made to extend static methods to temporal networks, which would overlook time-series features like dynamic changes and temporal dependencies between node connections, thus causing loss of temporal information. In this paper, we propose NFI-SGAT, a novel N ode F eature I nitialization-based S treaming graph learning model with G raph A ttention ne T work. To mitigate the loss of temporal information problem, NFI-SGAT continuously updates node representations with a novel streaming temporal graph learning module based on the initial node features. To capture nodes’ topology information comprehensively in temporal networks, we generate initial multi-scale feature vectors. To address the staleness problem, we utilize attention mechanism to aggregate information of nodes’ temporal neighborhood. Experimental results on eight real-world temporal networks demonstrate the effectiveness of the proposed framework.}, archive = {J_NEUCOM}, author = {Mingxuan Yan and Yuexing Han and Bing Wang}, doi = {10.1016/j.neucom.2025.129679}, journal = {Neurocomputing}, month = {5}, pages = {129679}, shortjournal = {Neurocomputing}, title = {NFI-SGAT: Node feature initialization-based streaming graph learning model with graph attention network for critical node identification in temporal networks}, volume = {630}, year = {2025}, }
• (2025). Neuromorphic event-based recognition boosted by motion-aware learning. NEUCOM, 630, 129678. (www)

  @article{ , abstract = {Neuromorphic visual sensors (event cameras) that produce asynchronous responses to changes in pixel brightness offer significant advantages such as high dynamic range, low power consumption, and reduced data redundancy. These features enable breakthroughs in object recognition and detection under extreme conditions such as fluctuating brightness or rapid movements. Therefore, event cameras have attracted great attention in many mainstream applications such as autonomous driving, UAV, remote sensing, etc. However, their unique output format poses challenges to existing vision models. Typically, event signals are converted into event-based frames for processing by 2D convolutional neural networks (CNNs). While 2D CNNs are effective at capturing spatial semantics, they are less adept at encoding motion information. To overcome this, we introduce a motion-aware branch ( MAB ) as an auxiliary network for 2D CNNs. The MAB leverages self- and cross-attention mechanisms to integrate learned motion messages from event frame patches into the recognition backbone. We unify features from scattered 3D patches and dense 2D features into joint representations for multi-branch fusion through spatial and temporal feature guidance and alignment. The fused features are then input into the classifier for prediction. Extensive experiments demonstrate the quantitative and qualitative superiority of our approach compared to previous methods.}, archive = {J_NEUCOM}, author = {Yuhan Liu and Yongjian Deng and Bochen Xie and Hai Liu and Zhen Yang and Youfu Li}, doi = {10.1016/j.neucom.2025.129678}, journal = {Neurocomputing}, month = {5}, pages = {129678}, shortjournal = {Neurocomputing}, title = {Neuromorphic event-based recognition boosted by motion-aware learning}, volume = {630}, year = {2025}, }
• (2025). MSOF: A main and secondary bi-directional optical flow feature method for spotting micro-expression. NEUCOM, 630, 129676. (www)

  @article{ , abstract = {Micro-expressions have the potential to reveal genuine emotions, making them particularly important in fields such as public safety and healthcare. However, spotting micro-expressions is challenging due to their subtle movements and brief duration. Existing micro-expression spotting methods need more accuracy in feature extraction and more ability to capture tiny movements. To address these problems, we propose a novel feature extraction method based on optical flow direction, named the Main and Secondary bi-directional Optical Flow feature (MSOF) extraction method. Firstly, to comprehensively encompass compelling motion features, we propose separately extracting optical flow features in the main and secondary directions of the region of interest. Next, to obtain smooth optical flow curves and better present the trend of these curves, we introduce low-pass filtering and empirical mode decomposition to remove high-frequency noise. Subsequently, to improve the accuracy of spotting candidate micro-expression segments, we devise a method for determining thresholds to screen the optical flow curves. Finally, we construct a non-maximum suppression method to determine the final expression intervals and achieve micro-expression spotting to process the candidate expression segments. We perform experiments on the CAS(ME) 2 and SAMM-LV datasets, and the results demonstrate that our method is competitive and outperforms the existing methods. In addition, we validate on the MEGC2022 test set, which shows that the overall F1-score of our method outperforms in the first place.}, archive = {J_NEUCOM}, author = {Henian Yang and Shucheng Huang and Mingxing Li}, doi = {10.1016/j.neucom.2025.129676}, journal = {Neurocomputing}, month = {5}, pages = {129676}, shortjournal = {Neurocomputing}, title = {MSOF: A main and secondary bi-directional optical flow feature method for spotting micro-expression}, volume = {630}, year = {2025}, }
• (2025). Measuring hierarchically-organized interactions in dynamic networks through spectral entropy rates: Theory, estimation, and illustrative application to physiological networks. NEUCOM, 630, 129675. (www)

  @article{ , abstract = {Recent advances in signal processing and information theory are boosting the development of new approaches for the data-driven modeling of complex network systems. In the fields of Network Physiology and Network Neuroscience where the signals of interest are rich of oscillatory content, the spectral representation of network systems is essential to ascribe interactions to specific oscillations with physiological meaning. The present work introduces a coherent framework integrating several information dynamics approaches to quantify node-specific, pairwise and higher-order interactions in network systems. A hierarchical organization of interactions of different order is established using measures of entropy rate, mutual information rate and O-information rate to quantify the dynamics of individual nodes, the links between pairs of nodes, and the redundant/synergistic hyperlinks in groups of nodes. All measures are formulated in the time domain and expanded to the spectral domain to obtain frequency-specific information. The practical computation of all measures is favored presenting a toolbox that implements parametric and non-parametric estimators and includes statistical validation approaches. The framework is illustrated using theoretical examples where the properties of the measures are displayed in benchmark simulated network systems, and representative multivariate time series in the context of Network Neuroscience and Network Physiology.}, archive = {J_NEUCOM}, author = {Laura Sparacino and Yuri Antonacci and Gorana Mijatovic and Luca Faes}, doi = {10.1016/j.neucom.2025.129675}, journal = {Neurocomputing}, month = {5}, pages = {129675}, shortjournal = {Neurocomputing}, title = {Measuring hierarchically-organized interactions in dynamic networks through spectral entropy rates: Theory, estimation, and illustrative application to physiological networks}, volume = {630}, year = {2025}, }
• (2025). A simplified adaptive fuzzy Min–Max neural network for pattern classification. NEUCOM, 630, 129668. (www)

  @article{ , abstract = {Hyperbox classifiers have been extensively applied in pattern classification tasks. The fuzzy min–max (FMM) neural network, as a prominent hyperbox classifier, has undergone significant research over the past three decades. Despite its effectiveness, the conventional FMM algorithm relies on a fixed expansion coefficient, which often leads to suboptimal performance because it cannot adapt to varying data distributions. Additionally, the membership function in FMM typically depends on sensitive parameters that need to be preset, further complicating its practical application. To address these limitations and improve the adaptability and classification accuracy of FMM, we propose a Simplified Adaptive Fuzzy Min–Max Neural Network (SAFMM) for pattern classification, where all parameters are adaptively chosen. Our contributions are twofold: Firstly, a novel FMM learning algorithm is introduced, where hyperboxes expand adaptively instead of using a fixed coefficient, and cluster and class-specific hyperboxes are simultaneously applied to avoid overlap. Secondly, considering uneven data distributions within hyperboxes, a new adaptive membership function based on data core is defined for classification. The proposed SAFMM model is evaluated on benchmark datasets and a real-world case of magnetic flux leakage (MFL) defect depth estimation in pipeline non-destructive testing (NDT). Experimental results demonstrate that SAFMM can generate fewer hyperboxes while ensuring classification accuracy and exhibits an absolute advantage for quick classification as the dataset size increases.}, archive = {J_NEUCOM}, author = {Mingrui Fu and Shuai Chen and Xiaoxiao Wei and Jinsong Du and Wei Wang and Jinhai Liu}, doi = {10.1016/j.neucom.2025.129668}, journal = {Neurocomputing}, month = {5}, pages = {129668}, shortjournal = {Neurocomputing}, title = {A simplified adaptive fuzzy Min–Max neural network for pattern classification}, volume = {630}, year = {2025}, }
• (2025). Deep learning for medical imaging super-resolution: A comprehensive review. NEUCOM, 630, 129667. (www)

  @article{ , abstract = {Medical image super-resolution has always been a challenging field, and high-quality medical images are crucial for disease diagnosis. In recent years, deep learning-based medical image super-resolution has achieved significant progress, resulting in numerous innovative breakthroughs. However, comprehensive reviews of the latest network architectures, their variants, and future perspectives in this field remain limited. Therefore, this study provides an extensive review of the current research status and development trends in medical image super-resolution technology, comprehensively examining the latest variants of various network models and the specific super-resolution problems they address to remedy this insufficiency. Firstly, we summarize the noise issues faced by medical imaging within clinical settings, as well as the methods of image degradation. Secondly, based on the principles of super-resolution algorithms, we categorize the existing methods into six types: convolutional neural networks, generative adversarial networks, Transformers, diffusion models, multi-contrast architectures, and collaborative learning architectures, and analyze them in detail. Subsequently, we review the commonly used datasets, loss functions, and evaluation metrics for medical image super-resolution. Additionally, we discuss the applications of super-resolution techniques across various medical imaging modalities, including CT and MRI. Finally, the challenges and future development directions are highlighted, including network structure optimization, improvement of evaluation metrics, clinical application potential, model robustness, and other key areas.}, archive = {J_NEUCOM}, author = {Hanguang Xiao and Zhiying Yang and Tianqi Liu and Shihong Liu and Xiaoxuan Huang and Jiahui Dai}, doi = {10.1016/j.neucom.2025.129667}, journal = {Neurocomputing}, month = {5}, pages = {129667}, shortjournal = {Neurocomputing}, title = {Deep learning for medical imaging super-resolution: A comprehensive review}, volume = {630}, year = {2025}, }
• (2025). Smooth policy iteration for zero-sum markov games. NEUCOM, 630, 129666. (www)

  @article{ , abstract = {Zero-sum Markov Games have been an efficient framework for robust reinforcement learning wherein a minimax problem is constructed to find the equilibrium policies. As a classic solution, the maximum operator is exactly calculated to find the worst-case performance over an uncertainty set, and then the minimum operator is iteratively solved to optimize the policy. However, it suffers from low efficiency to handle complex tasks as finding the maximum over large scale action spaces is usually cumbersome. In this paper, we propose the smooth policy iteration (SPI) to improve the solving efficiency of zero-sum Markov Games by replacing the maximum operator with the thoughtfully devised approximation function, i.e., Weight LogSumExp (WLSE). It is shown that SPI is capable of conducting the policy evaluation with fewer iterations and a controllable error bound. We prove the convergence of SPI based on the contraction mapping theorem and show it can reach the nearly optimal Nash equilibrium policies. To handle the continuous tasks like autonomous driving, we propose an efficient algorithm called smooth adversarial actor–critic (SAAC) by extending SPI with neural networks. The target value related to WLSE function is evaluated by the sampled trajectories and mean square error is constructed to optimize the value function. And the gradient-ascent-descent method is adopted to optimize the protagonist and adversary policies jointly. We verify our algorithm in both tabular games and continuous path tracking tasks. Results show that SPI can reduce the iteration number by 28.6% without changing the optimal policies, and SAAC can improve the training efficiency of robust policy by 8.3 times compared with mainstream methods.}, archive = {J_NEUCOM}, author = {Yangang Ren and Yao Lyu and Wenxuan Wang and Shengbo Eben Li and Zeyang Li and Jingliang Duan}, doi = {10.1016/j.neucom.2025.129666}, journal = {Neurocomputing}, month = {5}, pages = {129666}, shortjournal = {Neurocomputing}, title = {Smooth policy iteration for zero-sum markov games}, volume = {630}, year = {2025}, }
• (2025). Auto-compression transfer learning methodology for deep convolutional neural networks. NEUCOM, 630, 129661. (www)

  @article{ , abstract = {This paper presents an auto-compression methodology for deep convolutional neural networks (DCNN) based on transfer learning. We propose a novel strategy that applies hyperparameter tuning to the top of DCNN in the training phase while simultaneously carrying out a pruning technique. The above aims to reduce the time and resources required to generate a precise and lightweight convolutional neural network. We use a bioinspired pruning strategy that can delete complete layers of the architecture by using pre-trained translator layers. This results in the pruning step becoming another hyperparameter to adjust. This methodology virtually offers a group of sub-architectures with various compression rates. Therefore, the hyperparameter tuning step also performs the pruning step. We also introduced a custom metric based on the accuracy of the network and the compression rate of the architecture, allowing us to select the trade-off between them. Therefore, our proposed methodology only requires a base architecture and a new classification task to automatically yield the best (accurate and compressed) candidates during the hyperparameter search execution. As a result, we developed a Python library with multiple MobileNet, MobileNetV2, and EfficientNetB5 sub-architectures (pre-trained on ImageNet) as backbone feature extractors. We demonstrate its efficiency in three classification tasks, achieving compression levels up to 80% and competitive accuracy scores compared to their counterparts created with the classical transfer learning approach.}, archive = {J_NEUCOM}, author = {J.D. Camacho and Carlos Villaseñor and Javier Gomez-Avila and Carlos Lopez-Franco and Nancy Arana-Daniel}, doi = {10.1016/j.neucom.2025.129661}, journal = {Neurocomputing}, month = {5}, pages = {129661}, shortjournal = {Neurocomputing}, title = {Auto-compression transfer learning methodology for deep convolutional neural networks}, volume = {630}, year = {2025}, }
• (2025). Prompt suffix-attack against text-to-image diffusion models. NEUCOM, 630, 129659. (www)

  @article{ , abstract = {Text-to-image diffusion models (T2I DMs) have achieved excellent performance in image generation. However, the adversarial robustness of T2I DMs has not been sufficiently explored. Most existing works focus on guiding T2I DMs to generate not safe for work (NSFW) images, primarily aiming to bypass the safety checkers of T2I DMs rather than address the adversarial vulnerabilities inherent to the models themselves. While some existing studies have achieved perturbations in the content of the generated images by appending a certain number of suffix characters to the text prompts, which we call prompt suffix attack (PSA). However, they lack a systematic exploration of the underlying mechanisms of these attacks. In this work, we conduct a detailed study to better understand such character-level failure modes of the T2I DMs’ text encoders. To achieve this, we investigate the performance of various PSA strategies targeting the text encoder of Stable Diffusion. We incorporate four established algorithms, namely Particle Swarm Optimization (PSO), Simulated Annealing (SA), Hotflip and GPT-based approaches. Furthermore, we propose an integrated gradient-free algorithm, G&S, and systematically compare its performance with existing methods using quantitative metrics and visualization techniques. Experimental results demonstrate that G&S achieves significant advantages in character-level perturbation attacks. Subsequently, we use G&S to investigate the mechanism behind the character-level failure modes of SD’s text encoder, attributing these vulnerabilities to two key factors: the fragility of contextual encoding and the nonlinear effects within the semantic space. Additionally, through extensive experiments in both gray-box and black-box settings, we demonstrate that this vulnerability in text encoder is pervasive across T2I DMs, offering new insights for future research on the robustness of T2I DMs.}, archive = {J_NEUCOM}, author = {Siyun Xiong and Yanhui Du and Zhuohao Wang and Peiqi Sun}, doi = {10.1016/j.neucom.2025.129659}, journal = {Neurocomputing}, month = {5}, pages = {129659}, shortjournal = {Neurocomputing}, title = {Prompt suffix-attack against text-to-image diffusion models}, volume = {630}, year = {2025}, }
• (2025). MFEA-net: A pixel-adaptive multigrid finite element analysis neural network for efficient material response prediction. NEUCOM, 630, 129657. (www)

  @article{ , abstract = {In this study, a novel physics-guided machine learning model, MFEA-Net, is proposed for the efficient prediction of mechanical responses in both single and multi-phase material systems. The MFEA-Net employs a data-driven Pixel-Adaptive Convolution (PAC) structure, unifying the modeling of arbitrary complex multi-phase systems. This is a significant enhancement for the finite element analysis neural networks (FEA-Net) that solve responses at the pixel level. Additionally, by integrating a geometric multigrid structure, the proposed model achieves ultra-fast convergence rates, substantially reducing the computational time required compared to conventional iterative methods. Smoothers for each grid level are refined using a three-layer convolutional neural network, innovatively adapted from the Jacobi smoother kernel to enhance performance. These modifications ensure excellent generalization capabilities and facilitate a computational algorithm characterized by linear time complexity. Several numerical experiments are performed to demonstrate and verify the proposed method with benchmark methods. The proposed MFEA-Net exhibits remarkable improvement in convergence efficiency over traditional Jacobi-based iteration methods, enabling it to predict material response 3-4 orders of magnitude faster.}, archive = {J_NEUCOM}, author = {Changyu Meng and Houpu Yao and Yongming Liu}, doi = {10.1016/j.neucom.2025.129657}, journal = {Neurocomputing}, month = {5}, pages = {129657}, shortjournal = {Neurocomputing}, title = {MFEA-net: A pixel-adaptive multigrid finite element analysis neural network for efficient material response prediction}, volume = {630}, year = {2025}, }
• (2025). Assessing microbial colony counting: A deep learning approach with the AGAR image dataset. NEUCOM, 630, 129654. (www)

  @article{ , abstract = {This study introduces the Annotated Germs for Automated Recognition (AGAR) dataset, an image database designed to advance the automation of microbial colony detection and classification. It encompasses an extensive collection of 18,000 high-resolution images depicting five distinct microorganisms, meticulously curated as single or mixed cultures, taken under diverse lighting conditions using two different camera setups. The dataset categorizes images into countable, uncountable, and empty, with the countable images further labeled by expert microbiologists to mark colony locations and species identification, totaling 336,442 colonies. We detail the comprehensive methodology employed in the creation of the AGAR database, emphasizing the rigorous annotation process by a team of professional microbiologists who have worked with microbial collection and identification for over 20 years. This process ensures the high-quality data necessary for training machine learning models. Our investigation further explores the performance of deep neural network architectures for object detection, namely Faster R-CNN and Cascade R-CNN with four different backbones, in leveraging the AGAR dataset for the tasks of microbial colony localization and species classification. The results confirmed the great potential of deep learning methods to automate the process of microbe localization and classification based on Petri dish photos. AGAR stands as the first dataset of this kind, both in scale and in public availability, laying the groundwork for future advancements in the field of machine learning applied to microbiology ( https://agar.neurosys.com/ ).}, archive = {J_NEUCOM}, author = {Sylwia Majchrowska and Jarosław Pawłowski and Grzegorz Guła and Tomasz Bonus and Agata Hanas and Adam Loch and Agnieszka Pawlak and Justyna Roszkowiak and Tomasz Golan and Zuzanna Drulis-Kawa}, doi = {10.1016/j.neucom.2025.129654}, journal = {Neurocomputing}, month = {5}, pages = {129654}, shortjournal = {Neurocomputing}, title = {Assessing microbial colony counting: A deep learning approach with the AGAR image dataset}, volume = {630}, year = {2025}, }
• (2025). NeRF dynamic scene reconstruction based on motion, semantic information and inpainting. NEUCOM, 630, 129653. (www)

  @article{ , abstract = {In this work, we address the inherent limitations of Neural Radiance Field (NeRF) in synthesizing novel viewpoints within dynamic environments, particularly those compromised by moving objects. Such scenarios frequently yield reconstructions of suboptimal quality, characterized by blurriness and the presence of artifacts, which significantly undermines the fidelity of synthetic scenes. This limitation significantly restricts the potential applications of NeRF in autonomous driving contexts, such as scene editing, high-precision map construction, and related functionalities. To overcome these challenges, we propose a novel NeRF-based approach tailored to address the complexities associated with moving objects in monocular driving scenarios. The proposed approach combines optical flow analysis and semantic information to precisely detect and localize moving objects. This was then followed by an inpainting technique that guides the NeRF reconstruction process, effectively mitigating the adverse impacts of dynamic elements within the scene. Our model is further enhanced by incorporating depth and semantic data to refine the training process. We validate the efficacy of our approach through comprehensive experimentation on both synthetic and real-world driving datasets, as well as on challenging self-recorded realistic driving scenes. Our method achieves a performance improvement of up to 13% compared to previous state-of-the-art methods. Additionally, we verify the efficacy of our approach through comprehensive ablation analyses. Both the quantitative and qualitative results demonstrate the superiority especially in dynamic driving scenes, advancing the potential applications in autonomous driving contexts. Our code and self-collected data are available at https://github.com/GandalfTGrey/Nerf-KBS.git .}, archive = {J_NEUCOM}, author = {Jiaxuan Liu and Huan Cheng and Shuo Wang and Fang Zhao and Meng Li}, doi = {10.1016/j.neucom.2025.129653}, journal = {Neurocomputing}, month = {5}, pages = {129653}, shortjournal = {Neurocomputing}, title = {NeRF dynamic scene reconstruction based on motion, semantic information and inpainting}, volume = {630}, year = {2025}, }
• (2025). A non-deteriorating approach to improve natural language understanding in conversational agents. NEUCOM, 630, 129652. (www)

  @article{ , abstract = {The performance of Conversational Agents (CAs) relies heavily on intent classification and entity extraction. However, the effectiveness of these tasks is often hindered by the absence of an explicit model that accounts for their interdependence. To address this limitation, we introduce a novel approach that leverages the inter-dependency between intents and entities to improve the Exact Match Accuracy (EMA) of CAs. The approach evaluates the consistency of the output generated by the Natural Language Understanding (NLU) component based on a specification that outlines all valid combinations of intents and entities. If an inconsistent output is detected, the ranking of predicted intents is leveraged to determine the most probable intent that aligns with the identified entities. The technique guarantees the preservation or improvement of EMA and can be applied as a post-processing step in combination with any existing NLU method that returns a ranking of intents. The proposed approach was evaluated through an ablation study conducted using three different NLU models (DIET, DCA-Net and Bi-model) on four distinct conversational datasets: ATIS, SNIPS, NLU-Benchmark, and AWPS. The results demonstrated that our method led to a consistent improvement in the EMA across all NLU methods and datasets. The magnitude of the improvement varies depending on the method, dataset, and training data size, ranging from below 1% when using DCA-Net or Bi-model on the SNIPS dataset with full training data, to over 9% when using Bi-model or DIET on AWPS with 10% of the available training data.}, archive = {J_NEUCOM}, author = {Miguel Arevalillo-Herráez and Romina S. Albornoz-De Luise and Yuyan Wu}, doi = {10.1016/j.neucom.2025.129652}, journal = {Neurocomputing}, month = {5}, pages = {129652}, shortjournal = {Neurocomputing}, title = {A non-deteriorating approach to improve natural language understanding in conversational agents}, volume = {630}, year = {2025}, }
• (2025). Encoder augmentation for multi-task graph contrastive learning. NEUCOM, 630, 129650. (www)

  @article{ , abstract = {Graph contrastive learning (GCL), as one of the most popular self-supervised paradigms, has achieved a great deal of success in the field of graph representation learning. However, many GCL methods face a limitation in generating diverse enough views due to the absence of prior knowledge, leading to suboptimal performance. Moreover, the community structures are rarely considered. We argue that the semantic similarities among nodes within the same community substantially contribute to graph analyzing tasks. To address these issues, we propose an Encoder Augmentation for Multi-task Graph Contrastive Learning method, known as EA-MGCL. Specifically, we introduce a multi-layer information weighting aggregation method to enhance the diversity of the generated views within the encoder augmentation module. Additionally, we present a multi-task graph contrastive learning module to discern local community information. The proposed EA-MGCL effectively captures and leverages node-level, node-community-level, and community-level information via three contrastive learning strategies. Extensive experimental results on six real-world datasets demonstrate that EA-MGCL achieves superior performance compared to state-of-the-art methods. The codes of this work are available at https://github.com/ZZY-GraphMiningLab/EA-MGCL .}, archive = {J_NEUCOM}, author = {Xiaoyu Wang and Qiqi Zhang and Gen Liu and Zhongying Zhao and Hongzhi Cui}, doi = {10.1016/j.neucom.2025.129650}, journal = {Neurocomputing}, month = {5}, pages = {129650}, shortjournal = {Neurocomputing}, title = {Encoder augmentation for multi-task graph contrastive learning}, volume = {630}, year = {2025}, }
• (2025). Unsupervised super resolution using dual contrastive learning. NEUCOM, 630, 129649. (www)

  @article{ , abstract = {The unsupervised learning method is effective in addressing the real world image super resolution problem, as obtaining paired real world images is nearly impossible. The underlying idea of some methods is to train networks using synthetic datasets with various complex noise. Through image translation and cycle consistency, noisy images are transformed into clean images, which are then upscaled using a super resolution network to obtain the final results. While these methods yield some effectiveness, relying solely on cycle consistency makes it difficult to ensure that the network model learns sufficient latent features from noisy and clean images. In this paper, we present a novel network model that explores inter image features through dual contrastive learning and adversarial network framework, effectively denoising images for better super resolution results. The model consists of three coordinated branches: the Image Translation Branch (ITB), the Contrastive Learning Branch (CLB), and the Super Resolution Branch (SRB). Specifically, ITB employs cycle consistency to learn bidirectional mappings between images, transforming noisy images into clean ones. The CLB branch utilizes a dual contrastive learning strategy to extract latent features from noisy and clean images, pulling similar features closer and pushing dissimilar features apart, thereby guiding ITB to improve denoising. SRB uses a super resolution network to upscale the denoised images to HR images, yielding the final results. Extensive experiments conducted on multiple datasets, including AIM19 (PSNR:22.194/SSIM:0.595), NTIRE20 (PSNR:24.631/SSIM:0.622), and Urban100 (PSNR:21.274/SSIM:0.620), have confirmed the superiority of the proposed method.}, archive = {J_NEUCOM}, author = {Chao Wu and Yuan Jing}, doi = {10.1016/j.neucom.2025.129649}, journal = {Neurocomputing}, month = {5}, pages = {129649}, shortjournal = {Neurocomputing}, title = {Unsupervised super resolution using dual contrastive learning}, volume = {630}, year = {2025}, }
• (2025). Deep hierarchical learning on point clouds in feature space. NEUCOM, 630, 129647. (www)

  @article{ , abstract = {D ynamic G raph CNN (DGCNN) exhibits impressive performance on point cloud recognition benchmarks. However, by analyzing the operations it employed, we find three limitations that constrain the capability of DGCNN: (1) The shallow design of EdgeConv restricts its ability to capture fine-grained structural information; (2) The computational cost grows linearly with the stack of stages; (3) The max global aggregator discards most useful points. In this work, we introduce a hierarchical dynamic graph network dubbed HDGNet to address these issues. Our HDGNet inherits the dynamic feature learning property of DGCNN and extends it to other operations. First, we propose an EdgeConv++ module to extract neighborhood features in a powerful and efficient fashion. Second, we construct a hierarchical network based on farthest feature sampling (FFS) to reduce computational costs. Lastly, we design a general and effective weighted sort pooling (WS-Pooling) to aggregate the global features. Extensive experiments show that HDGNet outperforms the previous state-of-the-art on various benchmarks. For 3D classification, HDGNet reaches 94.3% and 88.6% on ModelNet40 and ScanObjectNN, surpassing DGCNN by 1.4% and 10.5%, respectively, while testing 3 × faster. For segmentation, HDGNet achieves 87.1% instance IoU on ShapeNetPart and 72.5% mIoU on S3DIS Area 5, being superior to the recent PointMetaBase and PointVector.}, archive = {J_NEUCOM}, author = {Lingbin Bu and Yifan Wang and Qiming Ma and Zhiwen Hou and Rong Wang and Fanliang Bu}, doi = {10.1016/j.neucom.2025.129647}, journal = {Neurocomputing}, month = {5}, pages = {129647}, shortjournal = {Neurocomputing}, title = {Deep hierarchical learning on point clouds in feature space}, volume = {630}, year = {2025}, }
• (2025). UniASM: Binary code similarity detection without fine-tuning. NEUCOM, 630, 129646. (www)

  @article{ , abstract = {Binary code similarity detection (BCSD) is widely used in various binary analysis tasks such as vulnerability search, malware detection, clone detection, and patch analysis. Recent studies have shown that the learning-based binary code embedding models perform better than the traditional feature-based approaches. However, previous studies have not delved deeply into the key factors that affect model performance. In this paper, we design extensive ablation studies to explore these influencing factors. The experimental results have provided us with many new insights. We have made innovations in both code representation and model selection: we propose a novel rich-semantic function representation technique to ensure the model captures the intricate nuances of binary code, and we introduce the first UniLM-based binary code embedding model, named UniASM, which includes two newly designed training tasks to learn representations of binary functions. The experimental results show that UniASM outperforms the state-of-the-art (SOTA) approaches on the evaluation datasets. The average scores of Recall@1 on cross-compilers, cross-optimization-levels, and cross-obfuscations have improved by 12.7 % , 8.5 % , and 22.3 % , respectively, compared to the best of the baseline methods. Besides, in the real-world task of known vulnerability search, UniASM outperforms all the current baselines.}, archive = {J_NEUCOM}, author = {Yeming Gu and Hui Shu and Fei Kang and Fan Hu}, doi = {10.1016/j.neucom.2025.129646}, journal = {Neurocomputing}, month = {5}, pages = {129646}, shortjournal = {Neurocomputing}, title = {UniASM: Binary code similarity detection without fine-tuning}, volume = {630}, year = {2025}, }
• (2025). Apollonius partitions based pursuit-evasion strategies via multi-agent reinforcement learning. NEUCOM, 630, 129643. (www)

  @article{ , abstract = {This paper considers a multi-agent pursuit-evasion game in which no less than two pursuers cooperate to capture a sensible evader. In contrast, the slower evader takes its strategy to extend life as long as possible. Since the classical Apollonius circle is effective in generating a capture strategy for multiple pursuers around a single evader scenario, we use Apollonius circle method to divide the entire state space into their respective dominant regions. The overlapping area of the Apollonius circles is assigned as the dominant region of the slower evader. In addition, we design a potential game model for this game, which guarantees the existence of the Nash equilibrium solution. In this paper, the evader’s strategy is derived using the dominant region determined by current global states while pursuers’ strategies are generated by the designed algorithm, where pursuers cooperate to shrink the evader’s dominant area to achieve capture during the motion. Finally, three simulations of different scenarios based on the QMIX algorithm demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Lei Xue and Qing Wang and Yongbao Wu and Xin Yuan and Jian Liu}, doi = {10.1016/j.neucom.2025.129643}, journal = {Neurocomputing}, month = {5}, pages = {129643}, shortjournal = {Neurocomputing}, title = {Apollonius partitions based pursuit-evasion strategies via multi-agent reinforcement learning}, volume = {630}, year = {2025}, }
• (2025). Small and dim target detection in infrared imagery: A review, current techniques and future directions. NEUCOM, 630, 129640. (www)

  @article{ , abstract = {While there has been significant progress in object detection using conventional image processing and machine learning algorithms, exploring small and dim target detection in the IR domain is a relatively new area of study. The majority of small and dim target detection methods are derived from conventional object detection algorithms, albeit with some alterations. The task of detecting small and dim targets in IR imagery is complex. The lack of distinct features, cluttered background with obscure details, and the variability of infrared signals due to thermodynamic fluctuations are the reasons why these targets often do not exhibit identifiable characteristics. The primary objective of this review is to highlight the progress made in this field. This is the first review in the field of small and dim target detection in infrared imagery, encompassing various methodologies ranging from conventional image processing to cutting-edge deep learning-based approaches. The authors have also introduced a taxonomy of such approaches. Both conventional image processing and deep learning approaches are further categorized into two subcategories: methodologies that utilize several frames for detection, and single-frame-based detection strategies. Our findings suggest that deep learning methods exhibit superior performance, particularly in cluttered environments, compared to conventional image processing methods for both of the aforementioned subcategories. In addition, a comprehensive compilation of various available datasets has also been provided. Furthermore, this review identifies the gaps and limitations in existing techniques, paving the way for future research and development in this area.}, archive = {J_NEUCOM}, author = {Nikhil Kumar and Pravendra Singh}, doi = {10.1016/j.neucom.2025.129640}, journal = {Neurocomputing}, month = {5}, pages = {129640}, shortjournal = {Neurocomputing}, title = {Small and dim target detection in infrared imagery: A review, current techniques and future directions}, volume = {630}, year = {2025}, }
• (2025). SRVC-LA: Sparse regularization of visual context and latent attention based model for video description. NEUCOM, 630, 129639. (www)

  @article{ , abstract = {Video description is one of the important generative tasks, which is to summarize the visual content and translate it into natural language. Nowadays, an increasing number of effective models have been developed for this task. Nevertheless, the visual and language features are combined and represented in a dense multi-modal feature space in popular works, making it easy for the model to become stuck in over fitting. This results in the model lacking sufficient generalization ability. A model with sparse regularization of visual context and latent attention (SRVC-LA) is proposed in this work. According to encoding padding sequence corresponding to the encoded video and then concatenation of visual contextual features, the visual regularization context is extracted for vision attention sparsity. Then it is attended by a latent attention mechanism, where the visual regularization context and the previous hidden state is combined and attended for multi-modal semantic alignment. Additionally, the visual and language features are combined with their respective latent attention features and fed to two branches for semantic compensation. Experiments on MSVD and MSR-VTT2016 datasets are conducted, and better performances are achieved compared to the baseline and other popular models, demonstrating the effectiveness and superiority of the proposed model.}, archive = {J_NEUCOM}, author = {Pengjie Tang and Jiayu Zhang and Hanli Wang and Yunlan Tan and Yun Yi}, doi = {10.1016/j.neucom.2025.129639}, journal = {Neurocomputing}, month = {5}, pages = {129639}, shortjournal = {Neurocomputing}, title = {SRVC-LA: Sparse regularization of visual context and latent attention based model for video description}, volume = {630}, year = {2025}, }
• (2025). CALMS: Constrained adaptive lasso with multi-directional signals for latent networks reconstruction. NEUCOM, 630, 129545. (www)

  @article{ , abstract = {The ubiquity of networks in the real world significantly influences dynamic behaviors, with empirical analysis often relying on known network structures to identify network effects. However, inferring latent network structures based solely on observed data remains a challenging inverse problem. To this end, we introduce a new method, Adaptive Lasso with Multi-directional Signals (ALMS), which transforms the network reconstruction problem into a high-dimensional estimation optimization process. ALMS shrinks the coefficients effectively to different directions by adaptive multi-directional penalties, which is able to reduce the estimation bias. Furthermore, we propose a constrained modification, Constrained Adaptive Lasso with Multi-directional Signals (CALMS) to improve both the accuracy and the interpretability of the estimated results. Theoretically, the oracle properties of both the ALMS estimator and the CALMS estimator are established. We demonstrate the excellent finite sample performances of both ALMS and CALMS via comprehensive simulation studies and empirical analysis involving the evolutionary ultimatum game, synchronization model and real experimental data.}, archive = {J_NEUCOM}, author = {Zhaoyu Xing and Huaiyu Tan and Wei Zhong and Lei Shi}, doi = {10.1016/j.neucom.2025.129545}, journal = {Neurocomputing}, month = {5}, pages = {129545}, shortjournal = {Neurocomputing}, title = {CALMS: Constrained adaptive lasso with multi-directional signals for latent networks reconstruction}, volume = {630}, year = {2025}, }
• (2025). Allowable delay set flexible fragmentation approach to passivity analysis of delayed neural networks. NEUCOM, 629, 129730. (www)

  @article{ , abstract = {This paper addresses the passivity issue of neural networks with a time-varying delay. We introduce an allowable delay set flexible fragmentation approach for constructing a novel Lyapunov–Krasovskii functional (LKF). Unlike some existing methods, this LKF is more flexible as it allows for different Lyapunov matrices in different allowable delay subsets. Based on the proposed LKF, and utilizing integral inequality approaches and zero equation techniques, several passivity criteria are derived for neural networks with a time-varying delay. Two numerical examples are finally provided to demonstrate the advantages of the proposed method.}, archive = {J_NEUCOM}, author = {Yun Chen and Chengda Lu and Xian-Ming Zhang}, doi = {10.1016/j.neucom.2025.129730}, journal = {Neurocomputing}, month = {5}, pages = {129730}, shortjournal = {Neurocomputing}, title = {Allowable delay set flexible fragmentation approach to passivity analysis of delayed neural networks}, volume = {629}, year = {2025}, }
• (2025). PLV-CSNet: Projected landweber variant unfolding network for image compressive sensing reconstruction. NEUCOM, 629, 129723. (www)

  @article{ , abstract = {Due to the powerful learning capability and fast processing speed of deep neural networks, a series of pure data-driven and deep unfolding networks for image reconstruction have emerged, achieving improved reconstruction quality. These reconstruction networks typically employ convolutional neural networks or residual neural networks to extract high-dimensional features of the dominant structure component. However, the edge and texture components in multi-dimensional features as well as the measurement residual generated at each iteration during the unfolding procedure are often neglected, which would affect the quality of image reconstruction. In this paper, a projected Landweber variant unfolding network (PLV-CSNet) is proposed for image compressive sensing reconstruction. A PLV algorithm is investigated and then unfolded to the PLV-Block, which consists of a thresholding module (TSM) and a progressive projecting module (PPM). The TSM utilizes the dense block to fuse multi-dimensional image features and the soft thresholding to eliminate image noise. The PPM combines the approximate message passing algorithm with deep neural networks to compute the projections of the approximation solution for images, as well as calculate the measurement residual generated during each iteration. Furthermore, a residual integration module (RIM) is designed to employ the measurement residuals to reconstruct the image residual which are then flexibly supplemented back into the reconstructed image. The effectiveness of PLV-CSNet is demonstrated in four standard benchmark datasets, and comparisons with classical image compressive sensing reconstruction networks show that our network could achieve higher reconstruction accuracy. Codes are available at: https://github.com/junp-hao/PLV-CSNet .}, archive = {J_NEUCOM}, author = {Junpeng Hao and Huang Bai and Xiumei Li and Marko Panic and Junmei Sun}, doi = {10.1016/j.neucom.2025.129723}, journal = {Neurocomputing}, month = {5}, pages = {129723}, shortjournal = {Neurocomputing}, title = {PLV-CSNet: Projected landweber variant unfolding network for image compressive sensing reconstruction}, volume = {629}, year = {2025}, }
• (2025). Trusted cross-view completion for incomplete multi-view classification. NEUCOM, 629, 129722. (www)

  @article{ , abstract = {In real-world scenarios, missing views is common due to the complexity of data collection. Therefore, it is inevitable to classify incomplete multi-view data. Although substantial progress has been achieved, there are still two challenging problems with incomplete multi-view classification: (1) Simply ignoring these missing views is often ineffective, especially under high missing rates, which can lead to incomplete analysis and unreliable results. (2) Most existing multi-view classification models primarily focus on maximizing consistency between different views. However, neglecting specific-view information may lead to decreased performance. To solve the above problems, we propose a novel framework called Trusted Cross-View Completion (TCVC) for incomplete multi-view classification. Specifically, TCVC consists of three modules: Cross-view Feature Learning Module (CVFL), Imputation Module (IM) and Trusted Fusion Module (TFM). First, CVFL mines specific-view information to obtain cross-view reconstruction features. Then, IM restores the missing view by fusing cross-view reconstruction features with weights, guided by uncertainty-aware information. This information is the quality assessment of the cross-view reconstruction features in TFM. Moreover, the recovered views are supervised by cross-view neighborhood-aware. Finally, TFM effectively fuses complete data to generate trusted classification predictions. Extensive experiments show that our method is effective and robust.}, archive = {J_NEUCOM}, author = {Liping Zhou and Shiyun Chen and Peihuan Song and Qinghai Zheng and Yuanlong Yu}, doi = {10.1016/j.neucom.2025.129722}, journal = {Neurocomputing}, month = {5}, pages = {129722}, shortjournal = {Neurocomputing}, title = {Trusted cross-view completion for incomplete multi-view classification}, volume = {629}, year = {2025}, }
• (2025). SFIMCO: Scalable fair influence maximization based on overlapping communities and optimization algorithms. NEUCOM, 629, 129687. (www)

  @article{ , abstract = {Influence Maximization (IM) is a fundamental problem in social networks, aiming to identify a small set of seed nodes that maximize the spread of influence. Fair Influence Maximization (FIM) extends this concept by incorporating fairness criteria, ensuring equitable influence distribution among diverse groups. Welfare fairness, a prominent fairness measure, strives to balance influence spread and fairness levels. However, optimizing welfare fairness objectives is computationally challenging and often limited to small-scale networks. While community structures are widely used to improve FIM performance, the overlapping nature of communities has been largely overlooked in the literature. This oversight can significantly impact the trade-off between computational complexity and fairness. To address these gaps, this study introduces Scalable FIM based on overlapping Communities and Optimization algorithms (SFIMCO). SFIMCO considers the influence potential of nodes within communities by distinguishing between overlapping and non-overlapping nodes. It employs an efficient optimization algorithm to select the seed nodes, ensuring a scalable solution. Meanwhile, the objective function for welfare fairness in SFIMCO is formulated as maximizing the weighted exponential welfare across all overlapping communities, ensuring scalable approximation of fair influence. Experimental evaluations on real-world datasets demonstrate that SFIMCO outperforms state-of-the-art methods in both scalability and fairness. Notably, it achieves a 2.54 % improvement in average influence spread compared to the best existing approach, highlighting its effectiveness in addressing the trade-offs between influence spread, fairness, and computational efficiency.}, archive = {J_NEUCOM}, author = {Lijuan Chen and Amin Rezaeipanah}, doi = {10.1016/j.neucom.2025.129687}, journal = {Neurocomputing}, month = {5}, pages = {129687}, shortjournal = {Neurocomputing}, title = {SFIMCO: Scalable fair influence maximization based on overlapping communities and optimization algorithms}, volume = {629}, year = {2025}, }
• (2025). IAFMVS: Iterative depth estimation with adaptive features for multi-view stereo. NEUCOM, 629, 129682. (www)

  @article{ , abstract = {Nowadays, few approaches exist for optimizing feature extraction in multi-view stereo networks, and refined feature map extraction is still a critical task for high-resolution 3D reconstruction. This paper proposes an efficient multi-view stereo processing method based on IterMVS to extract fine features and hold lightweight without consuming excessive memory and running time. Firstly, we retain the GRU depth probability estimation module to maintain the lightweight. Meanwhile, an enhancing feature extraction module based on hierarchical fusion is incorporated. This module utilizes deformable convolutions to capture three different-scale feature maps, then fused at multiple levels to strengthen per-pixel information. This strategy enables the network to handle more blurry boundaries and texture-less regions. Secondly, the attention-based feature-matching mechanism proposed during the initialization process learns image interactions by focusing on crucial information and improves the feature expressiveness and reconstruction accuracy. The evaluation based on the DTU dataset shows a 9% improvement over the IterMVS and surpasses several state-of-the-art methods. Experimental results on the Tanks & Temples benchmark dataset validate that our algorithm outperforms the IterMVS model.}, archive = {J_NEUCOM}, author = {Guyu Zhao and Huyixin Wei and Hongdou He}, doi = {10.1016/j.neucom.2025.129682}, journal = {Neurocomputing}, month = {5}, pages = {129682}, shortjournal = {Neurocomputing}, title = {IAFMVS: Iterative depth estimation with adaptive features for multi-view stereo}, volume = {629}, year = {2025}, }
• (2025). Deep differentiable symbolic regression neural network. NEUCOM, 629, 129671. (www)

  @article{ , abstract = {Deep learning (DL) methods have recently been applied to the symbolic regression (SR) challenge. However, for SR problems, DL methods are non-differentiable due to SR’s discrete search space, resulting in longer training time and challenging convergence. To address this, we introduce a new approach: Deep Differentiable Symbolic Regression Neural Network (DDSR-NN). Within DDSR-NN, every layer represents a continuous distribution of mathematical operators, and the connections between layers represent the coefficient between these operators. So, DDSR-NN represents the distribution of mathematical expression. To ensure efficiency and avoid overly complex results, DDSR-NN utilizes gate masks to prune results and rule masks to reduce search space. Furthermore, we developed a hybrid training algorithm, merging the gradient descent method with a risk-seeking policy gradient method. This ensures that DDSR-NN swiftly converges to the distribution of mathematical expressions with minimal loss. Experiments conducted on 35 distinct datasets revealed that DDSR-NN outperforms 23 baseline algorithms in striking a balance between accuracy and interpretability. Notably, when compared to the three neural network (NN) baselines, DDSR-NN consistently yielded more accurate outcomes.}, archive = {J_NEUCOM}, author = {Qiang Lu and Yuanzhen Luo and Haoyang Li and Jake Luo and Zhiguang Wang}, doi = {10.1016/j.neucom.2025.129671}, journal = {Neurocomputing}, month = {5}, pages = {129671}, shortjournal = {Neurocomputing}, title = {Deep differentiable symbolic regression neural network}, volume = {629}, year = {2025}, }
• (2025). Brain-inspired reward broadcasting: Brain learning mechanism guides learning of spiking neural network. NEUCOM, 629, 129664. (www)

  @article{ , abstract = {The secret of the brain’s efficient learning lies in its rich set of learning mechanisms, neuronal encoding and plasticity principles. However, the combination of learning mechanisms and plasticity principles has not yet been incorporated into spiking neural networks (SNN). Existing brain-inspired learning algorithms still have problems such as being unable to achieve global credit allocation and causing catastrophic forgetting in incremental learning. Drawing inspiration from the brain’s reward learning mechanism, this paper employs reinforcement learning (RL) to model reward mechanisms and achieve synaptic plasticity and presents a novel SNN learning algorithm called Brain-Inspired Reward Broadcasting Algorithm (BIRB). BIRB treats synapses as agents and guides agent actions to achieve plasticity through RL; it is biologically plausible to model plasticity in this way as it achieves global credit assignment by broadcasting reward signals, which aligns with the reward mechanism process. BIRB incorporates a innovative learning approach known as two-step learning, which effectively addresses the issue of catastrophic forgetting. This paper also mathematically derives the plasticity process and learning principles of BIRB. Moreover, experimental results demonstrate the exceptional performance of BIRB across multiple datasets, thereby highlighting its pivotal role in guiding synaptic weight changes and facilitating generalization.}, archive = {J_NEUCOM}, author = {Miao Wang and Gangyi Ding and Yunlin Lei and Yu Zhang and Lanyu Gao and Xu Yang}, doi = {10.1016/j.neucom.2025.129664}, journal = {Neurocomputing}, month = {5}, pages = {129664}, shortjournal = {Neurocomputing}, title = {Brain-inspired reward broadcasting: Brain learning mechanism guides learning of spiking neural network}, volume = {629}, year = {2025}, }
• (2025). Bi-branch network with region-wise dynamic convolution for image inpainting. NEUCOM, 629, 129663. (www)

  @article{ , abstract = {For image inpainting methods based on convolutional neural networks, once training is completed, the parameters of convolution kernels remain fixed, making it intractable to accommodate variations in the input. Additionally, these methods treat the missing and complete regions equally, leading to difficulties in accommodating the differences in feature distributions between regions. To address the aforementioned challenges, we propose a Bi-branch Network with Region-wise Dynamic Convolution (BNRDC), which consists of the upper and lower branches. The upper branch is cored with Dynamic Convolution Kernel Prediction (DCKP) module, which aims to predict dynamic convolution kernels. By generating several attention weights used for combining multiple convolution kernels, DCKP can dynamically adjust kernel parameters, responding to variations in the input. The lower branch is cored with Region Dynamic Convolution Attention (RDCA) module. RDCA performs region-wise convolutions with the help of dynamic convolution kernels to accommodate the difference in feature distributions. Furthermore, we design a Multi-Scale Feature Fusion (MSFF) module between the two branches to provide rich features for predicting dynamic convolution kernels. It fuses multi-scale features, from spatial and scale perspectives, based on dilated convolutions and a feature pyramid, respectively. Extensive experiments on Places2, CelebA and Dunhuang Challenge datasets demonstrate that the proposed method outperforms the state-of-the-art baselines. Our method can dynamically capture variations in the input and effectively align the features of missing regions with those of ground-truth images.}, archive = {J_NEUCOM}, author = {Guangyuan Huang and Rong Huang and Shubo Zhou and Zhijie Wang}, doi = {10.1016/j.neucom.2025.129663}, journal = {Neurocomputing}, month = {5}, pages = {129663}, shortjournal = {Neurocomputing}, title = {Bi-branch network with region-wise dynamic convolution for image inpainting}, volume = {629}, year = {2025}, }
• (2025). CoAtUNet: A symmetric encoder-decoder with hybrid transformers for semantic segmentation of breast ultrasound images. NEUCOM, 629, 129660. (www)

  @article{ , abstract = {The CoAtNet hybrid deep model yields state-of-the-art performance in image analysis by integrating the strengths of convolutional neural networks (CNNs) along with self-attention mechanisms, combining local feature extraction with global dependency modeling. In this paper, we introduce CoAtUNet, a novel semantic segmentation network that utilizes a symmetric encoder–decoder architecture in which CoAtNet is applied to both branches. CoAtUNet is specifically designed for tumor segmentation in breast ultrasound (BUS) images, aiming to enhance segmentation accuracy and robustness. By leveraging CNNs for detailed local feature extraction and Transformers for capturing long-range dependencies, CoAtUNet addresses the unique challenges of breast ultrasound images, such as noise, low contrast, and the variability in tumor shapes and sizes. This hybrid approach enables CoAtUNet to outperform traditional asymmetric configurations and other advanced segmentation models. Evaluations conducted on two public breast ultrasound datasets, UDIAT and BUSI, demonstrate the superior performance of CoAtUNet, achieving Dice coefficients of 79.12% and 76.21%, respectively. These results not only surpass existing tumor segmentation methods, but also highlight the potential of CoAtUNet to assist clinicians in precise tumor detection, facilitating more accurate and reliable diagnoses. The CoAtUNet implementation is publicly available at https://github.com/NadeemAlkilani .}, archive = {J_NEUCOM}, author = {Nadeem Zaidkilani and Miguel Angel Garcia and Domenec Puig}, doi = {10.1016/j.neucom.2025.129660}, journal = {Neurocomputing}, month = {5}, pages = {129660}, shortjournal = {Neurocomputing}, title = {CoAtUNet: A symmetric encoder-decoder with hybrid transformers for semantic segmentation of breast ultrasound images}, volume = {629}, year = {2025}, }
• (2025). A distance correlation-based approach to characterize the effectiveness of recurrent neural networks for time series forecasting. NEUCOM, 629, 129641. (www)

  @article{ , abstract = {Time series forecasting has received a lot of attention, with recurrent neural networks (RNNs) being one of the widely used models due to their ability to handle sequential data. Previous studies on RNN time series forecasting, however, show inconsistent outcomes and offer few explanations for performance variations among the datasets. In this paper, we provide an approach to link time series characteristics with RNN components via the versatile metric of distance correlation. This metric allows us to examine the information flow through the RNN activation layers to be able to interpret and explain their performance. We empirically show that the RNN activation layers learn the lag structures of time series well. However, they gradually lose this information over the span of a few consecutive layers, thereby worsening the forecast quality for series with large lag structures. We also show that the activation layers cannot adequately model moving average and heteroskedastic time series processes. Last, we generate heatmaps for visual comparisons of the activation layers for different choices of the network hyperparameters to identify which of them affect the forecast performance. Our findings can, therefore, aid practitioners in assessing the effectiveness of RNNs for given time series data without actually training and evaluating the networks.}, archive = {J_NEUCOM}, author = {Christopher Salazar and Ashis G. Banerjee}, doi = {10.1016/j.neucom.2025.129641}, journal = {Neurocomputing}, month = {5}, pages = {129641}, shortjournal = {Neurocomputing}, title = {A distance correlation-based approach to characterize the effectiveness of recurrent neural networks for time series forecasting}, volume = {629}, year = {2025}, }
• (2025). Multi-modal spiking tensor regression network for audio-visual zero-shot learning. NEUCOM, 629, 129636. (www)

  @article{ , abstract = {Recently, convolutional neural networks have got significant attention, particularly in the field of audio-visual zero-shot learning. It can accurately perceive and capture local features, which allows the model to effectively obtain the corresponding attributes. The original multilinear structure is disrupted when the tensor is flattened as it passes through the fully connected layers. Inspired by this, we introduce a multi-modal spiking tensor regression network (MSTR). MSTR incorporates tensor regression networks with tensor contractions and spiking neural networks featuring threshold adjustments, thus effectively handling temporal and spatial information. It facilitates fine-grained feature extraction while retaining high-dimensional spatial information. Specifically, we use Spiking Neural Networks (SNN) to encode temporal features, and Tensor Regression Networks (TRN) to encode spatial features. Our proposed Temporal-Spatial-Semantic Fusion block combines temporal, spatial, and semantic features for each modality. Finally, the fused audio and visual features pass through a series of cross-modal transformers, further exploring the inner relationship between each modalities. Experimental results on three benchmark datasets, ActivityNet, VGGSound, and UCF, demonstrate that MSTR demonstrates superiorities compared with state-of-the-art models, with significant improvements in harmonic mean (HM) scores on three datasets of 6.0%, 6.8%, and 2.2%, respectively. The code and pre-trained models are available at https://github.com/xia-zhe/MSTR .}, archive = {J_NEUCOM}, author = {Zhe Yang and Wenrui Li and Jinxiu Hou and Guanghui Cheng}, doi = {10.1016/j.neucom.2025.129636}, journal = {Neurocomputing}, month = {5}, pages = {129636}, shortjournal = {Neurocomputing}, title = {Multi-modal spiking tensor regression network for audio-visual zero-shot learning}, volume = {629}, year = {2025}, }
• (2025). Comparing synchronous and asynchronous UWB time-based localization systems for autonomous mobile robots. NEUCOM, 629, 129622. (www)

  @article{ , abstract = {Autonomous Mobile Robots (AMRs) are essential for automating internal transportation in Industry 4.0, relying on visual-based positioning systems to generate maps and compare their location with mapped reference obstacles. However, enhancing positioning accuracy is crucial for demanding industrial tasks. Combining vision-based systems with Ultra-Wideband (UWB) time-based positioning systems has gained research interest. In this sense, UWB based localization systems are commonly categorized into synchronous and asynchronous configurations. Synchronous systems, such as the Time of Arrival (TOA) and the Time Difference of Arrival (TDOA) systems, require clock synchronization but reduce signal path lengths. In contrast, asynchronous systems, such as Asynchronous Time Difference of Arrival (A-TDOA) and Single-Sided Two-Way Ranging (SS-TWR), eliminate synchronization needs but may involve longer signal paths. To address these differences, in this paper, we propose a methodology to fairly compare these configurations. The approach includes characterizing the error sources in each system and analyzing the optimal spatial arrangement of sensors under incremental time reply conditions. Results reveal the superiority of asynchronous configurations in low sensor density deployments, achieving up to a 54.8% reduction in localization error compared to synchronous systems. In addition, as the number of deployed nodes decreases, synchronous systems show up to a 28.9% increase in localization error to the analysis performed, whereas the asynchronous systems achieved a 24.8% increase for the A-TDOA, and only a 9.2% increase in the case of the SS-TWR method. These findings prompt further investigation into the suitability of synchronous positioning for minimizing errors in large-scale industrial applications.}, archive = {J_NEUCOM}, author = {Rubén Ferrero-Guillén and Javier Díez-González and Alberto Martínez-Gutiérrez and Hilde Perez and Joaquín Torres-Sospedra and Adriano Moreira}, doi = {10.1016/j.neucom.2025.129622}, journal = {Neurocomputing}, month = {5}, pages = {129622}, shortjournal = {Neurocomputing}, title = {Comparing synchronous and asynchronous UWB time-based localization systems for autonomous mobile robots}, volume = {629}, year = {2025}, }
• (2025). Graph transformer-based heterogeneous graph neural networks enhanced by multiple meta-path adjacency matrices decomposition. NEUCOM, 629, 129604. (www)

  @article{ , abstract = {Heterogeneous graph neural networks (HGNNs) are adept at processing data within multi-relational heterogeneous networks. Nonetheless, contemporary HGNNs that are relation-aware or utilize meta-path-based architectures struggle with recognizing long-range dependencies. Moreover, an over-extension in model depth frequently results in over-smoothing. To address these challenges, this paper proposes the Multiple meta-path Structural representation-aware Heterogeneous Graph Transformer (MSHGT). Firstly, the model constructs a latent fully-connected attention graph using node features, allowing all nodes to aggregate information from neighbor nodes of arbitrary hops through weighted aggregation via the attention graph. This endows the model with global perception and the ability to capture long-range dependency information, mitigating the issue of over-smoothing in node representations. Furthermore, to enhance MSHGT’s structure perception capability and further bolster its ability to capture heterogeneous connectivity patterns, a structural representation encoding method based on multiple meta-path adjacency matrices decomposition is proposed. This method processes adjacency matrices of different meta-path categories through singular value decomposition, extracting structure correlation information between nodes from various semantic perspectives and encapsulating it into structural representations that span multiple semantic dimensions. Subsequently, these structural representations and node attribute features are simultaneously integrated into the Transformer module, prompting the model to perform message-passing from both attribute and structure dimensions. Comprehensive experiments conducted on six heterogeneous graph datasets, covering node classification and link prediction tasks, demonstrate that MSHGT outperforms baseline methods significantly, confirming its efficacy in addressing the complexities of heterogeneous networks.}, archive = {J_NEUCOM}, author = {Shibin Li and Jun Gong and Shengnan Ke and Shengjun Tang}, doi = {10.1016/j.neucom.2025.129604}, journal = {Neurocomputing}, month = {5}, pages = {129604}, shortjournal = {Neurocomputing}, title = {Graph transformer-based heterogeneous graph neural networks enhanced by multiple meta-path adjacency matrices decomposition}, volume = {629}, year = {2025}, }
• (2025). The roles of internal dynamics and proprioceptive feedback in motor cortex during movement execution. NEUCOM, 629, 129551. (www)

  @article{ , abstract = {The motor cortex controls arm movements by sending commands to lower motor centers. The synaptic connections within the motor cortex (internal dynamics) are vital to generate motor commands during movement execution. However, recent studies suggest that proprioception also contribute to motor cortex processing. To investigate the contributions of internal dynamics and proprioceptive feedback to movement execution, we built a recurrent neural network model of the motor cortex; the model receives proprioceptive feedback from a virtual arm performing a delayed-reach task. We found that both internal dynamics and proprioceptive feedback contribute to the resemblance to the real motor cortex data. We then dissected their contributions by disrupting them separately. Internal dynamics dominate the generation of neural population activity, while proprioceptive feedback modulates neural responses and controls movement deceleration. Additionally, proprioceptive feedback improves the network’s robustness in noisy initial conditions. We further investigated the relative importance of the components in proprioceptive feedback and found that hand velocity is most important. Our findings could inform the development of neural prosthetics that can replicate the sensorimotor control of the biological system.}, archive = {J_NEUCOM}, author = {Hongru Jiang and Xiangdong Bu and Zhiyan Zheng and Huajin Tang and Xiaochuan Pan and Yao Chen}, doi = {10.1016/j.neucom.2025.129551}, journal = {Neurocomputing}, month = {5}, pages = {129551}, shortjournal = {Neurocomputing}, title = {The roles of internal dynamics and proprioceptive feedback in motor cortex during movement execution}, volume = {629}, year = {2025}, }
• (2025). Lightweight oriented object detection with dynamic smooth feature fusion network. NEUCOM, 628, 129725. (www)

  @article{ , abstract = {In remote sensing visual tasks, detection of small objects remains a challenge due to low resolution, complexity and scale variations. In addition, many of these detection tasks need to be deployed to front-end low-resource devices on Unmanned Aerial Vehicle (UAV), which requires that detection method should be lightweight. This paper presents a lightweight oriented object detection method for small objects in remote sensing images, which is named Dynamic Smooth Feature Fusion Network (DSFF-Net). DynamicConv (DC) module is designed to enhance the depth of the network by stacking and continuously fusing small modules to capture contextual information of small objects at lower computational cost. Smooth Attention (SA) module is developed to incorporate attention mechanisms along spatial height and width directions of feature maps. The SA module enhances spatial feature extraction by generating attention maps that emphasize the selection of critical object features while suppressing background noise. It is worth mentioning that the proposed DC and SA modules can be integrated into many classical object detection frameworks and enhance the detection performance of remote sensing small objects consistently. Extensive experiments verify the effectiveness of the proposed DSFF-Net.}, archive = {J_NEUCOM}, author = {Iftikhar Ahmad and Wei Lu and Si-Bao Chen and Jin Tang and Bin Luo}, doi = {10.1016/j.neucom.2025.129725}, journal = {Neurocomputing}, month = {5}, pages = {129725}, shortjournal = {Neurocomputing}, title = {Lightweight oriented object detection with dynamic smooth feature fusion network}, volume = {628}, year = {2025}, }
• (2025). A spatio-frequency cross fusion model for deepfake detection and segmentation. NEUCOM, 628, 129683. (www)

  @article{ , abstract = {The misuse of deepfake technology has led to a crisis of trust in society, with a growing expectation for more reliable evidence to support deepfake detection results. To address this need, we propose a novel spatio-frequency cross fusion (SFCF) model for deepfake detection and segmentation tasks. This model leverages the deep fusion of spatial and frequency features to enhance the extraction of forgery traces, thereby achieving more accurate deepfake detection and segmentation. Our model includes two innovative modules: the multi-scale high-frequency (MSHF) module and the feature cross fusion (FCF) module. The MSHF module utilizes the stationary wavelet transform (SWT) to extract multi-scale high-frequency features from images. The FCF module employs an attention mechanism to facilitate information exchange between the spatial stream and the frequency stream. Comprehensive evaluations on multiple datasets demonstrate that SFCF achieves excellent results in both deepfake detection and segmentation tasks, outperforming existing segmentation methods in terms of visual quality and accuracy in predicting manipulated regions.}, archive = {J_NEUCOM}, author = {Junshuai Zheng and Yichao Zhou and Ning Zhang and Xiyuan Hu and Kaiwen Xu and Dongyang Gao and Zhenmin Tang}, doi = {10.1016/j.neucom.2025.129683}, journal = {Neurocomputing}, month = {5}, pages = {129683}, shortjournal = {Neurocomputing}, title = {A spatio-frequency cross fusion model for deepfake detection and segmentation}, volume = {628}, year = {2025}, }
• (2025). Dynamic event-triggering adaptive dynamic programming for robust stabilization of partially unknown nonlinear systems. NEUCOM, 628, 129673. (www)

  @article{ , abstract = {In this paper, a new dynamic event-triggering (DET) mechanism based on adaptive dynamic programming (ADP) is developed to deal with the robust control problem of partially unknown uncertain systems. First, this paper completes the transition from the robust control problem to the optimal control problem by designing a nominal system. Meanwhile, the use of integral reinforcement learning (IRL) eliminates the need for prior knowledge of drift dynamics. Then, to improve resource utilization, a static event-triggering (SET) scheme is designed. Subsequently, a DET scheme is developed on the basis of SET to further improve resource utilization. It is proven that the developed DET controller guarantees the robustness of the partially unknown uncertain system. The neural network (NN) weight estimation errors are uniformly ultimately bounded (UUB) while the Zeno behavior is successfully avoided. Finally, an experiment is provided to demonstrate that the proposed DET algorithm has the fewest triggering samples while guaranteeing robustness.}, archive = {J_NEUCOM}, author = {Yishen Hong and Xue Shan and Derong Liu and Yonghua Wang}, doi = {10.1016/j.neucom.2025.129673}, journal = {Neurocomputing}, month = {5}, pages = {129673}, shortjournal = {Neurocomputing}, title = {Dynamic event-triggering adaptive dynamic programming for robust stabilization of partially unknown nonlinear systems}, volume = {628}, year = {2025}, }
• (2025). Lightweight expression recognition combined attention fusion network with hybrid knowledge distillation for occluded e-learner facial images. NEUCOM, 628, 129656. (www)

  @article{ , abstract = {In online education, it is crucial to monitor the students’ learning status timely and accurately. Facial expression recognition serves as the main tool for assessing their engagement levels. Existing algorithms still have some issues when directly applied in online learning. These issues primarily involve the loss of facial features due to occlusion, which directly affects the accuracy of expression recognition and at the same time, the expression recognition models require a large number of parameters and significant computational power, making them difficult to deploy and apply effectively on mobile devices with limited hardware resources. We aims to address the two issues by a two-stage framework: training process of an occluded facial expression recognition model, and the compression process of the occlusion facial expression recognition. To be more specific, in the first stage, we propose an occlusion facial expression recognition model based on attention fusion (AFNet), which adopts a multi-branch spatial attention network, extracts local facial features, automatically perceives facial occluded regions, reduces the weight of the occluded areas, and enhance robustness to occlusion by combining with a randomly masked channel network. Meanwhile, a feature pyramid network is introduced to extract global multi-scale features. In the second stage, we propose a hybrid model compression algorithm based on multi-layer knowledge distillation (MKD). We introduce a spatial attention network to focus on the important knowledge, reducing the information loss during knowledge distillation. Experimental results on five datasets show that the AFNet and MKD outperformed the baseline.}, archive = {J_NEUCOM}, author = {Yan Chen and Kexuan Li and Feng Tian and Ganglin Wei and Morteza Seberi}, doi = {10.1016/j.neucom.2025.129656}, journal = {Neurocomputing}, month = {5}, pages = {129656}, shortjournal = {Neurocomputing}, title = {Lightweight expression recognition combined attention fusion network with hybrid knowledge distillation for occluded e-learner facial images}, volume = {628}, year = {2025}, }
• (2025). Multilevel diverse feature aggregation network for salient object detection. NEUCOM, 628, 129648. (www)

  @article{ , abstract = {The task of Salient Object Detection (SOD) based on deep learning focuses on comprehending the contrast relationship between the object's inherent state and the current scene, thereby identifying salient objects. In recent years, the majority of research explore the aggregation of multilevel features to achieve improvements in detection performance. However, multilevel features are often undiscriminatingly aggregated, neglecting the differentiation among these features. In addition, there is uncertainty in predicting salient objects, which leads to confusion between salient and non-salient pixels. In this paper, we propose a novel network called the Multilevel Diverse Feature Aggregation Network (MDFANet). The network achieves optimization of differentiation among multilevel features and addresses the uncertainty in predictions, thereby effectively highlighting clear and complete salient objects while suppressing interference from non-salient regions. This results in outstanding performance in complex scenarios. The network adopts an encoder-decoder structure, with ResNet-50 as the backbone for extracting multilevel features in the encoder. Meanwhile, five modules are designed in the decoder to transform these multilevel feature representations into accurate and clear saliency maps. The Diverse Feature Activation (DFA) module is designed to capture saliency information and enhance the network's ability to recognize salient objects in multiple directions and channels, thus obtaining more structural and detailed information about salient objects. The Context Attention (CA) module is used to obtain comprehensive contextual semantic information, and thus learn the semantic correlation between different salient objects. The Edge Assist (EA) module is used to refine and supplement the boundary information of salient objects, thereby improving the network's understanding of pixels around the contours of salient objects. The Multilevel Feature Fusion (MFF) module effectively utilizes multilevel saliency features, reduces differences between features, aggregates contextual semantic information and diverse saliency information to generate more representative saliency features, effectively overcomes the problem of confusion between salient and non-salient pixels, and reduces uncertainty in prediction results. The Self Improve (SI) module refines the aggregated features and identifies feature spaces and channels with high responsiveness to salient objects to generate more accurate and clearer saliency maps. Experimental results show that our proposed approach performs favorably against 13 state-of-the-art representative methods on six benchmark datasets.}, archive = {J_NEUCOM}, author = {Qiaoning Yang and Jiahao Zheng and Juan Chen}, doi = {10.1016/j.neucom.2025.129648}, journal = {Neurocomputing}, month = {5}, pages = {129648}, shortjournal = {Neurocomputing}, title = {Multilevel diverse feature aggregation network for salient object detection}, volume = {628}, year = {2025}, }
• (2025). ACF-r+: An asymmetry-sensitive method for image-text retrieval enhanced by cross-modal fusion and re-ranking based on contrastive learning. NEUCOM, 628, 129642. (www)

  @article{ , abstract = {The task of multi-modal retrieval between the image and text modality is to find pertinent information from a designated image or textual corpus. The principal challenge lies in the integration of multi-modal representations and the discernment of fine-grained distinctions among various modalities, with the goal of identifying analogous content and filtering out irrelevant information. However, the existing methods predominantly concentrated on the unification of semantic representations and the alignment of concepts across multiple modalities, while the subtle differences between modalities have not been extensively investigated, which resulted in the issue of information asymmetry. To address this problem, a novel asymmetry-sensitive method based on contrastive learning was proposed in this paper, for achieving unified multi-modal semantic representations while being able to distinguish fine-grained semantic differences by generating corresponding positive and negative samples tailored to different asymmetry types. Then, we introduced a hierarchical cross-modal fusion method to integrate semantic features at both the global and local levels by the multi-modal attention mechanism, for serving to enhance the alignment of conceptual representations across different modalities. On this basis, a re-ranking strategy was proposed by utilizing cross-modal bidirectional retrieval information. This algorithm exploits the K-nearest neighbors of the initial retrieval results to perform a reverse search and combines bidirectional retrieval information, for further improving the performance of image-text retrieval tasks. Extensive experiments conducted on the MSCOCO and Flickr30K datasets, verified the effectiveness of the proposed method. Code is publicly available at https://github.com/Gzy1112/ACF-R .}, archive = {J_NEUCOM}, author = {Ziyu Gong and Yihua Huang and Chunhua Yu and Peng Dai and Xing Ge and Yiming Shen and Yafei Liu}, doi = {10.1016/j.neucom.2025.129642}, journal = {Neurocomputing}, month = {5}, pages = {129642}, shortjournal = {Neurocomputing}, title = {ACF-r+: An asymmetry-sensitive method for image-text retrieval enhanced by cross-modal fusion and re-ranking based on contrastive learning}, volume = {628}, year = {2025}, }
• (2025). A comprehensive overview of topic modeling: Techniques, applications and challenges. NEUCOM, 628, 129638. (www)

  @article{ , abstract = {The growing need of gaining valuable information from large collections of text data has attracted a considerable interest from researchers in the field of natural language processing (NLP) and machine learning. Topic modeling is an unsupervised NLP technique that aims to extract hidden themes within a corpus of textual documents. This paper provides a thorough and comprehensive review of topic modeling techniques from classical methods such as latent sematic analysis to most cutting-edge neural approaches and transformer-based methods. We also introduce topic modeling applications in various domains such as text classification, bioinformatics and sentiment analysis. Real-world case studies are examined to show the versatility of topic modeling in discovering hidden structures within text and unstructured data to facilitate the process of knowledge discovery. The authors also address current challenges facing topic modeling such as interpretability and scalability, and discuss future research directions. This review aims to provide both academic researchers and practitioners with a good understanding of topic modeling techniques, applications, datasets, and research trends.}, archive = {J_NEUCOM}, author = {Mustapha Hankar and Mohammed Kasri and Abderrahim Beni-Hssane}, doi = {10.1016/j.neucom.2025.129638}, journal = {Neurocomputing}, month = {5}, pages = {129638}, shortjournal = {Neurocomputing}, title = {A comprehensive overview of topic modeling: Techniques, applications and challenges}, volume = {628}, year = {2025}, }
• (2025). A brief survey of deep learning-based models for CircRNA-protein binding sites prediction. NEUCOM, 628, 129637. (www)

  @article{ , abstract = {CircRNAs are a particular single-stranded, circular structure and “non-coding” RNA molecules, with various biological functions. Existing studies have demonstrated the fundamental role of circRNAs in gene expression regulation and their significant involvement in the development of diverse complex diseases. Predicting the protein binding sites in circRNA can aid in comprehending the regulation mechanism involved in circRNA-protein binding during gene expression and facilitate the investigation of potential diagnosis and treatment strategies for complex diseases. This review begins by introducing the concept and functions of circRNAs, as well as their involvement in gene expression regulation. Then, some critical and publicly accessible databases about circRNA annotation, protein annotation, circRNA-protein binding were listed. Next, we present a brief introduction to the computational model for predicting circRNA-protein binding, followed by model performance comparison and suggestions for non-computer science experts on model selection. Finally, we examine the problems, limitations, and advantages of computational models and explore the further direction of circRNA-protein prediction, such as developing new and complex computational models, introducing complex biological sequence encoding schemes, and integrating additional biological data related to circRNA-protein binding.}, archive = {J_NEUCOM}, author = {Zhen Shen and Lin Yuan and Wenzheng Bao and Siguo Wang and Qinhu Zhang and De-Shuang Huang}, doi = {10.1016/j.neucom.2025.129637}, journal = {Neurocomputing}, month = {5}, pages = {129637}, shortjournal = {Neurocomputing}, title = {A brief survey of deep learning-based models for CircRNA-protein binding sites prediction}, volume = {628}, year = {2025}, }
• (2025). Dynamic subtask representation and assignment in cooperative multi-agent tasks. NEUCOM, 628, 129535. (www)

  @article{ , abstract = {In cooperative multi-agent reinforcement learning, achieving scalability is a critical objective, often pursued through parameter sharing. However, parameter sharing often leads to homogeneous agent behaviors, thereby reducing strategic diversity. Task decomposition emerges as a viable solution to achieve both scalability and diversity in multi-agent systems, but existing methods struggle with intricate environments due to reliance on prior knowledge or ignoring agent-specific characteristics. To bridge this gap, we introduce a novel framework for learning Dynamic Subtask Representation and Assignment (DSRA) in cooperative multi-agent tasks. We first leverage a variational autoencoder with a multivariate Gaussian distribution, enabling agents to generate dynamic and distinct subtask representations. After that, we incorporate an ability encoder to optimize the subtask assignment process in alignment with each agent’s abilities. To encourage diversity while sharing experiences, we further empower agents to develop individual policies for each subtask, promoting specialization and encouraging knowledge sharing, especially among agents engaged in similar tasks. Empirical results on Level-based Foraging, Predator–Prey, and Starcraft II micromanagement challenges demonstrate that DSRA significantly improves learning performance compared to existing MARL methods.}, archive = {J_NEUCOM}, author = {Chenlong You and Yingbo Wu and Junpeng Cai and Qi Luo and Yanbing Zhou}, doi = {10.1016/j.neucom.2025.129535}, journal = {Neurocomputing}, month = {5}, pages = {129535}, shortjournal = {Neurocomputing}, title = {Dynamic subtask representation and assignment in cooperative multi-agent tasks}, volume = {628}, year = {2025}, }
• (2025). A novel multi-level hierarchy optimization algorithm for pipeline inner detector speed control. NEUCOM, 627, 129715. (www)

  @article{ , abstract = {This paper proposes a novel nature-inspired algorithm called Multi-Level Hierarchy Optimization (MLHO) for solving optimization problems over continuous space. The MLHO algorithm is inspired by the hierarchy of nature, especially the hierarchy of biological populations. The entire algorithm structure is divided into four levels for iterative optimization, and the work of each level is global direction guidance, optimization-seeking task allocation, local optimal exploration, and broad domain exploration. Differential variation strategy and dynamic inertia factor are also designed to solve the problem of decreasing population diversity and slow convergence speed at the late stage of evolution. In order to validate and analyze the performance of MLHO, numerical experiments were conducted on benchmark problems in each dimension of CEC'20. In addition, comparisons with 4 state-of-the-art (SOTA) algorithms are executed. The results show that the performance of MLHO is significantly superior to, or at least comparable to the SOTA algorithms. At the same time, the feasibility and effectiveness of MLHO are also demonstrated for the speed control problem of the pipeline inner detector.}, archive = {J_NEUCOM}, author = {Jinze Liu and Jian Feng and Huaguang Zhang and Shengxiang Yang}, doi = {10.1016/j.neucom.2025.129715}, journal = {Neurocomputing}, month = {4}, pages = {129715}, shortjournal = {Neurocomputing}, title = {A novel multi-level hierarchy optimization algorithm for pipeline inner detector speed control}, volume = {627}, year = {2025}, }
• (2025). StereoSqueezeNet: With fewer parameters but higher accuracy than SqueezeNet. NEUCOM, 627, 129677. (www)

  @article{ , abstract = {Convolutional neural networks (CNNs) have evolved from the initial LeNet to date, and network models have become increasingly deep and comprehensive. It has been proven that deeper networks have better fitting effects, but the corresponding parameter size and computational complexity increase rapidly. With the continuous development of mobile Internet technology, portable devices have been rapidly popularized, and users have put forward more and more demands. Thus, how to design efficient and high-performance lightweight convolutional neural networks (CNNs) is the key to solve this challenging problem. Recently, this type of convolutional neural networks (CNNs)--lightweight convolutional neural networks (CNNs), which adopt the design concept of compression networks and maintain high accuracy with fewer parameters, has attracted increasing attention. SqueezeNet is a lightweight CNN adapting to edge device deployment. Its number of parameters is only 1/50 of AlexNet, but it achieves the same accuracy as AlexNet. In order to make the network more lightweight, inspired by SqueezeNet, MobileNet, SENet, SKNet, AlexNet, etc., in this paper we propose StereoSqueezeNet, using much fewer parameters but achieving even better accuracy than SqueezeNet.}, archive = {J_NEUCOM}, author = {Qiaoyan Sun and Jianfei Chen}, doi = {10.1016/j.neucom.2025.129677}, journal = {Neurocomputing}, month = {4}, pages = {129677}, shortjournal = {Neurocomputing}, title = {StereoSqueezeNet: With fewer parameters but higher accuracy than SqueezeNet}, volume = {627}, year = {2025}, }
• (2025). Nonlinearly activated gradient neuronet with fixed-time convergence for online equality-constrained quadratic programming. NEUCOM, 627, 129608. (www)

  @article{ , abstract = {The objective of this study is to solve the equality-constrained quadratic programming (ECQP) problem online. This problem frequently appears in various fields. In this regard, a nonlinearly activated gradient neuronet (GN) with fixed-time convergence is designed and investigated by adhibiting two novel activation functions. Detailed theoretical derivations and simulation observations sustain its fixed-time convergence and superiority compared with existing GN models in terms of convergence. Moreover, a robotic path tracking example is successfully conducted, and the corresponding results further showcase the superior convergence performance of the proposed fixed-time convergent GN model.}, archive = {J_NEUCOM}, author = {Xuanjiao Lv and Ke Chen and Junying Yuan}, doi = {10.1016/j.neucom.2025.129608}, journal = {Neurocomputing}, month = {4}, pages = {129608}, shortjournal = {Neurocomputing}, title = {Nonlinearly activated gradient neuronet with fixed-time convergence for online equality-constrained quadratic programming}, volume = {627}, year = {2025}, }
• (2025). Addressing bayes imbalance in partial label learning via range adaptive graph guided disambiguation. NEUCOM, 627, 129606. (www)

  @article{ , abstract = {Partial label learning is a multi-class classification problem in weakly supervised learning. Each instance corresponds to multiple labels, only one of which is the ground-truth label. To correctly identify the ground-truth label, a common approach is label disambiguation, i.e. finding the ground-truth label by distinguishing the label confidence of different labels. Recently, graph-based disambiguation strategies have shown excellent results by utilizing feature space to generate label confidence. However, these approaches focus only on nearest neighbor samples. When the instances suffer from Bayes imbalance, i.e. different label instance distributions are coupled to each other, resulting in some instances being surrounded by instances from other labels, the algorithm’s performance may be degraded. We show a novel partial label learning approach based on range adaptive graph guided disambiguation , revealing that expanding the range of graph structure can effectively reduce the negative effect of Bayes imbalance. Meanwhile, we add label distribution constraint to the label disambiguation process to weaken the negative impact of label noise and class imbalance. Numerous experiments show that range adaptive graph guided disambiguation and label distribution constraint can improve the performance of the algorithm when facing the problem of data imbalance.}, archive = {J_NEUCOM}, author = {Zhenbing Liu and Zhaoyuan Zhang and Haoxiang Lu and Wenhao Wang}, doi = {10.1016/j.neucom.2025.129606}, journal = {Neurocomputing}, month = {4}, pages = {129606}, shortjournal = {Neurocomputing}, title = {Addressing bayes imbalance in partial label learning via range adaptive graph guided disambiguation}, volume = {627}, year = {2025}, }
• (2025). Deep learning-based depression recognition through facial expression: A systematic review. NEUCOM, 627, 129605. (www)

  @article{ , abstract = {Depression is a type of prevalent mental illness that can lead to suicidal or self-harm behaviors in severe cases. Recently, depression recognition has garnered extensive attention from the deep learning community due to its urgent need to assist conventional diagnostic methods. Deep learning-based depression recognition through facial expression (DL-FEDR) is one of the most popular research directions. Therefore, this paper tries to summarize advances from 2017 to 2024 in DL-FEDR. We focus on (1) Various approaches of DL-FEDR are divided into two categories: spatial and spatial–temporal features. (2) These methods are analyzed from metrics results, ethical privacy, application scenarios and technological advancements. (3) Present challenges and future directions of DL-FEDR systems are discussed.}, archive = {J_NEUCOM}, author = {Xiaoming Cao and Lingling Zhai and Pengpeng Zhai and Fangfei Li and Tao He and Lang He}, doi = {10.1016/j.neucom.2025.129605}, journal = {Neurocomputing}, month = {4}, pages = {129605}, shortjournal = {Neurocomputing}, title = {Deep learning-based depression recognition through facial expression: A systematic review}, volume = {627}, year = {2025}, }
• (2025). Data-driven machinery fault diagnosis: A comprehensive review. NEUCOM, 627, 129588. (www)

  @article{ , abstract = {In this era of advanced manufacturing, it is now more crucial than ever to diagnose machine faults as early as possible to guarantee their safe and efficient operation. With the increasing complexity of modern industrial processes, traditional machine health monitoring approaches cannot provide efficient performance. With the massive surge in industrial big data and the advancement in sensing and computational technologies, data-driven machinery fault diagnosis solutions based on machine/deep learning approaches have been used ubiquitously in manufacturing applications. Timely and accurately identifying faulty machine signals is vital in industrial applications for which many relevant solutions have been proposed and are reviewed in many earlier articles. Despite the availability of numerous solutions and reviews on machinery fault diagnosis, existing works often lack several aspects. Most of the available literature has limited applicability in a wide range of manufacturing settings due to their concentration on a particular type of equipment or method of analysis. Additionally, discussions regarding the challenges associated with implementing data-driven approaches, such as dealing with noisy data, selecting appropriate features, and adapting models to accommodate new or unforeseen faults, are often superficial or completely overlooked. Thus, this survey provides a comprehensive review of the articles using different types of machine learning approaches for the detection and diagnosis of various types of machinery faults, highlights their strengths and limitations, provides a review of the methods used for predictive analyses, comprehensively discusses the available machinery fault datasets, introduces future researchers to the possible challenges they have to encounter while using these approaches for fault diagnosis and recommends the probable solutions to mitigate those problems. The future research prospects are also pointed out for a better understanding of the field. We believe that this article will help researchers and contribute to the further development of the field.}, archive = {J_NEUCOM}, author = {Dhiraj Neupane and Mohamed Reda Bouadjenek and Richard Dazeley and Sunil Aryal}, doi = {10.1016/j.neucom.2025.129588}, journal = {Neurocomputing}, month = {4}, pages = {129588}, shortjournal = {Neurocomputing}, title = {Data-driven machinery fault diagnosis: A comprehensive review}, volume = {627}, year = {2025}, }
• (2025). Long-tailed recognition via key attribute learning. NEUCOM, 627, 129586. (www)

  @article{ , abstract = {Deep learning models often struggle with datasets exhibiting long-tailed distributions, where the majority of data is concentrated in a few categories, leaving many with very few samples. This imbalance results in models favouring well-represented categories, leading to poorer performance for those with fewer instances. Existing methodologies focus on addressing class-wise imbalance but disregard the attribute-wise disparities. By assigning equal weight to each instance within a class, these approaches overlook the long-tailed distribution of attributes, thus underrepresenting information from infrequent attributes. The reduction in feature diversity consequently diminishes model performance. To address this challenge, we introduce an innovative methodology, namely Key Attribute Learning (KAL). It emphasises the importance of less common attributes by utilising the Instance Diversity Index (IDI) to assess and prioritise attribute diversity for each instance. KAL effectively expands feature margins among categories and addresses the overfitting problem. Our results demonstrate that KAL is non-invasive in both single-model and Mixture of Experts (MoE) settings. Implementing our method on BalPoE, we attained state-of-the-art (SOTA) performance on CIFAR-100-im100, ImageNet-LT, and iNaturalist datasets, showcasing its broad applicability and significant improvements across both balanced and diverse test distributions.}, archive = {J_NEUCOM}, author = {Yu Fu and Jungong Han and Xiang Chang and Changrui Chen and Changjing Shang and Qiang Shen}, doi = {10.1016/j.neucom.2025.129586}, journal = {Neurocomputing}, month = {4}, pages = {129586}, shortjournal = {Neurocomputing}, title = {Long-tailed recognition via key attribute learning}, volume = {627}, year = {2025}, }
• (2025). Spatio-temporal graph-based self-labeling for video anomaly detection. NEUCOM, 627, 129576. (www)

  @article{ , abstract = {Video anomaly detection (VAD) aims to identify abnormal events in a video sequence. Existing methods achieve VAD by learning the decision boundary between the normal space and the abnormal space pre-defined in the training data. However, these methods trend to neglect the distribution gap between the pre-defined abnormal space and the real one, which lead to overfitting on the normal space or bias toward the pre-defined abnormal space. In this paper, we propose a spatio-temporal graph-based self-labeling method that not only focuses on the pre-defined abnormal space but considers the real abnormal space, enabling it to capture the decision boundary between the normal space and a complementary space, called as the not-normal space. We first construct a spatio-temporal graph (ST-Graph) based on the objects of input video and utilize a spatio-temporal graph convolution network (ST-GCN) to model the interaction between objects. We then propose a self-labeling-based learning mechanism that encourages the proposed ST-GCN to record the normal events while abstaining from labeling the pseudo-abnormal events, thereby aggregating the pre-defined and real abnormal spaces into not-normal space. To evaluate the model performance on localizing anomalous objects and capturing interactions between objects, we further introduce an object-level criterion that bridges frame-level and pixel-level criteria. Our method is validated on three datasets and achieves state-of-the-art frame-level AUC results on Avenue (92.5%), and outperforms existing ST-Graph-based methods on UCSD Ped2 (96.5%) and ShanghaiTech (76.8%).}, archive = {J_NEUCOM}, author = {Meng Xing and Zhiyong Feng and Yong Su and Yiming Zhang and Changjae Oh and Valeriya Gribova and Vladimir Fedorovich Filaretoy and Deshuang Huang}, doi = {10.1016/j.neucom.2025.129576}, journal = {Neurocomputing}, month = {4}, pages = {129576}, shortjournal = {Neurocomputing}, title = {Spatio-temporal graph-based self-labeling for video anomaly detection}, volume = {627}, year = {2025}, }
• (2025). MDANet: A multi-stage domain adaptation framework for generalizable low-light image enhancement. NEUCOM, 627, 129572. (www)

  @article{ , abstract = {Low-light image enhancement is a vital process that seeks to enhance the visibility and clarity of images taken under suboptimal lighting conditions. Traditional methods for low-light enhancement rely on altering illumination levels but struggle with local variations in lighting levels and complex scenes, often producing artifacts, color distortions, or unnatural appearances. Recent deep learning-based methods have improved performance through robust feature extraction but usually rely on single-domain datasets that cannot generalize to all lighting conditions present in real-life situations. To address these generalization challenges, we propose MDANet: a Multi-Stage Domain Adaptation framework that sequentially implements light, focus, and contextual domain adaptation. First, MDANet employs an Adaptive Lighting Enhancement Module that combines an innovative U-Net architecture with cross-domain contrastive learning for lighting invariance. Second, this framework utilizes large language models (LLMs) for context-aware adjustments while incorporating depth map constraints within an innovative Region-Focused Enhancement Module that selectively targets and processes key focal regions. Third, a Contextual Adaptation Module employs multitask meta-learning to ensure robustness and superior performance across a range of contextual scenarios. Experimental results demonstrate that MDANet consistently outperforms both traditional and state-of-the-art methods by offering both improved image quality as well as generalization across diverse low-light scenarios. Experimental results show that MDANet achieves state-of-the-art performance across multiple datasets, including MUSIQ scores of 38.234 on LOL, 39.811 on SID, 39.391 on SICE, and 39.971 on RELLISUR, outperforming all compared methods in both perceptual and quantitative metrics.}, archive = {J_NEUCOM}, author = {Jianhui Wang and Yangfan He and Kun Li and Sida Li and Lan Zhao and Jun Yin and Miao Zhang and Tianyu Shi and Xueqian Wang}, doi = {10.1016/j.neucom.2025.129572}, journal = {Neurocomputing}, month = {4}, pages = {129572}, shortjournal = {Neurocomputing}, title = {MDANet: A multi-stage domain adaptation framework for generalizable low-light image enhancement}, volume = {627}, year = {2025}, }
• (2025). Multimodal hypergraph network with contrastive learning for sentiment analysis. NEUCOM, 627, 129566. (www)

  @article{ , abstract = {Multimodal Sentiment Analysis (MSA) is the process of relying on multimodal information, such as text, audio, and visual, to determine a subject’s affective tendencies. While many recent studies have adopted graph-based techniques for MSA, they have yet to fully explore the sentimental interactions both within unimodal temporal steps and across individual modalities. To address the limitations arising from the isolation of unimodal hypergraphs in affective relationship mining, this paper proposes a multimodal hypergraph network based on contrastive learning. It constructs the hypergraph structure by utilizing the sequential time steps of all three modalities as a collection of nodes, aiming to explore the multidimensional affective relationships across uni-, bi-, and tri-modalities. Specifically, this paper first generates the initial hypergraph structure using a correlation-based hypergraph construction method to ensure the effectiveness of the constructed hypergraph. Then, both supervised and unsupervised contrastive learning methods are designed to optimize feature learning and the structure of the multimodal hypergraph, adaptively and simultaneously capturing the relationships among the time-series nodes of uni-, bi-, and tri-modalities. The proposed methods in this paper have demonstrated the advantages and effectiveness by conducting a large number of comparative and validation experiments on the English CMU-MOSI and CMU-MOSEI datasets as well as the Chinese CH-SIMS dataset.}, archive = {J_NEUCOM}, author = {Jian Huang and Kun Jiang and Yuanyuan Pu and Zhengpeng Zhao and Qiuxia Yang and Jinjing Gu and Dan Xu}, doi = {10.1016/j.neucom.2025.129566}, journal = {Neurocomputing}, month = {4}, pages = {129566}, shortjournal = {Neurocomputing}, title = {Multimodal hypergraph network with contrastive learning for sentiment analysis}, volume = {627}, year = {2025}, }
• (2025). Object-fabrication targeted attack for object detection. NEUCOM, 627, 129561. (www)

  @article{ , abstract = {Recent studies have demonstrated that object detection networks are usually vulnerable to adversarial examples. Generally, adversarial attacks for object detection can be categorized into targeted and untargeted attacks. Compared with untargeted attacks, targeted attacks present greater challenges and all existing targeted attack methods launch the attack by misleading detectors to mislabel the detected object as a specific wrong label. However, since these methods must depend on the presence of the detected objects within the victim image, they suffer from limitations in attack scenarios and attack success rates. In this paper, we propose a targeted feature space attack method that can mislead detectors to ‘fabricate’ extra designated objects regardless of whether the victim image contains objects or not. Specifically, we introduce a guided image to extract coarse-grained features of the target objects and design an innovative dual attention mechanism to filter out the critical features of the target objects efficiently. The attack performance of the proposed method is evaluated on MS COCO and BDD100K datasets with FasterRCNN and YOLOv5. Evaluation results indicate that the proposed targeted feature space attack method shows significant improvements in terms of image-specific, universality, and generalization attack performance, compared with the previous targeted attack for object detection.}, archive = {J_NEUCOM}, author = {Xuchong Zhang and Changfeng Sun and Haoliang Han and Hongbin Sun}, doi = {10.1016/j.neucom.2025.129561}, journal = {Neurocomputing}, month = {4}, pages = {129561}, shortjournal = {Neurocomputing}, title = {Object-fabrication targeted attack for object detection}, volume = {627}, year = {2025}, }
• (2025). Lag formation control for nonlinear second-order multiagent systems without and with external disturbances. NEUCOM, 627, 129547. (www)

  @article{ , abstract = {The lag and lag H ∞ formation control of nonlinear second-order multiagent systems (MASs) without and with communication delay are addressed. Based on inequality techniques and a distributed state feedback control strategy, a sufficient condition is presented for lag formation of the MAS. Moreover, a state feedback control scheme with communication delay is also developed to ensure the lag formation for the MAS. In addition, these control schemes are further extended to the MAS with external disturbances, and several lag H ∞ formation criteria for the MAS are also given. Finally, the devised distributed state feedback control protocols are verified via numerical examples.}, archive = {J_NEUCOM}, author = {Xiao-Lu Wen and Jin-Liang Wang and Kun Ling}, doi = {10.1016/j.neucom.2025.129547}, journal = {Neurocomputing}, month = {4}, pages = {129547}, shortjournal = {Neurocomputing}, title = {Lag formation control for nonlinear second-order multiagent systems without and with external disturbances}, volume = {627}, year = {2025}, }
• (2025). Neural-network-based adaptive fixed-time control for stochastic multi-agent systems. NEUCOM, 627, 129538. (www)

  @article{ , abstract = {This article deals with the fixed-time control design issue for stochastic multi-agent systems (MASs). First of all, a new practical fixed-time stability criterion in probability is proposed. Compared with existing works, the settling time is exclusively determined by design parameters, signifying that it can be calculated with precision. Utilizing this stability criterion, a fixed-time control strategy for stochastic MASs is designed, principally leveraging the backstepping control techniques and the radial basis function neural networks (RBF NNs). Additionally, the singularity problem in the control scheme is avoided by exploiting L’Hôpital’s rule. With the designed control strategy, the stochastic MASs achieve practical fixed-time stability. Furthermore, the tracking errors converge to an adjustable range near zero. The effectiveness of the proposed control strategy is verified by a series numerical simulation.}, archive = {J_NEUCOM}, author = {Ximin Li and An Yan and Shengqi Zhu and Dengxiu Yu and C.L. Philip Chen}, doi = {10.1016/j.neucom.2025.129538}, journal = {Neurocomputing}, month = {4}, pages = {129538}, shortjournal = {Neurocomputing}, title = {Neural-network-based adaptive fixed-time control for stochastic multi-agent systems}, volume = {627}, year = {2025}, }
• (2025). Learning general multi-agent decision model through multi-task pre-training. NEUCOM, 627, 129524. (www)

  @article{ , abstract = {Multi-agent reinforcement learning (MARL), involving collaboration and competition among multiple agents, demonstrates significant potential in addressing complex tasks and facilitating the cooperative nature of artificial intelligence systems. However, training MARL networks is a challenging task, demanding extensive computational resources to interact with complex environments, extract state representations, and acquire decision knowledge. Recent breakthroughs in large-scale pre-trained models have sparked our interest. We aim to discover shared knowledge in MARL through model pre-training. In multi-agent systems, the absence of shared state representations often arises due to differences in environments across tasks. This complicates the pre-training of common perception models. Nevertheless, there typically exists common decision logic among different multi-agent tasks, such as cooperation. Pre-training a general decision model for multi-agents proves to be an efficient solution. However, the coupling of perception and decision-making often impedes the individual pre-training of decision models. In this paper, we propose a novel approach based on multi-task pre-training to learn a general multi-agent decision model (GMADM). This approach decouples perception from decision-making, extracting shared decision logic across different tasks, enabling rapid learning of excellent policies on new tasks. Our approach demonstrates outstanding performance in the StarCraft Multi-Agent Challenge (SMAC) and Google Research Football (GRF) environments. Experimental results indicate that pre-trained general decision models can smoothly transfer to new tasks, significantly reducing training costs and improving performance. Additionally, visualization confirms the presence of common multi-agent decision knowledge in shared network layers, further validating the effectiveness of our approach. Code is available at https://github.com/wangjw55/GMADM .}, archive = {J_NEUCOM}, author = {Jiawei Wang and Lele Xu and Changyin Sun}, doi = {10.1016/j.neucom.2025.129524}, journal = {Neurocomputing}, month = {4}, pages = {129524}, shortjournal = {Neurocomputing}, title = {Learning general multi-agent decision model through multi-task pre-training}, volume = {627}, year = {2025}, }
• (2025). ZSDECNet: A zero-shot deep learning framework for image exposure correction. NEUCOM, 627, 129399. (www)

  @article{ , abstract = {When shooting street scenes at night, the captured images may be underexposed or overexposed, which seriously affects human visual perception. Therefore, exposure correction is required for these images. Most existing exposure correction methods rely heavily on reference images and the exposure correction results are not thorough. To address these issues, we propose a zero-shot multi-exposure correction method based on S-curves, called ZSDECNet. Our method is divided into two parts: multi-exposure correction and fusion. First, the illumination channel of the image and the corresponding inverted image are subjected to an initial exposure correction. Moreover, by exposure fusion technique, we select the best exposed area for exposure fusion from the exposure correction results of the input image, the illumination channel, and its inverted channel in order to obtain a visually optimal exposure-corrected image. The method is a zero-shot end-to-end training approach that does not require additional data. In addition, the S-type exposure correction curve corrects both underexposed and overexposed areas, which makes it possible to obtain more thorough exposure correction results. Experiments on existing datasets with various exposure conditions (underexposed and overexposed) and on a real nighttime street scene dataset show that our method outperforms state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Wenchao Li and Shuyuan Wen and Jinhao Zhu and Qiaofeng Ou and Yanchun Guo and Jiabao Chen and Bangshu Xiong}, doi = {10.1016/j.neucom.2025.129399}, journal = {Neurocomputing}, month = {4}, pages = {129399}, shortjournal = {Neurocomputing}, title = {ZSDECNet: A zero-shot deep learning framework for image exposure correction}, volume = {627}, year = {2025}, }
• (2025). Leveraging language models for automated distribution of review notes in animated productions. NEUCOM, 626, 129620. (www)

  @article{ , abstract = {During the production of an animated film, professionals at the animation studio prepare thousands of notes. These notes describe improvements and corrections identified by supervisors and directors during daily meetings where the film’s progress is reviewed. After each meeting, these notes are manually distributed to the appropriate departments that need to address them. Due to the manual nature of this process, many notes are not assigned correctly, and the identified issues are not addressed, reducing the final quality of the film. This article describes and compares several approaches to automatically distribute notes using multi-label text classification with different language models (LM). Implemented methods include logistic regression models, encoder-only models such as the BERT family, and decoder-only models such as Llama 2 including fine-tuning and QLoRA techniques. Training and inference were conducted on a local RTX-3090. The results of the different techniques have been compared, achieving a maximum average accuracy of 0.83 and an f1-score of 0.89 with the fine-tuned Multilingual BERT model. This demonstrates the validity of these models for multi-label text classification, as well as their usefulness in a hitherto unexplored area such as animation studios.}, archive = {J_NEUCOM}, author = {Diego Garcés and Matilde Santos and David Fernández-Llorca}, doi = {10.1016/j.neucom.2025.129620}, journal = {Neurocomputing}, month = {4}, pages = {129620}, shortjournal = {Neurocomputing}, title = {Leveraging language models for automated distribution of review notes in animated productions}, volume = {626}, year = {2025}, }
• (2025). Novel dynamic data-driven modeling based on feature enhancement with derivative memory LSTM for complex industrial process. NEUCOM, 626, 129619. (www)

  @article{ , abstract = {Effective feature extraction is important for accurate data-driven soft sensor modeling in large-scale dynamic industrial processes. However, due to the temporal, nonlinear, and high-dimensional nature of the data collected from large-scale industrial processes, traditional data-driven modeling methods often suffer from imperfect feature extraction. To overcome this challenge, this paper proposes a novel feature enhancement framework based on derivative memory long short-term memory (FEDM-LSTM) algorithm for soft sensor modeling. First, inspired by the proportional–integral–derivative control theory, the LSTM is equipped with a derivative gate that dynamically predicts future information. Combined with inherent gates that resemble the proportional and integral parts in traditional LSTM, the derivative memory LSTM (DM-LSTM) captures the dynamic information of the past, the present, and the future. Then, to adapt to multiple phases in complex industrial systems, a feature enhancement framework is designed for DM-LSTM in which features representing important dynamic information from previous phases are fed to the DM-LSTM as additional input in the current phase. Finally, the effectiveness of the proposed method is evaluated through two real industrial datasets, showcasing its ability to achieve high prediction accuracy.}, archive = {J_NEUCOM}, author = {Xiuli Zhu and Jiajun Xu and Zixuan Fu and Seshu Kumar Damarla and Peng Wang and Kuangrong Hao}, doi = {10.1016/j.neucom.2025.129619}, journal = {Neurocomputing}, month = {4}, pages = {129619}, shortjournal = {Neurocomputing}, title = {Novel dynamic data-driven modeling based on feature enhancement with derivative memory LSTM for complex industrial process}, volume = {626}, year = {2025}, }
• (2025). HGNN-NHSA: Heterogeneous graph neural network based on neighbor and high-order subgraph aggregation learning. NEUCOM, 626, 129618. (www)

  @article{ , abstract = {Heterogeneous graphs (HGs) have shown significant advantages in several complex real-world scenarios, such as social network analysis, recommender systems, and bioinformatics, due to their ability to efficiently model complex association relationships between different types of nodes and edges. However, in many real-world scenarios, due to reasons such as data privacy or information loss, node attributes in heterogeneous graphs are often incomplete, which challenges the performance of the model. In addition, most heterogeneous graph neural networks (HGNNs) currently have certain limitations in mining node attribute, such as capturing high-order isomorphic node attribute through meta-paths, but ignoring the influence of heterogeneous neighboring nodes. To address the above challenges, we propose a heterogeneous graph neural network based on neighbor and high-order subgraph aggregation learning. Specifically, a node attribute completion strategy was first proposed, which completes missing attributes of nodes through an encoder–decoder architecture; Secondly, a subgraph learning mechanism was designed to divide heterogeneous graphs into multiple neighbor and high-order subgraphs, and specialized node aggregation methods were provided for different types of subgraphs to capture richer local and global information; Finally, based on the cross-attention mechanism, subgraphs from different perspectives are fused to generate the final node representation for downstream tasks. Through extensive experiments, it has been demonstrated that our proposed model outperforms existing advanced methods on three real-world datasets, demonstrating superior performance.}, archive = {J_NEUCOM}, author = {Weiqing Sun and Yuanxun Cheng and Xiaohu Zhao}, doi = {10.1016/j.neucom.2025.129618}, journal = {Neurocomputing}, month = {4}, pages = {129618}, shortjournal = {Neurocomputing}, title = {HGNN-NHSA: Heterogeneous graph neural network based on neighbor and high-order subgraph aggregation learning}, volume = {626}, year = {2025}, }
• (2025). Policy distillation for efficient decentralized execution in multi-agent reinforcement learning. NEUCOM, 626, 129617. (www)

  @article{ , abstract = {Cooperative Multi-Agent Reinforcement Learning (MARL) addresses complex scenarios where multiple agents collaborate to achieve shared objectives. Training these agents in partially observable environments within the Centralized Training with Decentralized Execution (CTDE) framework remains challenging due to limited information access and the need for lightweight agent networks. To overcome these challenges, we introduce the Centralized Training and Policy Distillation for Decentralized Execution (CTPDE) framework. We propose a centralized dual-attention agent network that integrates global state and local observations to enable lossless value decomposition and prevent homogeneous agent behaviors. In addition, an efficient policy distillation method is proposed, in which a network of action value distribution is distilled from the centralized agent network, ensuring the efficiency of decentralized execution. The evaluation of CTPDE in benchmark environments demonstrates that the attention-based network achieves state-of-the-art performance during training. Moreover, the distilled agent network surpasses existing RNN-based methods and, in some cases, matches the capabilities of more complex architectures. These findings underscore the potential of CTPDE for advancing cooperative MARL tasks.}, archive = {J_NEUCOM}, author = {Yuhang Pei and Tao Ren and Yuxiang Zhang and Zhipeng Sun and Matys Champeyrol}, doi = {10.1016/j.neucom.2025.129617}, journal = {Neurocomputing}, month = {4}, pages = {129617}, shortjournal = {Neurocomputing}, title = {Policy distillation for efficient decentralized execution in multi-agent reinforcement learning}, volume = {626}, year = {2025}, }
• (2025). Learning-based aggregation of quasi-nonlinear fuzzy cognitive maps. NEUCOM, 626, 129611. (www)

  @article{ , abstract = {Quasi-Nonlinear Fuzzy Cognitive Maps (q-FCMs) are an algorithmic generalization of Fuzzy Cognitive Maps (FCMs) used for modeling and simulation. The key advantages of q-FCMs include their interpretability and hybrid reasoning capabilities where expert knowledge and historical data can be exploited to build the model. Another distinctive feature of neural cognitive mapping is that it allows the aggregation of different models that represent the same problem into a unified neural system. Unfortunately, existing aggregation algorithms focus on producing an aggregated model that resembles the structure of the individual q-FCMs, while neglecting the functional aspect. The ramification of this oversight is that the simulation results produced by the aggregated model often differ significantly from those generated by the individual models. In this paper, we introduce a parameterized learning-based method for aggregating q-FCMs that considers both aspects. Firstly, it ensures that the aggregated model’s weight matrix is reasonably similar to those associated with the individual models, thus maintaining the structural integrity of the aggregation. Secondly, it ensures that the aggregated model’s outputs closely align with those produced by the individual models when operating under the same initial conditions. The core of our aggregation method lies in an analytically derived loss function that is minimized using a gradient-based optimizer which approximates the Jacobian and Hessian while using a limited amount of memory. Extensive simulations on synthetically generated models and a case study with diverse structural properties and complexities demonstrate that our approach significantly outperforms representative state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Gonzalo Nápoles and Isel Grau and Agnieszka Jastrzebska and Yamisleydi Salgueiro}, doi = {10.1016/j.neucom.2025.129611}, journal = {Neurocomputing}, month = {4}, pages = {129611}, shortjournal = {Neurocomputing}, title = {Learning-based aggregation of quasi-nonlinear fuzzy cognitive maps}, volume = {626}, year = {2025}, }
• (2025). MAFNet: A multi-scale aligned fusion network for infrared small target detection. NEUCOM, 626, 129610. (www)

  @article{ , abstract = {Infrared small target detection (IRSTD) is a challenging problem that separates small targets from complex background. Existing IRSTD methods all only utilize the slightly barren single-scale input information for the detection of dim small targets, while multi-scale input architecture has not been fully investigated. To this end, we propose a Multi-scale Alignment Fusion Network, MAFNet, which simulates the behavior of the human eye when observing small targets, i.e. , the zoom-in strategy. Specifically, by designing the scale alignment fusion block, MAFNet can learn the mixed-scale semantics and fully perceive the subtle small target discrimination cues. In addition, considering the uncertainty caused by high-intensity misleading background noise, we construct a multi-field search block. It utilizes a rich receptive field to enhance the diversified expression of features, thereby helping the model to robustly segment small targets in complex background and effectively reduce the false-alarm rate. The experimental results on public datasets demonstrate the effectiveness of the proposed method, especially on low-resolution data, where it significantly outperforms other methods. The code is available at https://github.com/Jupiter-Wang/MAFNet .}, archive = {J_NEUCOM}, author = {Zhipeng Wang and Feng Zhou and Pengfei Lu and Wei Zheng and Fan Hao and Jiayang Yin}, doi = {10.1016/j.neucom.2025.129610}, journal = {Neurocomputing}, month = {4}, pages = {129610}, shortjournal = {Neurocomputing}, title = {MAFNet: A multi-scale aligned fusion network for infrared small target detection}, volume = {626}, year = {2025}, }
• (2025). A novel rumor detection method focusing on social psychology with graph attention network. NEUCOM, 626, 129609. (www)

  @article{ , abstract = {The proliferation of online social media has led to an increase in the spread of rumors. Current rumor detection methods have not adequately considered the impact of social psychology and have neglected to integrate text features with other characteristics. This paper introduces a multi-feature fusion rumor detection model, Social Psychology-based Graph ATtention network (SPGAT), designed to enhance the accuracy of rumor detection. In this model, social psychological features are extracted using large language models, encompassing user emotion, social identity, group emotional resonance, and social influence. These features aim to deeply capture the essential attributes of rumors. Concurrently, a multi-head dynamic graph attention convolutional network is constructed. This network amalgamates complex structural features with essential features, thereby effectively capturing spatial propagation features and significant features while paying attention to the high-dimensional hidden features of rumors. Furthermore, a neural network is designed to comprehensively integrate the high-dimensional features of rumors, and rumor detection is achieved through a fully connected layer. Extensive experiments are conducted on three public datasets. Compared to the latest typical detection methods, the proposed method demonstrates certain advantages. Specifically, the accuracy and F1 score of rumor detection are improved by 5.87% and 4.4% on average on Weibo, PHEME 5, and PHEME 9 datasets compared to the latest baselines, respectively. Meanwhile, we verify and analyze the key role of socio-psychological characteristics in rumor propagation, which provides strong support for an in-depth understanding of the rumor propagation mechanism.}, archive = {J_NEUCOM}, author = {Lina Li and Guoxing Liu and Yu Liu and Qinghe Yu and Cheng Luo and Nianfeng Li}, doi = {10.1016/j.neucom.2025.129609}, journal = {Neurocomputing}, month = {4}, pages = {129609}, shortjournal = {Neurocomputing}, title = {A novel rumor detection method focusing on social psychology with graph attention network}, volume = {626}, year = {2025}, }
• (2025). Multi-level hierarchical dynamic graph convolutional networks for motor imagery EEG analysis. NEUCOM, 626, 129594. (www)

  @article{ , abstract = {The human brain network consists of tightly connected modular nodes, including local functional areas and global functional connections. Graph Convolutional Networks (GCNs) have shown impressive capabilities in learning the topological relationships among electroencephalogram (EEG) channels during motor imagery tasks. However, existing GCN methods predominantly focus on a single-level functional connection mode, overlooking the diverse connection patterns across different functional modules and neglecting local functional dependencies. To address this limitation, this paper proposes a novel approach called the Multi-level Hierarchical Dynamic GCN (MHD-GCN) method, which aims to explore the multi-layer dynamic information among EEG channels within distinct functional modules. Firstly, the spatiotemporal attention module is employed to capture critical sequence fragments and spatial position information within EEG signals. Secondly, a multi-level wavelet convolutional block is utilized to capture distinct spectral features from corresponding subbands. Lastly, a pre-constructed topological space is leveraged for representation learning, comprising two branches dedicated to exploring global dynamic information and extracting local functional information, respectively. In each branch, the multi-level adjacency matrix is employed to enhance the expressive power of GCN, enabling the acquisition of discriminative features. Extensive experiments on publicly available datasets show that our proposed method performs better than state-of-the-art approaches on the Physionet dataset. Moreover, interpretability analysis is performed on the model, revealing active brain regions and important electrode pairs associated with movement. These findings provide valuable insights and practical guidance for future research in motor imagery tasks.}, archive = {J_NEUCOM}, author = {Xin Deng and Daijiang Xu and Huaxiang Huo and Xinyi Hong and Bin Liu}, doi = {10.1016/j.neucom.2025.129594}, journal = {Neurocomputing}, month = {4}, pages = {129594}, shortjournal = {Neurocomputing}, title = {Multi-level hierarchical dynamic graph convolutional networks for motor imagery EEG analysis}, volume = {626}, year = {2025}, }
• (2025). Enhancing U-net with low-rank attention skip block for 3D point cloud segmentation. NEUCOM, 626, 129593. (www)

  @article{ , abstract = {The U-Net framework has been widely applied in the field of 3D point cloud segmentation. Most existing methods primarily focus on employing more powerful encoders to enhance the understanding of point cloud data. However, some inherent operations in traditional encoders, such as down-sampling and local neighborhood feature aggregation, restrict the effective receptive field of the model to a relatively small region. To address this issue, we introduce a global Attention Skip Block (ASBlock) to improve the skip connection in the U-Net framework, facilitating feature fusion between each point and the global context. Moreover, inspired by the dense-connection idea exploited by some well-known works such as DenseNet and 3D DenseNet, we further extend the proposed ASBlock to a new version with dense connection that can integrate more global contextual information for better segmentation performance. Additionally, to reduce computational complexity, we compress the global attention model using approximate low-rank matrix decomposition, leading to the development of the Low-rank Attention Skip Block (LrASBlock). This module can be efficiently applied to large-scale datasets and seamlessly integrated into existing U-Net segmentation networks as a universal plug-and-play tool. Finally, extensive experimental results on multiple datasets demonstrate that the integration of LrASBlock can significantly improve segmentation performance of several typical U-Net-based methods. Code is available at https://github.com/Ysc156/LrASBlock .}, archive = {J_NEUCOM}, author = {Shoucheng Yan and Yang Chen and Wenfei Cao and Huibin Li}, doi = {10.1016/j.neucom.2025.129593}, journal = {Neurocomputing}, month = {4}, pages = {129593}, shortjournal = {Neurocomputing}, title = {Enhancing U-net with low-rank attention skip block for 3D point cloud segmentation}, volume = {626}, year = {2025}, }
• (2025). LEFuse: Joint low-light enhancement and image fusion for nighttime infrared and visible images. NEUCOM, 626, 129592. (www)

  @article{ , abstract = {Infrared and visible image fusion (IVIF) aims to represent scenes more richly and accurately by integrating information from both modalities. However, existing IVIF methods are typically designed for normal illumination conditions, aiming to achieve higher scores by maintaining close similarity to the source images. In night scenes, visible images often suffer from both low light and localized overexposure due to the dim environment and the interference from local light sources. These methods fail to explore the information hidden in the dark regions of visible images, resulting in fusion images that lack texture details, appear overall dark, and exhibit poor visual quality. To address this issue, we propose a novel image fusion network called LEFuse. LEFuse not only integrates complementary information from both visible and infrared images but also focuses on recovering hidden texture details in visible images. By doing so, LEFuse enhances the visibility and contrast of the fused image, resulting in a brighter and more vivid representation. To achieve this goal, we propose a set of unsupervised loss functions to drive the network’s learning. This set includes a maximum entropy-based fusion enhancement loss for both image fusion and low-light enhancement, as well as a perceptual loss to mitigate the impact of local overexposure in visible images on the fused result. These losses can be applied to any existing image fusion network, enhancing fused images without compromising fusion performance. Extensive experiments demonstrate that our LEFuse achieves promising results in terms of visual quality and quantitative evaluations, especially in nighttime environments. Our code is publicly available at https://github.com/cheng411523/LEFuse .}, archive = {J_NEUCOM}, author = {Muhang Cheng and Haiyan Huang and Xiangyu Liu and Hongwei Mo and Xiongbo Zhao and Songling Wu}, doi = {10.1016/j.neucom.2025.129592}, journal = {Neurocomputing}, month = {4}, pages = {129592}, shortjournal = {Neurocomputing}, title = {LEFuse: Joint low-light enhancement and image fusion for nighttime infrared and visible images}, volume = {626}, year = {2025}, }
• (2025). Coarse-to-fine text injecting for realistic image super-resolution. NEUCOM, 626, 129591. (www)

  @article{ , abstract = {Image Super-Resolution (ISR) aims at enhancing the resolution of a given image by guessing the RGB values of additional pixels from existing ones. Equipping with the pre-trained text-to-image diffusion models, i.e. Stable Diffusion (SD), and fine-tuning it for image-to-image purposes, recent ISR methods are able to leverage powerful prior knowledge to generate abundant details from the low-resolution observations. They abolished the capability of using precise text to control image generation and thus might lose the prior can be queried via text for ISR. In this work, we propose a plug-and-play text-injecting strategy that capitalizes on the inherent benefits of text guidance from a pre-trained SD model. Specifically, two levels of text prompts, including general and detailed descriptions, are automatically generated via existing large language models. By introducing a time-aware text injector, we are able to inject text features with varying granularity to generate progressively detailed results as diffusion unfolds. Based on the observation that the stochastic nature of the diffusion model may lead to different reconstruction results, we develop two test-time methods that conduct multiple runs with different initial noises to refine the fidelity or realism of reconstructed images. Additionally, we explore the interpretability of our method by visualizing the cross-attention maps in the denoising U-Net, which show an obvious correspondence between semantically meaningful words in the text prompt and the corresponding image features. Extensive experiments demonstrated the superiority of our method in balancing fidelity and realism over current state-of-the-art approaches.}, archive = {J_NEUCOM}, author = {Xiaoyu Chen and Chao Bai and Zhenyao Wu and Xinyi Wu and Qi Zou and Yong Xia and Song Wang}, doi = {10.1016/j.neucom.2025.129591}, journal = {Neurocomputing}, month = {4}, pages = {129591}, shortjournal = {Neurocomputing}, title = {Coarse-to-fine text injecting for realistic image super-resolution}, volume = {626}, year = {2025}, }
• (2025). A novel large language model enhanced joint learning framework for fine-grained sentiment analysis on drug reviews. NEUCOM, 626, 129589. (www)

  @article{ , abstract = {Patient feedback on drug reviews extracted from social media platforms and online forums provides genuine sentiment information regarding post-medication usage. The insights are invaluable for prospective patients seeking appropriate medical guidance and stakeholders within the biomedical industry aiming to improve products. This paper presents a novel joint learning framework for fine-grained aspect-based sentiment analysis on drug reviews. By leveraging prior biomedical knowledge from the domain-specific pre-trained large language model, we address the challenge of fine-grained aspect-based sentiment analysis by collaboratively integrating both coarse and fine-grained contextual features within the text content, capturing precise biomedical aspect terms and the corresponding sentiment polarities. To the best of our knowledge, this work pioneers the initial introduction of a joint learning framework based on the fine-tuned biomedical large language model for fine-grained aspect-based sentiment analysis within drug reviews. By conducting extensive experiments on publicly available drug review datasets and comparing the constructed architecture with state-of-the-art techniques, the joint learning framework outperforms baseline competitors across evaluation metrics.}, archive = {J_NEUCOM}, author = {Haochen Zou and Yongli Wang}, doi = {10.1016/j.neucom.2025.129589}, journal = {Neurocomputing}, month = {4}, pages = {129589}, shortjournal = {Neurocomputing}, title = {A novel large language model enhanced joint learning framework for fine-grained sentiment analysis on drug reviews}, volume = {626}, year = {2025}, }
• (2025). MoSCE-ReID: Mixture of semantic clustering experts for person re-identification. NEUCOM, 626, 129587. (www)

  @article{ , abstract = {This study advances the utilization of semantic information in person re-identification (ReID) by leveraging pre-trained vision-language models, addressing the current limitations in semantic information processing within ReID systems. While recent studies have explored CLIP integration for ReID tasks, their training approaches have inadvertently diminished semantic information by focusing primarily on indirect alignment between person IDs through text encoders and image features. Through comprehensive empirical analysis of semantic information’s role in pedestrian ReID, we propose MoSCE-ReID, a mixed semantic clustering expert model. The framework incorporates two key components: a learnable Attribute Group Weight Extractor (AGWE) and a Mixed of LoRA Expert (MoLE) module, designed specifically for attribute group feature extraction. The final ReID decisions are made through the synergistic integration of attribute group features and global features. Extensive experiments across multiple public datasets demonstrate that our approach, by effectively incorporating person attribute group semantic information, achieves substantial performance improvements in ReID tasks, exhibiting superior generalization capabilities compared to existing frameworks.}, archive = {J_NEUCOM}, author = {Kai Ren and Chuanping Hu and Hao Xi and Yongqiang Li and Jinhao Fan and Lihua Liu}, doi = {10.1016/j.neucom.2025.129587}, journal = {Neurocomputing}, month = {4}, pages = {129587}, shortjournal = {Neurocomputing}, title = {MoSCE-ReID: Mixture of semantic clustering experts for person re-identification}, volume = {626}, year = {2025}, }
• (2025). Brain dynamics simulation of schizophrenia with chaotic neural networks. NEUCOM, 626, 129585. (www)

  @article{ , abstract = {Structural brain networks underpin complex brain dynamics and cognitive functions, but abnormality in the network structures may disrupt neural activity in schizophrenia (SCHZ). The mechanism how structural alterations in brain networks lead to abnormal dynamics remain unclear. In this work, we construct a whole-brain dynamic model to simulate resting-state brain activity, using Aihara chaotic neurons to represent brain regions and structural connection strengths from diffusion spectrum magnetic resonance imaging as coupling weights across 83 regions defined by a standard atlas. Our simulations reveal oscillatory patterns shared across groups, highlighting beta-frequency activity in the default mode, basal ganglia, and somatomotor networks. In SCHZ patients, four right-hemisphere regions, the posterior cingulate, isthmus cingulate, temporal pole and amygdala, show transitions from uncertain states in the control group to stable beta-wave activity, consistent with prior EEG findings of elevated beta waves in SCHZ. Graph theory analysis reveals increased local connection strength in the four regions of the SCHZ group, suggesting that the structural changes promote pathological beta activity. Our results illuminate how region-specific alterations in structural brain networks disrupt its dynamics, offering insights into the neural mechanism driving functional abnormalities in SCHZ.}, archive = {J_NEUCOM}, author = {Guiyang Lv and Ping Zhu and Feiyan Chen and Kazuyuki Aihara and Guoguang He}, doi = {10.1016/j.neucom.2025.129585}, journal = {Neurocomputing}, month = {4}, pages = {129585}, shortjournal = {Neurocomputing}, title = {Brain dynamics simulation of schizophrenia with chaotic neural networks}, volume = {626}, year = {2025}, }
• (2025). A real-time machine learning-based methodology for short-term frequency nadir prediction in low-inertia power systems. NEUCOM, 626, 129583. (www)

  @article{ , abstract = {In the modern era, electricity is vital for societal advancement, driving economic growth and essential functions. However, the landscape of power systems is swiftly changing due to the integration of renewable energy sources and the decline of traditional synchronous generation, which reduces the total rotational inertia of the systems. This reduction in inertia leads to more frequent and severe frequency deviations, directly impacting power system behavior. Therefore, there is a pressing need to anticipate frequency grid disturbances to maintain stability and prevent disruptions. A machine learning approach is proposed to address this issue, providing accurate and responsive frequency forecasting in power systems. This paper introduces a novel methodology that leverages machine learning for short-term minimum frequency prediction, emphasizing efficiency and rapid response. A comprehensive experimentation process was conducted using several popular machine learning models, with their hyperparameters optimized through a Bayesian algorithm and evaluated via cross-validation. Results highlight the effectiveness of Decision Trees, offering a balance between efficiency and efficacy. Validation was conducted using the SCADA of a Typhoon HIL real-time simulator, verifying that the proposed methodology is suitable for real-time applications.}, archive = {J_NEUCOM}, author = {Manuel Jesús Jiménez-Navarro and Jose Miguel Riquelme-Dominguez and Manuel Carranza-García and Francisco M. González-Longatt}, doi = {10.1016/j.neucom.2025.129583}, journal = {Neurocomputing}, month = {4}, pages = {129583}, shortjournal = {Neurocomputing}, title = {A real-time machine learning-based methodology for short-term frequency nadir prediction in low-inertia power systems}, volume = {626}, year = {2025}, }
• (2025). Unified finite-time error analysis of soft Q-learning. NEUCOM, 626, 129582. (www)

  @article{ , abstract = {Soft Q-learning is one of the most commonly used reinforcement learning algorithms for various purposes, e.g., dealing with entropy-regularized Markov decision problems, reducing the overestimation bias, and improving explorations. Its effectiveness in practice has led to its widespread use; however, there has not been much theoretical study on soft Q-learning. This paper attempts to provide an integrated finite-time analytical approach for soft Q-learning from a control-theoretic perspective. We examine three different kinds of soft Q-learning algorithms that use the log-sum-exp operator, the Boltzmann operator, and the mellowmax operator, respectively. Utilizing dynamical switching system models, we obtain the finite-time error bounds of three soft Q-learning variants. We believe that our analysis can assist in a better understanding of soft Q-learning through links with switching system models.}, archive = {J_NEUCOM}, author = {Narim Jeong and Donghwan Lee}, doi = {10.1016/j.neucom.2025.129582}, journal = {Neurocomputing}, month = {4}, pages = {129582}, shortjournal = {Neurocomputing}, title = {Unified finite-time error analysis of soft Q-learning}, volume = {626}, year = {2025}, }
• (2025). Learning semantical dynamics and spatiotemporal collaboration for human pose estimation in video. NEUCOM, 626, 129581. (www)

  @article{ , abstract = {Temporal modeling and spatio-temporal collaboration are pivotal techniques for video-based human pose estimation. Most state-of-the-art methods adopt optical flow or temporal difference, learning local visual content correspondence across frames at the pixel level, to capture motion dynamics. However, such a paradigm essentially relies on localized pixel-to-pixel similarity, which neglects the semantical correlations among frames and is vulnerable to image quality degradations ( e.g. occlusions or blur). Moreover, existing approaches often combine motion and spatial (appearance) features via simple concatenation or summation, leading to practical challenges in fully leveraging these distinct modalities. In this paper, we present a novel framework that learns multi-level semantical dynamics and dense spatio-temporal collaboration for multi-frame human pose estimation. Specifically, we first design a Multi-Level Semantic Motion Encoder using a multi-masked context and pose reconstruction strategy. This strategy stimulates the model to explore multi-granularity spatiotemporal semantic relationships among frames by progressively masking the features of (patch) cubes and frames. We further introduce a Spatial-Motion Mutual Learning module which densely propagates and consolidates context information from spatial and motion features to enhance the capability of the model. Extensive experiments demonstrate that our approach sets new state-of-the-art results on three benchmark datasets, PoseTrack2017, PoseTrack2018, and PoseTrack21.}, archive = {J_NEUCOM}, author = {Runyang Feng and Haoming Chen}, doi = {10.1016/j.neucom.2025.129581}, journal = {Neurocomputing}, month = {4}, pages = {129581}, shortjournal = {Neurocomputing}, title = {Learning semantical dynamics and spatiotemporal collaboration for human pose estimation in video}, volume = {626}, year = {2025}, }
• (2025). Gl-MambaNet: A global-local hybrid mamba network for medical image segmentation. NEUCOM, 626, 129580. (www)

  @article{ , abstract = {In recent years, deep learning has made significant progress in medical image segmentation, but it still faces challenges in capturing global features and managing high computational complexity. The Mamba architecture has garnered considerable attention in medical image segmentation tasks due to its ability to capture global information with low computational overhead while maintaining high segmentation accuracy. Based on this, we propose a dual-path network design that combines CNN and Mamba architectures, utilizing a global-local feature fusion module (GLFFM), a multi-stage attention enhancement module (MSAEM), and a semantic feature alignment enhancement module (SFAEM) to significantly improve semantic consistency and multi-scale feature fusion. Specifically, GLFFM is employed to integrate the complementary features of CNN and Mamba, MSAEM enhances feature representation through channel and spatial attention mechanisms, while SFAEM precisely activates relevant features using group attention mechanisms and ensures consistency in dimensions and resolution across skip connection feature maps. Experimental validation on three publicly available medical image datasets, namely GlaS, PH2, and DSB2018, demonstrates that the proposed model, GL-MambaNet, outperforms existing state-of-the-art methods in terms of segmentation accuracy, memory usage, and computational efficiency.}, archive = {J_NEUCOM}, author = {Xiaoyan Kui and Shen Jiang and Qinsong Li and Yifei Peng and Zhipeng Hu and Beiji Zou}, doi = {10.1016/j.neucom.2025.129580}, journal = {Neurocomputing}, month = {4}, pages = {129580}, shortjournal = {Neurocomputing}, title = {Gl-MambaNet: A global-local hybrid mamba network for medical image segmentation}, volume = {626}, year = {2025}, }
• (2025). Incremental value iteration for optimal output regulation of linear systems with unknown exosystems. NEUCOM, 626, 129579. (www)

  @article{ , abstract = {This paper addresses the optimal output regulation problem for discrete-time linear systems with completely unknown dynamics and unmeasurable exosystem states. The primary objective is to design incremental dataset-based value iteration (VI) reinforcement learning algorithms to derive both state feedback and output feedback controllers. In the context of data-driven optimal control, existing approaches typically require either the exosystem state to be measurable or the design of an autonomous system to reconstruct it. In contrast, this work proposes an incremental dataset-based VI algorithm, which eliminates the need for exosystem state measurement or reconstruction. Additionally, the proposed method allows for the selection of an arbitrary initial admissible control policy, thereby overcoming the challenge of requiring an initial admissible control in policy iteration algorithms. Furthermore, the system state is reconstructed using the incremental dataset, and an optimal output feedback controller is developed based on the proposed VI algorithm. The theoretical convergence of the dataset-based incremental VI algorithm is rigorously analyzed, and comprehensive simulations are conducted to validate its effectiveness.}, archive = {J_NEUCOM}, author = {Chonglin Jing and Chaoli Wang and Dong Liang and Yujing Xu and Longyan Hao}, doi = {10.1016/j.neucom.2025.129579}, journal = {Neurocomputing}, month = {4}, pages = {129579}, shortjournal = {Neurocomputing}, title = {Incremental value iteration for optimal output regulation of linear systems with unknown exosystems}, volume = {626}, year = {2025}, }
• (2025). A cauchy loss-incorporated nonnegative latent factorization of tensors model for spatiotemporal traffic data recovery. NEUCOM, 626, 129575. (www)

  @article{ , abstract = {Intelligent transport systems (ITS) work effectively with complete spatiotemporal traffic data. However, in the actual traffic flow data collection scenarios, it is common for traffic data to be incomplete or corrupted due to communication or sensor malfunctions, etc., which extremely restricts the development of ITS. A nonnegative latent factorization of tensors (NLFT) model is an effective and typical approach for missing spatiotemporal traffic data recovery. However, existing NLFT models ignore the adverse effects of outliers ineluctably mixed with collected incomplete data, which inevitably cause degradation of recovery accuracy. To tackle this issue, this paper proposes a Cauchy Loss-incorporated Non-negative Latent Factorization of Tensors (CL-NLFT) model with three-fold ideas: (a) building the objective function based on the Cauchy loss to attain significant robustness and high recovery accuracy, (b) Taking an alternating direction method of multipliers principle to design an efficient set of parameters learning schemes, (c) achieving self-adaption control of hyperparameter with practicality. As the empirical study on four traffic average speed datasets collected from different cities, the proposed CL-NLFT model achieves significant performance gain in terms of accuracy and computational efficiency compared to the state-of-the-art models.}, archive = {J_NEUCOM}, author = {Hao Wu and Jiajia Mi}, doi = {10.1016/j.neucom.2025.129575}, journal = {Neurocomputing}, month = {4}, pages = {129575}, shortjournal = {Neurocomputing}, title = {A cauchy loss-incorporated nonnegative latent factorization of tensors model for spatiotemporal traffic data recovery}, volume = {626}, year = {2025}, }
• (2025). Adaptive node similarity for DropEdge. NEUCOM, 626, 129574. (www)

  @article{ , abstract = {There are two principal impediments of expanding deep graph convolutional networks (GCNs) due to the assumption of smoothness, i.e., over-fitting and over-smoothing. DropEdge methods relieve the convergence speed and reduce the information loss by randomly dropping a specific rate of edges, hence effectively alleviates these two issues and has been widely used to many backbone models. However, thanks to the blindness and potential risks of randomly removing edges, current DropEdge methods often remove important edges and retain unimportant edges, this inevitably reduce the accuracy of learning results. In order to tackle the challenges in previous DropEdge methods, this paper proposes a precise removal technique through the node similarity, which is closely related to edges. Specifically, we employ the hybrid optimal node similarity to drop edges, on the one hand, the edges that severely affect over-fitting and over-smoothing of nodes, i.e., the edges with high node similarity, are removed; on the other hand, the edges that are outliers and noisy, i.e., the edges with a large difference in similarity to normal nodes, are also removed. Therefore, our methods significantly alleviate over-fitting and over-smoothing, accurately reduce the impact of outliers and noise, more importantly, our methods is a generic skill that can be deployed current GCN and its variants. Experimental results on seven benchmark datasets including three assortative datasets and four disassortative datasets show that our methods outperforms the state-of-the-art methods, improve the performance by a large margin especially for disassortative graphs.}, archive = {J_NEUCOM}, author = {Yangcai Xie and Jiecheng Li and Shichao Zhang}, doi = {10.1016/j.neucom.2025.129574}, journal = {Neurocomputing}, month = {4}, pages = {129574}, shortjournal = {Neurocomputing}, title = {Adaptive node similarity for DropEdge}, volume = {626}, year = {2025}, }
• (2025). Stability analysis of inertial delayed neural network with delayed impulses via dynamic event-triggered impulsive control. NEUCOM, 626, 129573. (www)

  @article{ , abstract = {This paper investigates the stability of inertial delayed neural network under dynamic event-triggered impulsive control (DETIC). We innovate by generating the impulsive sequence through DETIC and incorporating impulsive delays, thereby enhancing the model’s practical relevance. Our methodology involves a two-step process: first, we transform the inertial neural network into a first-order differential form using appropriate vector transformations. Then, leveraging Lyapunov-based dynamic event-triggered control, we derive sufficient conditions for both uniform stability and uniform asymptotic stability of the system. To ensure practical applicability, we establish specific parameter constraints for the DETIC mechanism that precludes the Zeno phenomenon. To demonstrate the accuracy and efficacy of our theoretical results, we present two simulation examples.}, archive = {J_NEUCOM}, author = {Mengyao Shi and Lulu Li and Jinde Cao and Liang Hua and Mahmoud Abdel-Aty}, doi = {10.1016/j.neucom.2025.129573}, journal = {Neurocomputing}, month = {4}, pages = {129573}, shortjournal = {Neurocomputing}, title = {Stability analysis of inertial delayed neural network with delayed impulses via dynamic event-triggered impulsive control}, volume = {626}, year = {2025}, }
• (2025). FASTCC: A lightweight human pose detection method leveraging SimCC. NEUCOM, 626, 129571. (www)

  @article{ , abstract = {As a focal area within computer vision algorithms, human pose estimation algorithms find applications in diverse fields such as security and virtual reality. Achieving a balance between speed and accuracy is imperative for practical applications. Existing methods often present a trade-off between high accuracy and real-time performance. In response, this paper introduces the Fast Coordinate Classification (FastCC) detection head. It employs a shared fully connected Transformer for global self-attention operations on feature layers from the backbone network. The spatial attention coordinate encoder then outputs the coordinates of the horizontal and vertical axes of the keypoints, which are subsequently combined to derive the actual keypoint positions. Experimental evaluations conducted on the COCO and MPII datasets demonstrate that our detection head enhances the accuracy of human pose estimation algorithms while maintaining a lightweight design, outperforming the traditional heatmap method.}, archive = {J_NEUCOM}, author = {Yi Li and Jialin Zhang and Yongtao Wang and Dou Quan and YaBo Yan and Qinghai Yang}, doi = {10.1016/j.neucom.2025.129571}, journal = {Neurocomputing}, month = {4}, pages = {129571}, shortjournal = {Neurocomputing}, title = {FASTCC: A lightweight human pose detection method leveraging SimCC}, volume = {626}, year = {2025}, }
• (2025). Investigating the effects of recursion in convolutional layers using analytical methods. NEUCOM, 626, 129570. (www)

  @article{ , abstract = {Most Convolutional Neural Networks (CNNs) consist of a number of stages of decreasing spatial resolution and increasing channel dimension between succeeding stages, each stage is composed of convolutional blocks that are repeated a number of times. Previous research on very simple CNNs consisting of a number of convolutional layers in each stage demonstrated that the introduction of feedback loops around convolutional layers can improve results. This paper studies the effectiveness of recursion on convolutional blocks in a more general setting and aims at explaining the results. Four recent models, namely ResNet, Inception, MobileNet and DenseNet are considered in this study. It is found that for all but DenseNet, the recursive version produces results that are similar or better than their feedforward counterpart when the number of convolutional blocks are preserved. To understand this finding and to discriminate the functional behaviors of the feedforward and recursive counterparts, we embark on three investigations: (1) measuring the evolution of the contextualization of the neurons of the last layer using the effective receptive field concept; (2) comparing the position and the size of the global coverage of the networks using class activation maps; and (3) analyzing the evolution of the organization of the feature space prior to the classifier using the Silhouette score. The investigations reveal that the recursion of a convolutional block shares many similarities with the behavior of a sequence of that block, indicating that a recursive alternative consisting of a single physical layer, can be regarded as a “faithful simulation” of its deeper‘feedforward counterpart. We conclude that except for densely connected models, the recursion of convolutional blocks is a safe and powerful alternative enhancing modern network architectures.}, archive = {J_NEUCOM}, author = {Johan Chagnon and Markus Hagenbuchner and Ah Chung Tsoi and Franco Scarselli}, doi = {10.1016/j.neucom.2025.129570}, journal = {Neurocomputing}, month = {4}, pages = {129570}, shortjournal = {Neurocomputing}, title = {Investigating the effects of recursion in convolutional layers using analytical methods}, volume = {626}, year = {2025}, }
• (2025). Sparse bayesian based NARX modeling of cortical response: Introducing information entropy for enhancing the stability. NEUCOM, 626, 129569. (www)

  @article{ , abstract = {In this paper, an innovative Sparse Bayesian Learning (SBL)-based modeling approach incorporating Information Entropy (IE) to enhance the stability is developed to create Nonlinear Auto-Regressive model with eXogenous input, aiming to address the challenges of low estimation accuracy, limited computational efficiency and insufficient sparsity in existing methods. This development is conducive to capture the key features of cortical responses when focusing on neural activity, providing more accurate results for studying brain mechanisms. By introducing identity transformations and optimizing parameter update and stopping strategies, both computational efficiency and estimation accuracy of the SBL algorithm are effectively improved, where the iterative matrix within ISBL is refined by the introduced IE, which further strengthens the algorithm performance at low Signal-to-Noise Ratio levels. Extensive evaluation demonstrates the proposed method reduces the error by 48 %, decreases the traditional SBL method's runtime by 70 %, and achieves the sparsest result while maintaining structural accuracy, which shows significant competitiveness in accuracy, efficiency and sparsity as compared to other state-of-the-art methods. Moreover, the analysis of real EEG signals indicates that the brain's response follows a fundamental rhythm pattern of adaptation to both active and passive tasks, and such adaptive process can be effectively captured by the proposed sparse model through the combination of linear and nonlinear terms, each serving distinct roles. These findings offer a novel insight into the human sensorimotor system, which indicates the great potential of the proposed method in assessing sensorimotor impairments and exploring effective clinical intervention method.}, archive = {J_NEUCOM}, author = {Nan Zheng and Yurong Li and Wuxiang Shi and Qiurong Xie}, doi = {10.1016/j.neucom.2025.129569}, journal = {Neurocomputing}, month = {4}, pages = {129569}, shortjournal = {Neurocomputing}, title = {Sparse bayesian based NARX modeling of cortical response: Introducing information entropy for enhancing the stability}, volume = {626}, year = {2025}, }
• (2025). One-step multi-view spectral clustering with subspaces fusion on grassmann manifold. NEUCOM, 626, 129568. (www)

  @article{ , abstract = {Multi-view clustering garners extensive attention due to the increasing significance of accumulated multi-view data. However, current algorithms model diversity and consistency of various views with sparsity constraints, which fail to fully explore structure and relations among multiple views. To address this limitation, we propose a novel O ne-step M ulti-view S pectral C lustering with S ubspaces F usion on G rassmann M anifold (OMSC_SFGM), where similarity graph construction, consistency and diversity learning, and clustering are incorporated into a unified joint learning model. Specifically, OMSC_SFGM constructs a similarity graph for each view using self-representation learning, which is decomposed into diverse and consistent parts. To precisely model the diversity of views, OMSC_SFGM characterizes the diversity of each view using the subspace of other views, where subspace distance on the Grassmann manifold is adopted to quantify correlations among views. This explicitly addresses the relationships between views, enhancing feature quality by maintaining geometric structures across views. Moreover, OMSC_SFGM proposes a one-step spectral strategy that integrates representation learning and clustering with spectral rotation. Comprehensive experiments on eleven real-world datasets demonstrate that the proposed algorithm outperforms baseline methods in multi-view clustering.}, archive = {J_NEUCOM}, author = {Zengfa Dou and Haodong Ren and Yaxiong Ma and Yue Gao and Guohua Huang and Xiaoke Ma}, doi = {10.1016/j.neucom.2025.129568}, journal = {Neurocomputing}, month = {4}, pages = {129568}, shortjournal = {Neurocomputing}, title = {One-step multi-view spectral clustering with subspaces fusion on grassmann manifold}, volume = {626}, year = {2025}, }
• (2025). Saturation function-based fixed/predefined-time synchronization of impulsive quaternion-valued neural networks. NEUCOM, 626, 129565. (www)

  @article{ , abstract = {This article investigates the fixed-time (FXT) and predefined-time (PDT) synchronization issues for a class of impulsive quaternion-valued neural networks (QVNNs). First, two FXT and PDT stability criteria are constructed using average impulse intervals and a new Lyapunov function, where its derivative satisfies the switching differential inequality. Then, the quaternion-valued saturation function is proposed for the first time to construct the control protocol, which differs from the discontinuous sign function based controller due to its continuity and can effectively reduce the chattering phenomenon. Besides, to overcome the large computational and implementation difficulties of the separation method, the non-separation method is used to deal with QVNNs. Based on the inequality technique and the Lyapunov method, sufficient conditions for FXT/PDT synchronization of QVNNs are given, and more precise settling time (ST) estimations are derived by optimal value theory. Finally, these theoretical results are further illustrated via numerical examples and are applied to image encryption.}, archive = {J_NEUCOM}, author = {Jingjing You and Abdujelil Abdurahman and Lvming Liu}, doi = {10.1016/j.neucom.2025.129565}, journal = {Neurocomputing}, month = {4}, pages = {129565}, shortjournal = {Neurocomputing}, title = {Saturation function-based fixed/predefined-time synchronization of impulsive quaternion-valued neural networks}, volume = {626}, year = {2025}, }
• (2025). Multi-granularity classification of upper gastrointestinal endoscopic images. NEUCOM, 626, 129564. (www)

  @article{ , abstract = {The classification of upper gastrointestinal endoscopic images is crucial for enhancing diagnostic performance and improving patient survival rates. Current research in this area has primarily focused on classifying a limited number of coarse-grained categories. To address the scarcity of multi-granularity datasets in this field, a multi-granularity dataset named “multi-granularity labeled upper gastrointestinal endoscopy image dataset” is proposed. Additionally, a multi-granularity collaborative classification network (MGCN) is introduced, which mimics the learning process of physicians who progressively learn medical image features from coarse to fine, ultimately collaborating across multiple granularities to determine the image category at different granularity levels. The proposed framework extracts features from coarse-to-fine granularity and leverages the correlations between different granularities to facilitate mutual enhancement in feature extraction. It learns features at various granularities through a weakly supervised multi-granularity self-attention module and classifies these features from fine to coarse using a multi-granularity collaborative loss module. Extensive experiments conducted on the proposed dataset demonstrate that MGCN achieves state-of-the-art (SOTA) performance across all granularity levels. Specifically, MGCN achieves an accuracy of 85.73% in fine-grained classification, 93.49% in medium-grained classification, and 99.01% in coarse-grained classification, outperforming existing methods such as HRN, PMG and SnapMix. These results highlight the effectiveness of the proposed framework in leveraging multi-granularity features for hierarchical classification. Currently, the multi-granularity labeled upper gastrointestinal endoscopy image dataset is not publicly available because the dataset website is still under construction. Once the system is fully developed, the dataset will be made publicly accessible through this platform to facilitate further research and clinical applications.}, archive = {J_NEUCOM}, author = {Hua Wei and Dongyang Yi and Shuyu Hu and Guobin Zhu and Yi Ding and Minghui Pang}, doi = {10.1016/j.neucom.2025.129564}, journal = {Neurocomputing}, month = {4}, pages = {129564}, shortjournal = {Neurocomputing}, title = {Multi-granularity classification of upper gastrointestinal endoscopic images}, volume = {626}, year = {2025}, }
• (2025). UniSORT: Tracking with unified cue integration and refined association. NEUCOM, 626, 129563. (www)

  @article{ , abstract = {Multi-object tracking (MOT) aims to detect and associate all objects of interest across frames in a video sequence. Despite the impressive performance achieved by existing tracking methods, crowded and frequent occlusions remain challenging issues in MOT. In response to these challenges, we introduce a novel approach, which leverages unified cue integration and refined association to enhance tracking performance in occlusion scenarios. Our method integrates both strong and weak cues, incorporating weak cues such as confidence, pseudo-depth, and height to compensate for the failure of strong cues (geometric and appearance features) when objects are occluded. Moreover, we design a multi-stage association mechanism based on the classification score of detections, separating objects with clear appearances in low-score detections, thus selecting more suitable cost matrices for different subsets of detection boxes. Our method achieves outstanding performance across multiple test set benchmarks, including MOT17 and MOT20, as well as on DanceTrack, where interactions and occlusions are prevalent and severe.}, archive = {J_NEUCOM}, author = {Modi Li and Rui Li and Shaofeng Liang and Yingjing Shi}, doi = {10.1016/j.neucom.2025.129563}, journal = {Neurocomputing}, month = {4}, pages = {129563}, shortjournal = {Neurocomputing}, title = {UniSORT: Tracking with unified cue integration and refined association}, volume = {626}, year = {2025}, }
• (2025). Reinforcement learning-based receding-horizon control for optimal output regulation problem with sensor and actuator attacks. NEUCOM, 626, 129560. (www)

  @article{ , abstract = {This paper considers the output regulation problem for unknown linear discrete-time systems subject to hybrid cyber attacks on both sensor-controller and controller-actuator channels. A distinctive receding-horizon control is proposed to not only predict the future input based on the tampered data collected through the sensor-controller channel, but also recover the compromised input at the actuator end, capable of simultaneously mitigating hybrid attacks and achieving the asymptotic tracking of a dynamic reference. It is essentially a model-free optimal approach in the sense that the control parameters are designed under the condition that the system matrices are unknown and solved through the constrained optimization problem by combining the output regulation theory and reinforcement learning technique.}, archive = {J_NEUCOM}, author = {Xiran Cui and Yi Dong and Chaoli Wang and Shengyuan Xu}, doi = {10.1016/j.neucom.2025.129560}, journal = {Neurocomputing}, month = {4}, pages = {129560}, shortjournal = {Neurocomputing}, title = {Reinforcement learning-based receding-horizon control for optimal output regulation problem with sensor and actuator attacks}, volume = {626}, year = {2025}, }
• (2025). PseudoNeuronGAN: Unpaired synthetic image to pseudo-neuron image translation for label-free neuron instance segmentation. NEUCOM, 626, 129559. (www)

  @article{ , abstract = {Accurate neuron instance segmentation is of great significance in the field of neuroscience. The prerequisite for obtaining high-precision segmentation results using deep learning models is to have a large number of labeled datasets. However, in areas such as the dentate gyrus of the hippocampus where tens of thousands of neurons are aggregated, neuroscientists are unable to label neuron pixels. In this paper, we propose a pipeline for label-free neuron instance segmentation. Firstly, PseudoNeuronGAN, an unpaired synthetic image to pseudo-neuron image translation network, is proposed. Without requiring any manual labeling, synthetic cell images with known centroid labels and real neuron images are sufficient to generate a pseudo-neuron dataset. Since centroid labels are constraints to prevent neuron loss during the translation process, they are consistent in both the synthetic dataset and the generated pseudo-neuron dataset, and can be set as labels for pseudo-neuron images to train deep learning networks to predict the centroids of real neurons. Finally, based on the detected neuron centroids, neuron instance segmentation can be obtained by using competitive region growing algorithm. Experiments show that our pipeline succeeds in performing neuron instance segmentation without the need for manual annotations. PseudoNeuronGAN to generate a labeled pseudo-neuron dataset will greatly reduce the tedious labeling work by neuroscientists, and the accuracy of centroid labels is no longer biased by subjective factors. In terms of instance segmentation performance, the average F-score calculated by classical deep learning models trained on the pseudo-neuron dataset exceeds the average F-score trained on a limited number of real neuron dataset, reflecting the high quality of the generated pseudo-neuron dataset. Our critical code of PseudoNeuronGAN is available at https://github.com/zhenzhen89/PseudoNeuronGAN .}, archive = {J_NEUCOM}, author = {Zhenzhen You and Ming Jiang and Zhenghao Shi and Cheng Shi and Shuangli Du and Minghua Zhao and Anne-Sophie Hérard and Nicolas Souedet and Thierry Delzescaux}, doi = {10.1016/j.neucom.2025.129559}, journal = {Neurocomputing}, month = {4}, pages = {129559}, shortjournal = {Neurocomputing}, title = {PseudoNeuronGAN: Unpaired synthetic image to pseudo-neuron image translation for label-free neuron instance segmentation}, volume = {626}, year = {2025}, }
• (2025). Privacy-preserving face attribute classification via differential privacy. NEUCOM, 626, 129556. (www)

  @article{ , abstract = {The development of face attribute recognition technology has enhanced the intelligence capabilities in the retail industry. Merchants use the surveillance system to capture customers’ face images, and analyze their basic characteristics to provide accurate product recommendations and optimize product configurations. However, these captured face images may contain sensitive visual information, especially identity-related data, which could lead to potential security and privacy risks. Current methods for face privacy protection cannot fully support privacy preserving face attributes classification. To this end, this paper proposes a privacy protection scheme that employs differential privacy in the frequency domain to mitigate risks in face attribute classification systems. Our main goal is to take the frequency domain features perturbed with differential privacy as the input of the face attribute classification model to resist privacy attacks. Specifically, the proposed scheme first transforms the original face image into the frequency domain using the discrete cosine transform (DCT) and removes the DC components that contain the visual information. Then the privacy budget allocation in the differential privacy framework is optimized based on the loss of the face attribute classification network. Finally, the corresponding differential privacy noise is added to the frequency representation. The utilization of differential privacy theoretically provides privacy guarantees. Sufficient experimental results show that the proposed scheme can well balance the privacy-utility.}, archive = {J_NEUCOM}, author = {Xiaoting Zhang and Tao Wang and Junhao Ji and Yushu Zhang and Rushi Lan}, doi = {10.1016/j.neucom.2025.129556}, journal = {Neurocomputing}, month = {4}, pages = {129556}, shortjournal = {Neurocomputing}, title = {Privacy-preserving face attribute classification via differential privacy}, volume = {626}, year = {2025}, }
• (2025). Node importance evaluation in heterogeneous network based on attention mechanism and graph contrastive learning. NEUCOM, 626, 129555. (www)

  @article{ , abstract = {Recently, heterogeneous networks have attracted widespread attention as a modeling approach for complex networks. However, the complex structure and diversity of semantic relations of heterogeneous networks pose challenges to node importance evaluation. In addition, existing methods use SIR model to set true labels for all nodes, which increases the complexity. In light of this, this paper proposes a node importance evaluation in heterogeneous network based on attention mechanism and graph contrastive learning method(AGCL). Specifically, AGCL contains two core parts: node embedding and node importance evaluation. Considering the complex structure of heterogeneous networks, node embedding employs graph contrastive learning to comprehensively extract structural features through neighbor and cross-domain views. Node importance evaluation integrates an attention mechanism to calculate the global and local importance of nodes separately. To reduce the complexity of setting the true labels, a subset of nodes is sampled based on the power-law distribution of node degrees for model training. The evaluation of AGCL is conducted in three real-world networks utilizing three metrics: the maximum connected subgraph node ratio, the maximum propagation range in the linear threshold model, and the maximum propagation range in the independent cascade model. The experiment results demonstrate that AGCL outperforms state-of-the-art techniques in evaluating node importance.}, archive = {J_NEUCOM}, author = {Jian Shu and Yiling Zou and Hui Cui and Linlan Liu}, doi = {10.1016/j.neucom.2025.129555}, journal = {Neurocomputing}, month = {4}, pages = {129555}, shortjournal = {Neurocomputing}, title = {Node importance evaluation in heterogeneous network based on attention mechanism and graph contrastive learning}, volume = {626}, year = {2025}, }
• (2025). Enhancing few-shot out-of-distribution intent detection by reducing attention misallocation. NEUCOM, 626, 129554. (www)

  @article{ , abstract = {Few-shot out-of-distribution (OOD) intent detection aims to detect in-distribution (ID) intents and reject OOD intents with few ID training data. This data scarcity often causes traditional OOD intent detection models to rely on spurious correlations, leading to attention misallocation, where the model focuses on irrelevant features instead of capturing meaningful, intent-specific information. To address these challenges, we propose LiVaS , a novel framework designed for few-shot OOD intent detection. LiVaS enhances label semantics through label interpretation, generates diverse ID samples using a variational autoencoder (VAE), and synthesizes OOD samples to improve decision boundaries. By integrating supervised contrastive learning (SCL) with cross-entropy (CE) loss, the framework encourages the learning of intent-specific features and reduces attention misallocation. Experimental results demonstrate that LiVaS achieves an average improvement of 6% over state-of-the-art methods on two widely used OOD intent detection datasets.}, archive = {J_NEUCOM}, author = {Heng-yang Lu and Jia-ming Zhang and Yuntao Du and Chang Xia and Chongjun Wang and Wei Fang and Xiao-jun Wu}, doi = {10.1016/j.neucom.2025.129554}, journal = {Neurocomputing}, month = {4}, pages = {129554}, shortjournal = {Neurocomputing}, title = {Enhancing few-shot out-of-distribution intent detection by reducing attention misallocation}, volume = {626}, year = {2025}, }
• (2025). MABQN: Multi-agent reinforcement learning algorithm with discrete policy. NEUCOM, 626, 129552. (www)

  @article{ , abstract = {Cooperative multi-agent reinforcement learning (MARL) for continuous control has diverse applications in real-world scenarios. Most of those MARL algorithms focus on enhancing performance through policy-based paradigms, the challenge of low sample efficiency caused by the continuous nature remains underexplored. To address this issue, we propose the M ulti- A gent B ranching Q - N etworks (MABQN) algorithm, an improved QMIX architecture integrating action discretization and value decomposition. MABQN reduces the policy search space by progressively discretizing the continuous action space and decoupling action dimensions, thereby improving learning efficiency. Moreover, it employs a centralized hypernetwork to decompose joint action values, mitigating the credit assignment problem. Experimental results demonstrate that MABQN outperforms other mainstream cooperative MARL algorithms across continuous, discrete, and hybrid action space tasks.}, archive = {J_NEUCOM}, author = {Qing Xie and Zicheng Wang and Yuyuan Fang and Yukai Li}, doi = {10.1016/j.neucom.2025.129552}, journal = {Neurocomputing}, month = {4}, pages = {129552}, shortjournal = {Neurocomputing}, title = {MABQN: Multi-agent reinforcement learning algorithm with discrete policy}, volume = {626}, year = {2025}, }
• (2025). A fast federated reinforcement learning approach with phased weight-adjustment technique. NEUCOM, 626, 129550. (www)

  @article{ , abstract = {Federated reinforcement learning (FRL) enables multiple agents to learn collaboratively without directly sharing their local data. This method addresses the data privacy concerns in the distributed systems. However, FRL faces challenges such as high communication costs, since it requires extensive interactions to achieve satisfied performance. Therefore, this paper develops a fast FRL method with a dynamic aggregation coefficient to reduce the communication load during the learning process. Diverging from traditional FRL techniques which rely on static averaging, our approach begins by setting the initial aggregation coefficient to the logarithm of the number of participating agents. This elevation can enhance the early integration of updates from distributed agents and facilitate a rapid initial learning phase. As communications progress, the aggregation coefficient linearly decreases, transitioning to an average aggregation by the end of the specified interval. This gradual reduction aligns individual learning updates more closely over time, shifting towards a unified global learning model. Furthermore, we implement a value-clipping strategy to constrain global updates within a predefined safe range, thus safeguarding against the potential overflow issues. The aggregation coefficient stabilizes after the initial aggressive integration phase to ensure the training stability. The boundedness analysis of the model aggregation confirms that, despite the high initial coefficient, the parameters of the global model remain within the manageable limits on the FRL server. This strategy is applicable to both tabular and deep learning methods. We validate the designed algorithm on navigation and control tasks, including heterogeneous environments where distinct state transitions and dynamics are designed for each agent. The experimental results demonstrate that our proposed approach achieves faster convergence across various environments.}, archive = {J_NEUCOM}, author = {Yiran Pang and Zhen Ni and Xiangnan Zhong}, doi = {10.1016/j.neucom.2025.129550}, journal = {Neurocomputing}, month = {4}, pages = {129550}, shortjournal = {Neurocomputing}, title = {A fast federated reinforcement learning approach with phased weight-adjustment technique}, volume = {626}, year = {2025}, }
• (2025). Contour-texture preservation transformer for face super-resolution. NEUCOM, 626, 129549. (www)

  @article{ , abstract = {In the field of computer vision, face super-resolution (FSR) technology is an important tool for enhancing the performance of basic tasks such as face recognition and video surveillance. However, when faced with complex face images, existing FSR methods often rely on the Transformer model to improve image quality through its powerful global modeling capabilities. Yet, they tend to be slightly insufficient in local feature extraction due to a lack of adequate local detail capture capabilities. To alleviate these problems, we propose a novel contour-texture preservation Transformer (CTP) method for FSR. This method consists of two key components: the multi-scale attention enhancement block (MSAEB), which captures and fuses image features of different scales to improve the detail level of feature representation, and provides high-quality input for the contour-texture Transformer enhancement block (CTTEB). Additionally, CTTEB integrates convolution operations to enhance local feature extraction and improve feature expression. The feed-forward network (FFN) we introduced ensures the full fusion of global and local information. By combining MSAEB and CTTEB into a residual progressive attention group (RPAG), the network gradually extracts and fuses multi-scale features, ultimately achieving dual preservation of contour structure and texture details. Experiments show that our method achieves the best results on LFW, FFHQ, CelebA, and Helen, with a 0.32 dB increase in PSNR, a 0.0126 increase in SSIM, and a 0.0056 increase in FSIM over the second-best model. Experiments on real-world datasets SCface and Chokepoint confirm that the CTP method excels in both FSR reconstruction and face recognition, verifying its effectiveness.}, archive = {J_NEUCOM}, author = {Ziyi Wu and Yanduo Zhang and Tao Lu and Kanghui Zhao and Jiaming Wang}, doi = {10.1016/j.neucom.2025.129549}, journal = {Neurocomputing}, month = {4}, pages = {129549}, shortjournal = {Neurocomputing}, title = {Contour-texture preservation transformer for face super-resolution}, volume = {626}, year = {2025}, }
• (2025). A contrastive learning strategy for optimizing node non-alignment in dynamic community detection. NEUCOM, 626, 129548. (www)

  @article{ , abstract = {Dynamic community detection, which focuses on tracking local topological variation with time, is crucial for understanding the changing affiliations of nodes to communities in complex networks. Existing researches fell short of expectations primarily due to their heavy reliance on clustering methods or evolutionary algorithms. The emergence of graph contrastive learning offers us a novel perspective and inspiration, which performed well in recognizing pattern at both the node-node and node-graph levels. However, there are still the following limitations in practice: (i) conventional data augmentations may undermine task-relevant information by bring in invalid views or false positive samples, leading the model toward weak discriminative representations. (ii) the non-alignment of nodes caused by dynamic changes also limits the expressive ability of GCL. In this paper, we propose a C ontrastive L earning strategy for O ptimizing N ode non-alignment in D ynamic Community Detection ( CL-OND ). Initially, we confirm the viability of utilizing dynamic adjacent snapshots as monitoring signals through graph spectral experiments, which eliminates the dependence of contrastive learning on traditional data augmentations. Subsequently, we construct an end-to-end dynamic community detection model and introduce a non-aligned neighbor contrastive loss to capture temporal properties and inherent structure of evolutionary graphs by constructing positive and negative samples. Furthermore, extensive experimental results demonstrate that our approach consistently outperforms others in terms of performance.}, archive = {J_NEUCOM}, author = {Xiaohong Li and Wanyao Shi and Qixuan Peng and Hongyan Ran}, doi = {10.1016/j.neucom.2025.129548}, journal = {Neurocomputing}, month = {4}, pages = {129548}, shortjournal = {Neurocomputing}, title = {A contrastive learning strategy for optimizing node non-alignment in dynamic community detection}, volume = {626}, year = {2025}, }
• (2025). GCTGNN: A forecasting method for time series based on graph neural networks and graph clustering. NEUCOM, 626, 129544. (www)

  @article{ , abstract = {Graph structure can better extract and represent the complex spatio-temporal relationships among multiple time series in forecasting tasks. Their accurate prediction, however, remains a challenge if each forecasting target is not only affected by the overall trend of the graph structure but also by the strong correlation among its local neighborhood nodes as the time series evolve. In this paper, we propose a forecasting method for time series based on Graph Clustering and Graph Neural Networks, named as the GCTGNN model. GCTGNN firstly clusters nodes of strong correlations within the graph’s local neighborhood, deriving multiple sub-graphs. Subsequently, it employs a graph neural network to extract the nodes’ feature information within each sub-graph. Then, a two-layer time series forecasting structure is introduced. The first layer learns the local change trend of each sub-graph over time which are then fused across sub-graphs through an attention mechanism, thereby deriving the global change features. The second layer then propagates these fused global features to each node within each sub-graph to obtain the forecasting results. Experimental results on different datasets show that GCTGNN outperforms other baseline models in the task of complex graph time series forecasting where local neighborhood nodes have strong correlations.}, archive = {J_NEUCOM}, author = {Xin Liu and Yapeng Meng and Feng Chen and Dengjian Qiao and Fan Wu}, doi = {10.1016/j.neucom.2025.129544}, journal = {Neurocomputing}, month = {4}, pages = {129544}, shortjournal = {Neurocomputing}, title = {GCTGNN: A forecasting method for time series based on graph neural networks and graph clustering}, volume = {626}, year = {2025}, }
• (2025). Transformer feature collapse of temporal action detection via multi-granularity semantic enhancement. NEUCOM, 626, 129543. (www)

  @article{ , abstract = {Transformer-based models for Temporal Action Detection have achieved significant performance improvements, where the Multi-Head Self-Attention (MHSA) mechanism has played a pivotal role. However, owing to MHSA’s tendency to map different patches into similar latent representations, existing methodologies are afflicted with the issue of temporal feature collapse, resulting in high similarity among temporal points and consequently increasing the difficulty in distinguishing action from background. To address the issue, we propose a Multi-granularity Semantic Enhancement (MSE) Block to learn multi-granularity semantic information from different feature spaces. The MSE Block comprises three core components: Local Discriminative Information Modeling (LDM), Global Temporal Information Modeling (GTM), and Adaptive Fusion Module (AFM). LDM facilitates the capture of discriminative information via a multi-scale convolutional group for local detail enhancement, subsequently, GTM employs the MHSA for global temporal context interaction, and AFM adaptively fuses all enhanced features to achieve multi-granularity semantic representation enhancement. Extensive experiments validate the superiority of our method, yielding state-of-the-art performance of 70.5% on THUMOS14, 39.3% on HACS, and 36.9% on ActivityNet-1.3. The code is released at https://github.com/XinAn9508/MSE_for_TAD .}, archive = {J_NEUCOM}, author = {Xin An and Peng Zhao and Guiqin Wang and Cong Zhao and Shusen Yang}, doi = {10.1016/j.neucom.2025.129543}, journal = {Neurocomputing}, month = {4}, pages = {129543}, shortjournal = {Neurocomputing}, title = {Transformer feature collapse of temporal action detection via multi-granularity semantic enhancement}, volume = {626}, year = {2025}, }
• (2025). LiDAR semantic segmentation with local consistency constrained KPConv LSTM. NEUCOM, 626, 129542. (www)

  @article{ , abstract = {As a fundamental task for autonomous driving, LiDAR point cloud semantic segmentation has been intensively studied in recent years. Despite the great progress, achieving satisfactory semantic segmentation is still very challenging due to the sparsity of LiDAR points and the shape diversity of the classes in the open world. In this paper we propose a local consistency constrained KPConv-LSTM module to benefit the existing methods. It enhances point-based LSTM with several KPConvs to strengthen and align the previous hidden features, thus improving the temporal feature propagation. A temporal weighting block is designed within the module to further reduce the error caused by the misalignment of moving objects. In addition, a special local consistency loss is proposed to encourage the local smoothness of the feature, thereby providing more consistent feature for temporal propagation in LSTM. We apply our method to various existing LiDAR semantic segmentation models. The experimental results on multiple datasets show that our method can produce notable improvements on all of them, validating the effectiveness and generality of the method.}, archive = {J_NEUCOM}, author = {Tingming Bai and Zhiyu Xiang and Xijun Zhao and Peng Xu and Tianyu Pu and Jingyun Fu}, doi = {10.1016/j.neucom.2025.129542}, journal = {Neurocomputing}, month = {4}, pages = {129542}, shortjournal = {Neurocomputing}, title = {LiDAR semantic segmentation with local consistency constrained KPConv LSTM}, volume = {626}, year = {2025}, }
• (2025). TPGRec: Text-enhanced and popularity-smoothing graph collaborative filtering for long-tail item recommendation. NEUCOM, 626, 129539. (www)

  @article{ , abstract = {GNN-based graph collaborative filtering methods have shown significant potential in recommendation systems, but they are often challenged by the long-tail effect due to exposure bias. While existing methods utilize techniques like contrastive learning, data augmentation and resampling as countermeasures, their reliance on ID-based embeddings can result in less informative representations and limit the model’s grasp of intricate neighbor relationships. Recent researches have attempted to improve overall recommendation performance by incorporating text information for items, but they usually rely on extra graph structures or complex calculations, increasing computational costs and lacking adequate consideration for long-tail items. In this paper, we propose TPGRec, a novel Graph collaborative filtering method jointly from the text enhancement and popularity smoothing perspectives, which simultaneously improves both overall and long-tail recommendation performance. Initially, we introduce a balancing mechanism applied to the graph structure and training set to reduce the influence of popular items. Upon this, a structural-level contrastive learning technique is proposed for graph representation learning, which captures complex structural relationships without introducing excessive noise to node representations. Furthermore, we develop a semantic-level contrastive learning strategy that effectively and economically integrates ID embeddings with textual data, establishing implicit semantic relationships and deepening the model’s understanding of items. Ultimately, we develop a popularity-balanced BPR optimization module to facilitate fair recommendation opportunities for items of varying popularity and promote the model’s discriminative power over hard negative samples. Comprehensive experiments on four real-world datasets have demonstrated the superiority of TPGRec compared with the state-of-the-art baselines. Our codes and datasets are available at Github: https://github.com/ycy89/MyTPGRec .}, archive = {J_NEUCOM}, author = {Chenyun Yu and Junfeng Zhao and Xuan Wu and Yingle Luo and Yan Xiao}, doi = {10.1016/j.neucom.2025.129539}, journal = {Neurocomputing}, month = {4}, pages = {129539}, shortjournal = {Neurocomputing}, title = {TPGRec: Text-enhanced and popularity-smoothing graph collaborative filtering for long-tail item recommendation}, volume = {626}, year = {2025}, }
• (2025). A survey of artificial intelligence in gait-based neurodegenerative disease diagnosis. NEUCOM, 626, 129533. (www)

  @article{ , abstract = {Recent years have witnessed an increasing global population affected by neurodegenerative diseases (NDs), which traditionally require extensive healthcare resources and human effort for medical diagnosis and monitoring. As a crucial disease-related motor symptom, human gait can be exploited to characterize different NDs. The current advances in artificial intelligence (AI) models enable automatic gait analysis for NDs identification and classification, opening a new avenue to facilitate faster and more cost-effective diagnosis of NDs. In this paper, we provide a comprehensive survey on recent progress of machine learning and deep learning based AI techniques applied to diagnosis of five typical NDs through gait. We provide an overview of the process of AI-assisted NDs diagnosis, and present a systematic taxonomy of existing gait data and AI models. Meanwhile, a novel quality evaluation criterion is proposed to quantitatively assess the quality of existing studies. Through an extensive review and analysis of 169 studies, we present recent technical advancements, discuss existing challenges, potential solutions, and future directions in this field. Finally, we envision the prospective utilization of 3D skeleton data for human gait representation and the development of more efficient AI models for NDs diagnosis. 2}, archive = {J_NEUCOM}, author = {Haocong Rao and Minlin Zeng and Xuejiao Zhao and Chunyan Miao}, doi = {10.1016/j.neucom.2025.129533}, journal = {Neurocomputing}, month = {4}, pages = {129533}, shortjournal = {Neurocomputing}, title = {A survey of artificial intelligence in gait-based neurodegenerative disease diagnosis}, volume = {626}, year = {2025}, }
• (2025). Text-guided multi-level interaction and multi-scale spatial-memory fusion for multimodal sentiment analysis. NEUCOM, 626, 129532. (www)

  @article{ , abstract = {With the rapid surge in multimodal data, multimodal sentiment analysis (MSA) has garnered increasing attention in recent years. It enhances the accuracy of sentiment polarity extraction by integrating information across diverse data modalities, thereby achieving comprehensive fusion of information and enhancing the precision of sentiment analysis. However, contemporary in-depth models often neglect the utilization of spacial and global memory information to assistant in emotion analysis. Furthermore, the predominant emphasis on textual modal tends to hinder the representation of information from visual and acoustic modalities. To mitigate these challenges, we propose T-MRMSM, a Text-guided Multi-level Representation Integration and Multi-scale Spacial-Memory Information Fusion model. More precisely, we introduce the three main modules: Scale-Global Information Extraction (S-GIE) module, Enhanced Memory (EM) module and Two Steps Cross-attention Fusion (2Steps-AT) module. The proposed model can ensure that extracting multi-level and multi-scale feature representation and enhance memory ability for the model, so as to provide the long-range and spatio information for intra-modal feature extraction and inter-modal representation fusion, and solve the dependence on text modes to a certain extent, while enhancing the influence factor of the other two modes. Extensive experiments on benchmark datasets (MOSI, MOSEI) demonstrate the significant gains of T-MRMSM over state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Xiaojiang He and Yanjie Fang and Nan Xiang and Zuhe Li and Qiufeng Wang and Chenguang Yang and Hao Wang and Yushan Pan}, doi = {10.1016/j.neucom.2025.129532}, journal = {Neurocomputing}, month = {4}, pages = {129532}, shortjournal = {Neurocomputing}, title = {Text-guided multi-level interaction and multi-scale spatial-memory fusion for multimodal sentiment analysis}, volume = {626}, year = {2025}, }
• (2025). Automatic visual enhancement of PTZ camera based on reinforcement learning. NEUCOM, 626, 129531. (www)

  @article{ , abstract = {Video surveillance systems have become indispensable for enhancing border security and effectively addressing threats such as illegal border crossings. Object detection plays a crucial role in real-time monitoring and event response in these systems. However, challenges arise in border scenarios, where targets exhibit small sizes, low resolutions, and limited extractable features, resulting in lower detection confidence. To overcome these limitations, we propose an advanced pan–tilt–zoom camera control method that does not require intrinsic camera parameters. The objective of this study is to accomplish visual enhancement tasks for low-confidence targets. The proposed method employs deep reinforcement learning techniques, integrating both discrete and continuous action spaces to enhance the generalization capability of agent decisions across diverse scenarios, thereby achieving optimal target monitoring. In addition, the introduction of the cutout feature fusion filter enables the agent to focus equally on each target in multitarget scenarios. In our experiments, we compared the proposed method with other approaches. The results demonstrate the superiority of the proposed method in various scenarios and object detection algorithms.}, archive = {J_NEUCOM}, author = {Hao Fang and Huanyu Liu and Jiazheng Wen and Zhonglin Yang and Junbao Li and Qi Han}, doi = {10.1016/j.neucom.2025.129531}, journal = {Neurocomputing}, month = {4}, pages = {129531}, shortjournal = {Neurocomputing}, title = {Automatic visual enhancement of PTZ camera based on reinforcement learning}, volume = {626}, year = {2025}, }
• (2025). A transformer-based dual contrastive learning approach for zero-shot learning. NEUCOM, 626, 129530. (www)

  @article{ , abstract = {The goal of zero-shot learning is to utilize attribute information for seen classes so as to generalize the learned knowledge to unseen classes. However, current algorithms often overlook the fact that the same attribute may exhibit different visual features across domains, leading to domain shift issues when transferring knowledge. Furthermore, in terms of visual feature extraction, networks like ResNet are not effective in capturing global information from images, adversely impacting recognition accuracy. To address these challenges, we propose an end-to-end Transformer-Based Dual Contrastive Learning Approach (TFDNet) for zero-shot learning. The network leverages the Vision Transformer (ViT) for extracting visual features and includes a mechanism for attribute localization to identify regions most relevant to the given attributes. Subsequently, it employs a dual contrastive learning method as a constraint, optimizing the learning process to better capture global feature representations. The proposed method makes the classifier more robust and enhances the ability to discriminate and generalize the unseen classes. Experimental results on three public datasets demonstrate the superiority of TFDNet over current state-of-the-art algorithms, validating its effectiveness in the field of zero-shot learning.}, archive = {J_NEUCOM}, author = {Yu Lei and Ran Jing and Fangfang Li and Quanxue Gao and Cheng Deng}, doi = {10.1016/j.neucom.2025.129530}, journal = {Neurocomputing}, month = {4}, pages = {129530}, shortjournal = {Neurocomputing}, title = {A transformer-based dual contrastive learning approach for zero-shot learning}, volume = {626}, year = {2025}, }
• (2025). Research on efficient data processing method for fog computing based on blockchain and federated learning. NEUCOM, 626, 129529. (www)

  @article{ , abstract = {Federated learning provides an effective distributed machine learning solution for "centralized" processing of decentralized data. However, the increasing size of data and complexity of learning models in the context of digitalization have placed higher demands on the computing power of training equipment. To address the above problems, a Fog Computing Blockchain Federated Learning (FC-BCFL) method is proposed to realize secure and efficient processing of massive data. Firstly, an incentive mechanism based on time reputation value and loss reputation value is designed with the help of blockchain to motivate clients with high-quality data to join the training and improve the accuracy of the model. Secondly, a fog node selection mechanism based on time reputation value is proposed and designed to select efficient fog nodes to train local models at the edge of the network using the perturbed data from clients processed by the local differential privacy technique in order to shorten the training time of federated learning and improve the training efficiency. In addition, a model aggregation LR-FedAvg algorithm is designed in this method, which weights the local model update parameters according to the loss reputation value to increase the weight of the high-precision local parameters in the global parameters so as to reduce the number of model training times and get the converged global optimal model faster. Finally, comparative experiments were conducted on the MNIST dataset, Fashion MNIST dataset, and CIFAR-10 dataset for three variables: training rounds, one iteration time, and total training time. It was concluded that the FC-BCFL method has been further optimized and improved in terms of model global accuracy, training rounds, and training time. This shows that the model can learn efficiently while ensuring data privacy, and a model with higher accuracy can be obtained with relatively fewer training rounds and iteration time. This verifies the effectiveness of the proposed FC-BCFL method.}, archive = {J_NEUCOM}, author = {Haiyan Kang and Bing Wu and Yiran Cao}, doi = {10.1016/j.neucom.2025.129529}, journal = {Neurocomputing}, month = {4}, pages = {129529}, shortjournal = {Neurocomputing}, title = {Research on efficient data processing method for fog computing based on blockchain and federated learning}, volume = {626}, year = {2025}, }
• (2025). Multi-level analyzation of imbalance to resolve non-IID-ness in federated learning. NEUCOM, 626, 129528. (www)

  @article{ , abstract = {Class imbalance is a common problem in deep learning that severely degrades performance. In federated learning (FL), it is a critical factor contributing to non-identically distributed data (non-IID). Building on several previous attempts, we define and analyze imbalance issues in FL at three levels: inter-case, inter-class, and inter-client. Inter-case imbalance addresses the imbalance in every single class; inter-class imbalance compares the number of data between different classes. Inter-client imbalance represents different skewness of local data between clients. Based on these concepts, we propose FedBB, which consists of two main components: (1) Positive Negative Balanced (PNB) loss function addresses the inter-case and inter-class imbalances in local training, enhancing generalization on highly skewed local client datasets. It optimizes both multi-label and multi-class classifications by assigning higher weights to minority cases or classes. (2) Client Balanced Reweighting (CBR) reweights clients based on inter-client imbalance during model aggregation, giving greater weight to models trained on less skewed datasets. Various experiments on X-ray and natural image datasets demonstrate that FedBB outperforms other algorithms in both performance and efficiency. Additionally, it requires limited statistical information, which is beneficial for privacy protection. Through ablation studies, we proved that PNB loss and CBR independently contribute to performance. As FedBB aims to build a global model that accurately classifies all classes, it can serve as a robust baseline for the generic and personalized FL.}, archive = {J_NEUCOM}, author = {Haengbok Chung and Jae Sung Lee}, doi = {10.1016/j.neucom.2025.129528}, journal = {Neurocomputing}, month = {4}, pages = {129528}, shortjournal = {Neurocomputing}, title = {Multi-level analyzation of imbalance to resolve non-IID-ness in federated learning}, volume = {626}, year = {2025}, }
• (2025). Robust data augmentation and contrast learning for debiased visual question answering. NEUCOM, 626, 129527. (www)

  @article{ , abstract = {The language prior problem in VQA causes the model to make predictions directly based on spurious correlations between questions and answers, causing the model’s performance to drop sharply outside the distribution. Current debiasing methods often achieve good out-of-distribution generalization capabilities at the expense of significant in-distribution performance degradation, while non-debiasing methods sacrifice a large amount of out-of-distribution performance to achieve high in-distribution performance. We propose a novel method combining multi-teacher knowledge distillation and contrastive learning (MKDCL) to solve the language prior problem in VQA. We propose a Question Answer Selection (QAS) module to select reasonable questions for images, which also determines the pseudo answers with multi-teacher’s weighted predictions. The Contrastive Learning Samples Generation (CLSG) module we propose synthesizes four types of positive and negative samples in visual and language modalities for contrastive learning, effectively increases the semantic dependency of the images while avoiding performance degradation due to spurious correlations between questions and answers. Our method is model-agnostic and achieves state-of-the-art performance (62.93%) on the language prior-sensitive VQA-CP v2 dataset while maintaining performance (65.43%) on the VQA v2 dataset.}, archive = {J_NEUCOM}, author = {Ke Ning and Zhixin Li}, doi = {10.1016/j.neucom.2025.129527}, journal = {Neurocomputing}, month = {4}, pages = {129527}, shortjournal = {Neurocomputing}, title = {Robust data augmentation and contrast learning for debiased visual question answering}, volume = {626}, year = {2025}, }
• (2025). DPCCNN: A new lightweight fault diagnosis model for small samples and high noise problem. NEUCOM, 626, 129526. (www)

  @article{ , abstract = {Obtaining large amounts of industrial data in real industrial processes is usually difficult. Meanwhile, it is difficult for deep learning-based lightweight fault diagnosis networks to obtain reliability diagnosis performance in the presence of high noise. To address these limitations, a new lightweight convolutional neural network (CNN) fault diagnosis framework, the dilated perceptually coupled convolutional neural network (DPCCNN), has been proposed for small samples and high noise problems. First, dilated layered interactive convolutional module (DLICM) is designed to obtain strong feature extraction capability, enhance the receptive field of small convolutional kernels by dilated convolution, and the self-attention mechanism compensates for the weak interactivity of deep convolution, which greatly reduces the number of parameters and computation of the model. Second, the global aggregation block (GAB) is designed to extract the contextual information of the feature map, which can focus on the basic contextual information without extensive computational requirements. The performance of this method is verified to be better than the current popular fault diagnosis models by the public dataset and the self-constructed rotor dataset, which still has good noise resistance in a lightweight framework and maintains high stability in small sample scenarios.}, archive = {J_NEUCOM}, author = {Jiabin Zhang and Zhiqian Zhao and Yinghou Jiao and Runchao Zhao and Xiuli Hu and Renwei Che}, doi = {10.1016/j.neucom.2025.129526}, journal = {Neurocomputing}, month = {4}, pages = {129526}, shortjournal = {Neurocomputing}, title = {DPCCNN: A new lightweight fault diagnosis model for small samples and high noise problem}, volume = {626}, year = {2025}, }
• (2025). Patch clustering-based spectral-transformer with multi-level fusion network for spectral reconstruction. NEUCOM, 626, 129525. (www)

  @article{ , abstract = {Spectral Reconstruction (SR) has attracted a lot of attention in recent years. A few studies have proposed spectral reconstruction techniques based on Transformers. However, the performance optimization of Transformers still is a vital issue. Furthermore, how to simultaneously explore the spatial and spectral information of hyperspectral image and fuse the features of different resolutions at different hierarchical levels is worth further investigation. Considering these issues, in this paper, we propose the Patch Clustering-Based Spectral-Transformer with Multi-level Fusion Network (PCFM). Specifically, to address the interaction of pixels with similar long-distance spectral information, a Patch-based Clustering Spectral Self-Attention (PCSS) mechanism is developed. After patching, each patch can focus on capturing local features, and clustering can reduce the computational complexity of processing large images. Furthermore, in order to fully consider the relationship between different clusters for PCSS, a Deal Conv Feed Attention (DCFA) module is designed, which adjusts the weight of the input data by calculating the correlation between clusters so that the model can better understand and handle these complex interactions. In addition, to address the various scales fusion issue, different from the previous methods, an Hierarchical Optimization Feature Fusion mechanism (HOFF) is introduced for merging images generated at different hierarchical levels. Finally, the Channel Relation Module (CRM) is employed to further bridge the gap between the reconstructed HSI and their ground truth. Extensive experimental results on the Pavia, Sandiga and Chikusei datasets indicate that this method outperforms the SOTA method in quantitative metrics and visual effects.}, archive = {J_NEUCOM}, author = {Changzhong Zou and Yikang Liu and Changwu Zou}, doi = {10.1016/j.neucom.2025.129525}, journal = {Neurocomputing}, month = {4}, pages = {129525}, shortjournal = {Neurocomputing}, title = {Patch clustering-based spectral-transformer with multi-level fusion network for spectral reconstruction}, volume = {626}, year = {2025}, }
• (2025). FLRNet: A bio-inspired three-stage network for camouflaged object detection via filtering, localization and refinement. NEUCOM, 626, 129523. (www)

  @article{ , abstract = {Camouflaged Object Detection (COD) aims to segment camouflaged objects, which poses a more challenging task than generic object detection due to the high intrinsic similarity between the foreground and background. In this paper, we propose a bio-inspired three-stage network to uniformly highlight the complete camouflaged objects with explicit boundaries by information filtering, global localization, and progressive refinement. Specifically, we propose the local filtering module (LFM) to mimic the filtering stage by eliminating redundant information using cross-level information to reduce background interference. Furthermore, we design a global localization module (GLM) to determine positions of potential camouflaged objects by directly and progressively aggregating filtered features across different levels. Finally, to fully capture discriminative subtle features, we employ the progressive refinement module (PRM) to distinguish the details and structures by adopting a dual-attention strategy, progressively correcting prediction errors. Quantitative and qualitative experimental results demonstrate that our proposed FLRNet outperforms the state-of-the-art COD methods.}, archive = {J_NEUCOM}, author = {Yilin Zhao and Qing Zhang and Yuetong Li}, doi = {10.1016/j.neucom.2025.129523}, journal = {Neurocomputing}, month = {4}, pages = {129523}, shortjournal = {Neurocomputing}, title = {FLRNet: A bio-inspired three-stage network for camouflaged object detection via filtering, localization and refinement}, volume = {626}, year = {2025}, }
• (2025). Joint conditional diffusion model for image restoration with mixed degradations. NEUCOM, 626, 129512. (www)

  @article{ , abstract = {Image restoration is rather challenging in adverse weather conditions, especially when multiple degradations occur simultaneously. Blind image decomposition was proposed to tackle this issue, however, its effectiveness heavily relies on the accurate estimation of each component. Although diffusion-based models exhibit strong generative abilities in image restoration tasks, they may generate irrelevant contents when the degraded images are severely corrupted. To address these issues, we leverage physical constraints to guide the whole restoration process, where a mixed degradation model based on atmosphere scattering model is constructed. Then we formulate our Joint Conditional Diffusion Model (JCDM) by incorporating the degraded image and degradation mask to provide precise guidance. To achieve better color and detail recovery results, we further integrate a refinement network to reconstruct the restored image, where Uncertainty Estimation Block (UEB) is employed to enhance the features. Extensive experiments performed on both multi-weather and weather-specific datasets demonstrate the superiority of our method over state-of-the-art competing methods. The code will be available at https://github.com/mengyu212/JCDM .}, archive = {J_NEUCOM}, author = {Yufeng Yue and Meng Yu and Luojie Yang and Tong Liu}, doi = {10.1016/j.neucom.2025.129512}, journal = {Neurocomputing}, month = {4}, pages = {129512}, shortjournal = {Neurocomputing}, title = {Joint conditional diffusion model for image restoration with mixed degradations}, volume = {626}, year = {2025}, }
• (2025). Art creator: Steering styles in diffusion model. NEUCOM, 626, 129511. (www)

  @article{ , abstract = {Large-scale text-to-image (T2I) generative models are extensively used in the art and creative industries because of their remarkable capability in generating high-quality images. The generation of ideal images in a single attempt is nearly impossible, necessitating complex and precise post-image editing. However, stylization pose significant challenges in post-editing. In this context, we introduce the Art Creator, which facilitates style controls based on a simple description or a single image. Art Creator enables nuanced image style edits, alterations in painting materials, colors, and brushstrokes, and understanding of high-level attributes such as object shapes. Furthermore, we manually annotated and released a dataset named ChinArt, comprising over 20,000 eastern artworks, aiming to address the gap in the global art creation domain. We showcase the quality and efficiency of our method across various art style creations.}, archive = {J_NEUCOM}, author = {Shan Tang and Wenhua Qian and Peng Liu and Jinde Cao}, doi = {10.1016/j.neucom.2025.129511}, journal = {Neurocomputing}, month = {4}, pages = {129511}, shortjournal = {Neurocomputing}, title = {Art creator: Steering styles in diffusion model}, volume = {626}, year = {2025}, }
• (2025). RCEAE: A role correlation-enhanced model for event argument extraction. NEUCOM, 626, 129504. (www)

  @article{ , abstract = {Event argument extraction (EAE) is an important information extraction task, which aims to retrieve arguments from texts and classify them into predefined argument roles. In practice, argument roles are usually annotated with numerous types, which often arises schema-specific nature and long-tail type nature. Existing studies explore the role correlations without full utilization, which may limit the capability of argument extraction. This paper investigates two crucial correlations potentially benefiting to the task, namely intra- and inter-event role correlations . The intra-event role correlations consider the role dependency within an event, thus helping to capture event schema-specific nature. The inter-event role correlations leverage the relevance among roles across events, thus helping to learn beneficial features from other roles. To achieve the above ideas, we propose RCEAE from a new role correlation-enhanced perspective. Particularly, for intra-event role correlations, we devise a prompt-based cross encoder to capture role correlations from an event prompt, and retrieve arguments considering event schema information. For inter-event role correlations, we devise a multi-view graph-based role encoder to build relevance for role representations, accessing to beneficial features not only from their training data but also from their related roles across different event types. By incorporating both the correlation knowledge, we predict event arguments with a role-specific interactive decoder. We conduct experiments on three public benchmarks, ACE, RAMS and WIKIEVENTS. Empirical results show RCEAE achieves state-of-the-art F1 on all benchmarks and demonstrates the effectiveness of incorporating both intra- and inter-event role correlations.}, archive = {J_NEUCOM}, author = {Yiming Hei and Jiawei Sheng and Shu Guo and Lihong Wang and Qian Li and Jianwei Liu and Yizhong Liu and Prayag Tiwari}, doi = {10.1016/j.neucom.2025.129504}, journal = {Neurocomputing}, month = {4}, pages = {129504}, shortjournal = {Neurocomputing}, title = {RCEAE: A role correlation-enhanced model for event argument extraction}, volume = {626}, year = {2025}, }
• (2025). Global mittag-leffler synchronization of discontinuous memristor-based fractional-order fuzzy inertial neural networks with mixed delays. NEUCOM, 626, 129475. (www)

  @article{ , abstract = {The stability analysis of most typical neural networks with mixed delays most bases on the Razumikhin theorem, where the considered models are relatively simple and the relevant results are not applicable to infinite distributed delay. In this paper, the global Mittag-Leffler synchronization (GMLS) of discontinuous memristor-based fractional-order fuzzy inertial neural networks (MFFINNs) with uncertainties and mixed delays is investigated. To deal with the inertia term, a variable replacement method is adopted to transform the original MFFINNs into two equivalent lower order subsystems. Compared to previous direct analysis method based on system itself, the order-reduced method retains the sequential property of Caputo fractional derivative and can be applied to other fractional-order nonlinear systems with higher order. The differential inclusion theory and set valued mapping are introduced for non-smooth analysis, and a generalized fractional-order Halanay inequality is introduced to handle with the infinite distributed delay. Based on Lyapunov stability theory and generalized inequality, two control strategies are established to achieve GMLS of MFFINNs. Ultimately, the practicability of the proposed method is verified through simulation results.}, archive = {J_NEUCOM}, author = {Haihe Pan and Chengdai Huang and Jinde Cao and Heng Liu}, doi = {10.1016/j.neucom.2025.129475}, journal = {Neurocomputing}, month = {4}, pages = {129475}, shortjournal = {Neurocomputing}, title = {Global mittag-leffler synchronization of discontinuous memristor-based fractional-order fuzzy inertial neural networks with mixed delays}, volume = {626}, year = {2025}, }
• (2025). IDEAL: Interpretable-by-design ALgorithms for learning from foundation feature spaces. NEUCOM, 626, 129464. (www)

  @article{ , abstract = {The advance of foundation models (FM) makes it possible to avoid parametric tuning for transfer learning, taking advantage of pretrained feature spaces. In this study, we define a framework called IDEAL (Interpretable-by-design DEep learning ALgorithms) which tackles the problem of interpretable transfer learning by recasting the standard supervised classification problem into a function of similarity to a set of prototypes derived from the training data. This framework generalises previously-known prototypical approaches, such as ProtoPNet, xDNN and DNC, and decomposes the overall problem into two inherently connected stages: (A) feature extraction (FE), which maps the raw features of real-world data into a latent space, and (B) identification of representative prototypes and decision making based on similarity and association between the query and the prototypes. This addresses the issue of interpretability (stage B) while retaining the benefits of pretrained deep learning (DL) models. On a range of datasets (CIFAR-10, CIFAR-100, CalTech101, STL-10, Oxford-IIIT Pet, EuroSAT), we demonstrate, through an extensive set of experiments, how the choice of the latent space, prototype selection, and finetuning of the latent space affect accuracy and generalisation of the models on transfer learning scenarios for different backbones. Building upon this knowledge, we demonstrate that the proposed framework helps achieve an advantage over state-of-the-art baselines in class-incremental learning. The key findings can be summarised as follows: (1) the setting allows interpretability through prototypes, (2) lack of finetuning helps circumvent the issue of catastrophic forgetting, allowing efficient class-incremental transfer learning, while mitigating the issue of confounding bias, and (3) ViT architectures narrow the gap between finetuned and non-finetuned models allowing for transfer learning in a fraction of time without finetuning of the feature space on a target dataset with iterative supervised methods.}, archive = {J_NEUCOM}, author = {Plamen Angelov and Dmitry Kangin and Ziyang Zhang}, doi = {10.1016/j.neucom.2025.129464}, journal = {Neurocomputing}, month = {4}, pages = {129464}, shortjournal = {Neurocomputing}, title = {IDEAL: Interpretable-by-design ALgorithms for learning from foundation feature spaces}, volume = {626}, year = {2025}, }
• (2025). Unsupervised representation learning with hebbian synaptic and structural plasticity in brain-like feedforward neural networks. NEUCOM, 626, 129440. (www)

  @article{ , abstract = {Neural networks that can capture key principles underlying brain computation offer exciting new opportunities for developing artificial intelligence and brain-like computing algorithms. Such networks remain biologically plausible while leveraging localized forms of synaptic learning rules and modular network architecture found in the neocortex. Compared to backprop-driven deep learning approches, they provide more suitable models for deployment of neuromorphic hardware and have greater potential for scalability on large-scale computing clusters. The development of such brain-like neural networks depends on having a learning procedure that can build effective internal representations from data. In this work, we introduce and evaluate a brain-like neural network model capable of unsupervised representation learning. It builds on the Bayesian Confidence Propagation Neural Network (BCPNN), which has earlier been implemented as abstract as well as biophysically detailed recurrent attractor neural networks explaining various cortical associative memory phenomena. Here we developed a feedforward BCPNN model to perform representation learning by incorporating a range of brain-like attributes derived from neocortical circuits such as cortical columns, divisive normalization, Hebbian synaptic plasticity, structural plasticity, sparse activity, and sparse patchy connectivity. The model was tested on a diverse set of popular machine learning benchmarks: grayscale images (MNIST, F-MNIST), RGB natural images (SVHN, CIFAR-10), QSAR (MUV, HIV), and malware detection (EMBER). The performance of the model when using a linear classifier to predict the class labels fared competitively with conventional multi-layer perceptrons and other state-of-the-art brain-like neural networks.}, archive = {J_NEUCOM}, author = {Naresh Ravichandran and Anders Lansner and Pawel Herman}, doi = {10.1016/j.neucom.2025.129440}, journal = {Neurocomputing}, month = {4}, pages = {129440}, shortjournal = {Neurocomputing}, title = {Unsupervised representation learning with hebbian synaptic and structural plasticity in brain-like feedforward neural networks}, volume = {626}, year = {2025}, }
• (2025). Interpretable modelling and visualization of biomedical data. NEUCOM, 626, 129405. (www)

  @article{ , abstract = {Applications of interpretable machine learning (ML) techniques on medical datasets facilitate early and fast diagnoses, along with getting deeper insight into the data. Furthermore, the transparency of these models increase trust among application domain experts. Medical datasets face common issues such as heterogeneous measurements, imbalanced classes with limited sample size, and missing data, which hinder the straightforward application of ML techniques. In this paper we present a family of prototype-based (PB) interpretable models which are capable of handling these issues. Moreover we propose a strategy of harnessing the power of ensembles while maintaining the intrinsic interpretability of the PB models, by averaging over the model parameter manifolds. All the models were evaluated on a synthetic (publicly available dataset) in addition to detailed analyses of two real-world medical datasets (one publicly available). The models and strategies we introduce address the challenges of real-world medical data, while remaining computationally inexpensive and transparent. Moreover, they exhibit similar or superior in performance compared to alternative techniques.}, archive = {J_NEUCOM}, author = {S. Ghosh and E.S. Baranowski and M. Biehl and W. Arlt and P. Tiňo and K. Bunte}, doi = {10.1016/j.neucom.2025.129405}, journal = {Neurocomputing}, month = {4}, pages = {129405}, shortjournal = {Neurocomputing}, title = {Interpretable modelling and visualization of biomedical data}, volume = {626}, year = {2025}, }
• (2025). MW-FixMatch: A class imbalance semi-supervised learning algorithm based on re-weighting. NEUCOM, 626, 129385. (www)

  @article{ , abstract = {Semi-supervised learning for image classification is an important research area in computer vision. These algorithms typically assume that both labeled and unlabeled datasets are class-balanced and share the same distribution. However, when there is an imbalance in the class distribution, it can significantly affect their performance. To address this issue, we propose MW-FixMatch, a novel approach that better adjusts the semi-supervised learning process in the presence of class imbalance. It utilizes a weight network to balance the contribution of labeled and unlabeled data, and the parameters of this network are learned from a class-balanced sampled set. We tested our approach on several publicly available image datasets with class imbalance and consistently achieved superior results across multiple experiments.}, archive = {J_NEUCOM}, author = {Xiaoqing Zheng and Weijie Hong and Dengde Chen and Anke Xue and Yaguang Kong}, doi = {10.1016/j.neucom.2025.129385}, journal = {Neurocomputing}, month = {4}, pages = {129385}, shortjournal = {Neurocomputing}, title = {MW-FixMatch: A class imbalance semi-supervised learning algorithm based on re-weighting}, volume = {626}, year = {2025}, }
• (2025). A comprehensive review of network pruning based on pruning granularity and pruning time perspectives. NEUCOM, 626, 129382. (www)

  @article{ , abstract = {The prevalence of deep learning has resulted in the widespread deployment of deep neural networks. However, due to the explosive growth in data volume and advancements in hardware device performance, network models have also undergone a surge in both depth and width. The pursuit of model compression has gained significant traction, driven by the need to deploy models on edge and mobile devices. Network pruning is a straightforward yet highly effective method of model compression, accomplished by directly eliminating redundant components within the network. Various pruning algorithms have emerged in the past few years, especially since the birth of the Lottery Ticket Hypothesis, the traditional three-step pruning paradigm has been broken, resulting in many new pruning paradigms. Consequently, our primary focus in this review centers on recent pruning literature based on pruning time. Furthermore, we have extensively examined works related to the Lottery Ticket Hypothesis. Our comprehensive review demonstrates that both pruning at initialization and pruning during training can produce promising results, while also alleviating the significant resource and time investments usually associated with the iterative training and fine-tuning inherent in traditional pruning methods. In conclusion, we have undertaken a thorough exploration of the pruning domain and provided insights into potential avenues for the advancement of pruning algorithms.}, archive = {J_NEUCOM}, author = {Kehan Zhu and Fuyi Hu and Yuanbing Ding and Wei Zhou and Ruxin Wang}, doi = {10.1016/j.neucom.2025.129382}, journal = {Neurocomputing}, month = {4}, pages = {129382}, shortjournal = {Neurocomputing}, title = {A comprehensive review of network pruning based on pruning granularity and pruning time perspectives}, volume = {626}, year = {2025}, }
• (2025). Mixture of experts models for multilevel data: Modeling framework and approximation theory. NEUCOM, 626, 129357. (www)

  @article{ , abstract = {Multilevel data are prevalent in many real-world applications. However, it remains an open research problem to identify and justify a class of models that flexibly capture a wide range of multilevel data. Motivated by the versatility of the mixture of experts (MoE) models in fitting regression data, in this article we extend upon the MoE and study a class of mixed MoE (MMoE) models for multilevel data. Under some regularity conditions, we prove that the MMoE is dense in the space of any continuous mixed effects models in the sense of weak convergence. As a result, the MMoE has a potential to accurately resemble almost all characteristics inherited in multilevel data, including the marginal distributions, dependence structures, regression links, random intercepts and random slopes. In a particular case where the multilevel data is hierarchical, we further show that a nested version of the MMoE universally approximates a broad range of dependence structures of the random effects among different factor levels.}, archive = {J_NEUCOM}, author = {Tsz Chai Fung and Spark C. Tseung}, doi = {10.1016/j.neucom.2025.129357}, journal = {Neurocomputing}, month = {4}, pages = {129357}, shortjournal = {Neurocomputing}, title = {Mixture of experts models for multilevel data: Modeling framework and approximation theory}, volume = {626}, year = {2025}, }
• (2025). Warped convolutional neural networks for large homography transformation with psl(3) algebra. NEUCOM, 626, 129312. (www)

  @article{ , abstract = {Homography has a fundamental and elegant relationship with the projective special linear group and its embedding Lie algebra structure. However, the integration of homography and algebraic expressions in neural networks remains largely unexplored. In this paper, we propose Warped Convolution Neural Networks to effectively learn and represent the homography by psl ( 3 ) algebra with group convolution. Specifically, six commutative subgroups within the P S L ( 3 ) group are composed to form a homography. For each subgroup, a warp function is proposed to bridge the Lie algebra structure to its corresponding parameters in homography. By taking advantage of the warped convolution, homography learning is formulated into several simple pseudo-translation regressions. Our proposed method enables learning features that are invariant to significant homography transformations through exploration along the Lie topology. Moreover, it can be easily plugged into other popular CNN-based methods and empower them with homography representation capability. Through extensive experiments on benchmark datasets such as POT, S-COCO, and MNIST-Proj, we demonstrate the effectiveness of our approach in various tasks like classification, homography estimation, and planar object tracking.}, archive = {J_NEUCOM}, author = {Xinrui Zhan and Wenyu Liu and Risheng Yu and Jianke Zhu and Yang Li}, doi = {10.1016/j.neucom.2024.129312}, journal = {Neurocomputing}, month = {4}, pages = {129312}, shortjournal = {Neurocomputing}, title = {Warped convolutional neural networks for large homography transformation with psl(3) algebra}, volume = {626}, year = {2025}, }
• (2025). Enhancing music audio signal recognition through CNN-BiLSTM fusion with de-noising autoencoder for improved performance. NEUCOM, 625, 129607. (www)

  @article{ , abstract = {This study presents an advanced framework for music audio signal recognition that combines Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, and Noise Reduction Auto-encoder models to significantly improve accuracy and robustness. The core innovation is a novel noise reduction auto-encoder that integrates CNN and BiLSTM architectures, enabling superior recognition performance under varying noise levels and environmental conditions. The proposed framework, validated on several datasets including the Zhvoice, Common Voice, and LibriSpeech, demonstrates higher accuracy compared to existing methods. In addition, an optimized CNN architecture called Faster Region-based CNN with Multi-scale Information (FRCNN-MSI) is developed for efficient speech feature extraction, which shows significant improvements in noisy environments. The BiLSTM model is further enhanced with an attention mechanism that improves sequence modeling and contextual relationship capture. Together, these advances establish our approach as a robust solution to real-world speech recognition challenges, with potential implications for improving speech recognition systems in diverse applications.}, archive = {J_NEUCOM}, author = {Xiaoying Mao and Ye Tian and Tairan Jin and Bo Di}, doi = {10.1016/j.neucom.2025.129607}, journal = {Neurocomputing}, month = {4}, pages = {129607}, shortjournal = {Neurocomputing}, title = {Enhancing music audio signal recognition through CNN-BiLSTM fusion with de-noising autoencoder for improved performance}, volume = {625}, year = {2025}, }
• (2025). MPLR-CapsNet: A novel capsule network with multi-line parallel features and logical reasoning for occluded pedestrian detection. NEUCOM, 625, 129595. (www)

  @article{ , abstract = {Occluded pedestrian detection is a key technology in computer vision. The parts of pedestrians covered are random, and diverse clothes, postures, and varying scales make occluded pedestrian detection a challenging task. In order to solve the above issues, we propose a novel capsule network model, MPLR-CapsNet, based on multi-line parallel features and logical reasoning, aiming to achieve accurate detection of occluded pedestrians in dense scenes. This network model introduces RS module to learn diversified key feature information in a directional manner, and use Mix-squash function to enhance the anti-interference and generalization abilities of the network model. Furthermore, the local pruning strategy is adopted to retain only the important low-level capsules connected to high-level capsules in the dynamic routing stage, effectively reducing the training parameters of the network model. In addition, we propose a capsule reasoning module based on the hyper-sausage metric model in the reasoning detection stage, transforming the object category and coordinates into the attribute representations of capsule entities, which significantly enhance the autonomous reasoning and discriminant abilities of the network model. Extensive experiments conducted on two standard occluded pedestrian datasets validate the advancement of the proposed method, outperforming existing state-of-the-art solutions. Our model achieves Top-1 accuracies of 97.6 % and 94.5 % on the CityPersons and CrowdHuman datasets, respectively.}, archive = {J_NEUCOM}, author = {Jingwei Cao and Guoyang Hou and Jiawang Lv and Tianlin Gao and Liming Di and Chengtao Zhang}, doi = {10.1016/j.neucom.2025.129595}, journal = {Neurocomputing}, month = {4}, pages = {129595}, shortjournal = {Neurocomputing}, title = {MPLR-CapsNet: A novel capsule network with multi-line parallel features and logical reasoning for occluded pedestrian detection}, volume = {625}, year = {2025}, }
• (2025). PEVA-net: Prompt-enhanced view aggregation network for zero/few-shot multi-view 3D shape recognition. NEUCOM, 625, 129590. (www)

  @article{ , abstract = {Large vision–language models have impressively promote the performance of 2D visual recognition under zero/few-shot scenarios. In this paper, we focus on exploiting the large vision–language model, i.e., CLIP, to address zero/few-shot 3D shape recognition based on multi-view representations. The key challenge for both tasks is to generate a discriminative descriptor of the 3D shape represented by multiple view images under the scenarios of either without explicit training (zero-shot 3D shape recognition) or training with a limited number of data (few-shot 3D shape recognition). We analyze that both tasks are relevant and can be considered simultaneously. Specifically, leveraging the descriptor which is effective for zero-shot inference to guide the tuning of the aggregated descriptor under the few-shot training can significantly improve the few-shot learning efficacy. Hence, we propose Prompt-Enhanced View Aggregation Network (PEVA-Net) to simultaneously address zero/few-shot 3D shape recognition. Under the zero-shot scenario, we propose to leverage the prompts built up from candidate categories to enhance the aggregation process of multiple view-associated visual features. The resulting aggregated feature serves for effective zero-shot recognition of the 3D shapes. Under the few-shot scenario, we first exploit a ViT encoder to aggregate the view-associated visual features into a global descriptor. To tune the encoder, we propose a self-feature-distillation scheme via a feature distillation loss by treating the zero-shot descriptor as the guidance signal for the few-shot descriptor. This scheme can significantly enhance the few-shot learning efficacy. Without any pre-training process, our PEVA-Net can produce the state-of-the-art zero-shot 3D shape recognition performance on ModelNet40, ModelNet10 and ShapeNetCore 55 datasets with the accuracy of 84.48% , 93.50% and 74.65% . Under the 16-shot setting of ModelNet40, the proposed PEVA-Net also sets the state-of-the-art recognition accuracy of 90.64% . Extensive ablation experiments are conducted to analyze the superiority of the proposed PEVA-Net.}, archive = {J_NEUCOM}, author = {Dongyun Lin and Yi Cheng and Aiyuan Guo and Shangbo Mao and Yiqun Li}, doi = {10.1016/j.neucom.2025.129590}, journal = {Neurocomputing}, month = {4}, pages = {129590}, shortjournal = {Neurocomputing}, title = {PEVA-net: Prompt-enhanced view aggregation network for zero/few-shot multi-view 3D shape recognition}, volume = {625}, year = {2025}, }
• (2025). Advances in instance segmentation: Technologies, metrics and applications in computer vision. NEUCOM, 625, 129584. (www)

  @article{ , abstract = {Instance segmentation is an advanced technique in computer vision that focuses on identifying and classifying each individual object in an image at the pixel level. Unlike semantic segmentation, which groups pixels of similar objects without distinguishing between different instances, instance segmentation assigns unique labels to each object, even if they are of the same class. This makes it possible not only to detect the presence and category of objects in an image but also to locate each specific instance and clearly distinguish them from each other. This problem not only advances the technical and theoretical understanding of how machines see and process digital images, but also has a direct impact on various industries and sectors where computer vision is an essential part of the system. In this paper, we present the current deep learning-based technologies, the metrics used for their evaluation, and a review of general and concrete datasets in general and drone-specific contexts. The results of this study provide a compendium of easily deployable deep learning-based technologies. This review paper aims to accelerate the process of understanding and using instance segmentation technologies for the reader.}, archive = {J_NEUCOM}, author = {José M. Molina and Juan P. Llerena and Luis Usero and Miguel A. Patricio}, doi = {10.1016/j.neucom.2025.129584}, journal = {Neurocomputing}, month = {4}, pages = {129584}, shortjournal = {Neurocomputing}, title = {Advances in instance segmentation: Technologies, metrics and applications in computer vision}, volume = {625}, year = {2025}, }
• (2025). ViT-CAPS: Vision transformer with contrastive adaptive prompt segmentation. NEUCOM, 625, 129578. (www)

  @article{ , abstract = {Real-time segmentation plays an important role in numerous applications, including autonomous driving and medical imaging, where accurate and instantaneous segmentation influences essential decisions. The previous approaches suffer from the lack of cross-domain transferability and the need for large amounts of labeled data that prevent them from being applied successfully to real-world scenarios. This study presents a new model, ViT-CAPS, that utilizes Vision Transformers in the encoder to improve segmentation performance in challenging and large-scale scenes. We employ the Adaptive Context Embedding (ACE) module, incorporating contrastive learning to improve domain adaptation by matching features from support and query images. Also, the Meta Prompt Generator (MPG) is designed to generate prompts from aligned features, and it can segment in complicated environments without requiring much human input. ViT-CAPS has shown promising results in resolving domain shift problems and improving few-shot segmentation in dynamic low-annotation settings. We conducted extensive experiments on four well-known datasets, FSS-1000, Cityscapes, ISIC, and DeepGlobe, and achieved noteworthy performance. We achieved a performance gain of 4.6 % on FSS-1000, 4.2 % on DeepGlobe, 6.1 % on Cityscapes, and a slight difference of −3 % on the ISIC dataset compared to previous approaches. We achieved an average mean IoU of 60.52 and 69.3, which is 2.7 % and 5.1 % higher accuracy over state-of-the-art Cross-Domain Few-Shot Segmentation (CD-FSS) models on 1-shot and 5-shot settings respectively.}, archive = {J_NEUCOM}, author = {Khawaja Iftekhar Rashid and Chenhui Yang}, doi = {10.1016/j.neucom.2025.129578}, journal = {Neurocomputing}, month = {4}, pages = {129578}, shortjournal = {Neurocomputing}, title = {ViT-CAPS: Vision transformer with contrastive adaptive prompt segmentation}, volume = {625}, year = {2025}, }
• (2025). Permutation invariant self-attention infused U-shaped transformer for medical image segmentation. NEUCOM, 625, 129577. (www)

  @article{ , abstract = {The size and shape of organs in the human body vary according to factors like genetics, body size, proportions, health, lifestyle, gender, ethnicity, and race. Further, abnormalities due to cancer and chronic diseases also affect the size of organs and tumors. Moreover, the spatial location and area of these organs deviates along the transverse plane (Z plane) of the medical scans. Therefore, the generalizability and robustness of a computer vision framework over medical images can be improved if the framework is also encouraged to learn representations of the target areas regardless of their spatial location in input images. Hence, we propose a novel permutation invariant multi-headed self-attention (PISA) module to reduce a U-shaped transformer-based architecture Swin-UNet’s sensitivity towards permutation. We have infused this module in the skip connection of our architecture. We have achieved a mean dice score of 79.25 on the segmentations of 8 abdominal organs, better than most state-of-the-art algorithms. Moreover, we have analyzed the generalizability of our architecture over publicly available multi-sequence cardiac MRI datasets. When tested over a sequence unseen by the model during training, 25.1 % and 9.0 % improvement in dice scores were observed in comparison to the pure-CNN-based algorithm and pure transformer-based architecture, respectively, thereby demonstrating its versatility. Replacing the Self Attention module in a U-shaped transformer architecture with our Permutation Invariant Self Attention module produced noteworthy segmentations over shuffled test images, even though the module was trained solely on normal images. The results demonstrate the enhanced efficiency of the proposed module in imparting attention to target organs irrespective of their spatial positions.}, archive = {J_NEUCOM}, author = {Sanjeet S. Patil and Manojkumar Ramteke and Anurag S. Rathore}, doi = {10.1016/j.neucom.2025.129577}, journal = {Neurocomputing}, month = {4}, pages = {129577}, shortjournal = {Neurocomputing}, title = {Permutation invariant self-attention infused U-shaped transformer for medical image segmentation}, volume = {625}, year = {2025}, }
• (2025). SPADesc: Semantic and parallel attention with feature description. NEUCOM, 625, 129567. (www)

  @article{ , abstract = {Local feature detection and description are essential preliminary tasks in a multitude of computer vision applications. Despite the prowess of deep neural networks in feature extraction, they still grapple with challenges in capturing globally invariant and robust features, especially in dynamic scenes and areas with simplistic and repetitive geometric structures. This paper introduces a multi-scale feature fusion framework, SPADesc, which addresses these challenges by leveraging dynamic weighted fusion (DWF) and semantic priors. We integrate convolutional and self-attention mechanisms to bolster local feature detection and description in complex environments. Our approach employs a Parallel Convolution and Attention (PCA) module to generate descriptors that encompass both local and global scales. Additionally, a Semantic-Guided (SG) module is employed to produce class-aware global mask information, which implicitly guides the selection of keypoints and descriptors. By incorporating a Semantically Weighted (SW) loss function, we enhance the robustness and discriminative power of the descriptors. Extensive experimental results across various visual tasks demonstrate significant performance improvements, highlighting the superior adaptability and precision of our proposed model. The code is available at https://github.com/Diffcc/SPADesc .}, archive = {J_NEUCOM}, author = {Haijun Meng and Huimin Lu and Bozhi Ding and Qiangchang Wang}, doi = {10.1016/j.neucom.2025.129567}, journal = {Neurocomputing}, month = {4}, pages = {129567}, shortjournal = {Neurocomputing}, title = {SPADesc: Semantic and parallel attention with feature description}, volume = {625}, year = {2025}, }
• (2025). GPT4TFP: Spatio-temporal fusion large language model for traffic flow prediction. NEUCOM, 625, 129562. (www)

  @article{ , abstract = {Traffic flow prediction aims to anticipate the future usage levels of transportation, and is a pivotal component of intelligent transportation systems. Previous studies have mainly employed deep learning technologies to decode traffic flow data. These methods process the spatial and temporal embeddings of traffic flow data in a sequential, parallel, or single-feature manner. Although the structures of these models are becoming more and more complex, their accuracy has not improved. Recently, large language models (LLMs) have made significant progress in traffic flow prediction tasks due to their superior performance. However, although the spatio-temporal dependencies of traffic flow prediction can be captured by LLMs, they ignore the cross-relationships between spatio-temporal embeddings. To this end, we propose a spatio-temporal fusion large language model (GPT4TFP) for traffic flow prediction, which is divided into four components: the spatio-temporal embedding layer, the spatio-temporal fusion layer, the frozen pre-trained LLM layer, and the output linear layer. The spatio-temporal embedding layer embeds traffic flow data into the spatio-temporal representations required by traffic flow prediction. In the spatio-temporal fusion layer, we propose a spatio-temporal fusion strategy based on multi-head cross-attention to capture the cross-relationships between spatio-temporal embeddings. In addition, we introduce a frozen pre-trained strategy to fine-tune the LLM to improve the accuracy of traffic flow prediction. The experimental results on two traffic flow datasets show that the proposed model outperforms a set of state-of-the-art baseline models.}, archive = {J_NEUCOM}, author = {Yiwu Xu and Mengchi Liu}, doi = {10.1016/j.neucom.2025.129562}, journal = {Neurocomputing}, month = {4}, pages = {129562}, shortjournal = {Neurocomputing}, title = {GPT4TFP: Spatio-temporal fusion large language model for traffic flow prediction}, volume = {625}, year = {2025}, }
• (2025). DKETFormer: Salient object detection in optical remote sensing images based on discriminative knowledge extraction and transfer. NEUCOM, 625, 129558. (www)

  @article{ , abstract = {Generally, most methods for salient object detection in optical remote sensing images (ORSI-SOD) are based on convolutional neural networks (CNNs). However, CNNs, due to their architectural characteristics, can only encode local semantic information, which leads to a lack of exploration of discriminative features on a large scale. Therefore, to encode the long-term dependency within the detection image, enhance the extraction of discriminative knowledge, and transfer it at multiple scales, we introduce a Transformer architecture called DKETFormer. Specifically, DKETFormer utilizes the Transformer backbone to obtain multi-scale feature maps that have encoded long-term dependency relationships. Then, it constructs a decoder using the Cross-spatial Knowledge Extraction Module (CKEM) and the Inter-layer Feature Transfer Module (IFTM). The CKEM is capable of extracting discriminative information across receptive fields while preserving knowledge from each channel. It also utilizes global information encoding to calibrate channel weights, resulting in improved knowledge aggregation and capturing of pixel-level pairwise relationships. The IFTM utilizes encoded and extracted information from the backbone and CKEM, employing a self-attention mechanism with cosine similarity knowledge to model and propagate discriminative features. Finally, we generated the final detection map using a salient object detector. The results of comparative experiments and ablation experiments demonstrate the effectiveness of the proposed DKETFormer and its internal modules.}, archive = {J_NEUCOM}, author = {Yuze Sun and Hongwei Zhao and Jianhang Zhou}, doi = {10.1016/j.neucom.2025.129558}, journal = {Neurocomputing}, month = {4}, pages = {129558}, shortjournal = {Neurocomputing}, title = {DKETFormer: Salient object detection in optical remote sensing images based on discriminative knowledge extraction and transfer}, volume = {625}, year = {2025}, }
• (2025). Bilinear self-representation for unsupervised feature selection with structure learning. NEUCOM, 625, 129557. (www)

  @article{ , abstract = {Current feature selection approaches based on self-representation focus solely on representing either the sample space or the feature space. Consequently, a common challenge in such methods arises from the limited interaction between the feature and sample spaces, which can potentially result in incomplete discovery of hidden data information. To address this challenge, this paper introduces a novel mixture-level self-representation method called Bilinear Self-representation for Unsupervised Feature Selection with Structure Learning (BSUFSL). The newly proposed method, BSUFSL, distinguishes itself from existing feature selection approaches based on self-representation by utilizing the concept of bilinearity. This allows BSUFSL to capture self-representation in both the feature and sample spaces, integrating information from both domains and establishing an effective interaction between them to identify the most significant features. Meanwhile, BSUFSL incorporates local preservation to maintain geometric structure among features and global preservation to maximize data variance, both of which are essential aspects of structure learning. An efficient iterative optimization algorithm with theoretically guaranteed convergence is also developed to solve the BSUFSL model. Experimental results on eight benchmark datasets demonstrate the superior effectiveness of BSUFSL compared to state-of-the-art methods. Specifically, the numerical results reveal a significant improvement in clustering performance across several datasets. Furthermore, results from experiments confirm that BSUFSL enhances the clarity and accuracy of clustering by selecting discriminative features, thanks to its bilinear structure, which integrates feature and sample self-representation into a unified framework.}, archive = {J_NEUCOM}, author = {Hossein Nasser Assadi and Faranges Kyanfar and Farid Saberi-Movahed and Abbas Salemi}, doi = {10.1016/j.neucom.2025.129557}, journal = {Neurocomputing}, month = {4}, pages = {129557}, shortjournal = {Neurocomputing}, title = {Bilinear self-representation for unsupervised feature selection with structure learning}, volume = {625}, year = {2025}, }
• (2025). Crossing the chasm: A practical architecture augmentation for low-quality object detection. NEUCOM, 625, 129553. (www)

  @article{ , abstract = {The detection of objects in low-quality environments, such as underwater and low-light conditions, poses significant challenges due to poor visibility and complex degradations. Traditional object detectors exhibit suboptimal performance in these scenarios. We rethink the challenges of low-quality detection and design the A2HL architecture to address three key challenges. To the best of our knowledge, this is the first work to propose a general architectural enhancement strategy aimed at extending any conventional detector. The A2HL framework integrates three innovative modules: the Feature Augmentation Module, which boosts the feature representation capability of conventional detectors in low-quality environments; the Redundancy Elimination Module, which effectively removes redundant information during feature extraction; and the Shape Feature Extraction Module, which improves the detection of small and shape-dependent objects. Extensive experiments on underwater and low-light datasets demonstrate that A2HL significantly outperforms existing methods. The proposed architecture achieves notable improvements in mean Average Precision (mAP), with increases of over 5% and 10% for underwater and low-light scenes, respectively. The results affirm the effectiveness of A2HL in adapting conventional detectors to low-quality environments, providing a robust solution for practical applications in challenging detection scenarios.}, archive = {J_NEUCOM}, author = {Xinwei Xue and Haoze Zheng and Yuechao Gao and Tengyu Ma and Long Ma and Qi Jia}, doi = {10.1016/j.neucom.2025.129553}, journal = {Neurocomputing}, month = {4}, pages = {129553}, shortjournal = {Neurocomputing}, title = {Crossing the chasm: A practical architecture augmentation for low-quality object detection}, volume = {625}, year = {2025}, }
• (2025). A multi-scale graph-guided dynamic enhanced alignment network for mechanical fault diagnosis considering domain shift and data imbalance. NEUCOM, 625, 129546. (www)

  @article{ , abstract = {Transfer learning is widely applied in the cross-domain diagnosis of machines, often under the assumption of balanced data. However, machines mostly operate in the normal state in practical industry, and only limited fault samples can be collected before maintenance shutdowns, leading to class imbalance. This imbalance, combined with the inherent domain shift between the source and target domains, makes diagnosis challenging, including the neglect of minority fault classes and misalignment between domains. To address this, a Multi-scale graph-guided Dynamic enhanced alignment network (Mul-Dean) is proposed. Mul-Dean consists of two main components. One component, the Multi-scale Graph Construction Method (MGCM), tackles intra-domain imbalance by learning diverse and distinguishable relationships between classes. The other component, the Dynamic Enhancement Alignment Loss (DEAL) algorithm, reduces the sensitivity to proportion disparities of imbalance by introducing class-contrastive labels. Additionally, the inter-domain distance is proposed to quantify the alignment degree between domains. Experimental results on the bearing and reciprocating compressor datasets showed that Mul-Dean achieved better classification and alignment capability than the other 11 methods.}, archive = {J_NEUCOM}, author = {Xiaoxuan Fan and Lixiang Duan and Na Zhang}, doi = {10.1016/j.neucom.2025.129546}, journal = {Neurocomputing}, month = {4}, pages = {129546}, shortjournal = {Neurocomputing}, title = {A multi-scale graph-guided dynamic enhanced alignment network for mechanical fault diagnosis considering domain shift and data imbalance}, volume = {625}, year = {2025}, }
• (2025). A collaborative network via multi-head sparse and high-low frequency interaction for hyperspectral image classification. NEUCOM, 625, 129541. (www)

  @article{ , abstract = {Hyperspectral image (HSI) classification is an essential technique in hyperspectral remote sensing applications, and Transformers have shown great potential in this field. However, factors such as the distribution of land covers, spectral variability, and high dimensionality often cause many Transformers to lose information related to inter-class heterogeneity and boundary details. To address these issues, we propose a dual-branch Transformer network that combines sparsification and high-low frequency interaction, aiming to address challenges such as diverse surface cover and spectral feature mining with lower computational cost. The sparse Transformer branch is used to extract multi-scale local–global dependency features from HSIs, preserving fine-grained and wide-range contextual information. The high-low frequency interaction Transformer branch is combined to reveal the spectral response differences of objects in different spectral bands. Deep interaction between the two parallel branches is achieved by perceiving channel correlations and spatial positions. Extensive experiments on four public HSI datasets demonstrate the effectiveness and competitiveness of the proposed method, and the code is made publicly available at https://github.com/youngboy03/MSFI-CNet .}, archive = {J_NEUCOM}, author = {Qikang Liu and Shuaishuai Fan and Weiming Li and Songjie Wei and Yonghua Jiang}, doi = {10.1016/j.neucom.2025.129541}, journal = {Neurocomputing}, month = {4}, pages = {129541}, shortjournal = {Neurocomputing}, title = {A collaborative network via multi-head sparse and high-low frequency interaction for hyperspectral image classification}, volume = {625}, year = {2025}, }
• (2025). Predicted label distribution guided optimal transport for classification of remote sensing smoke-like scenes with noisy labels. NEUCOM, 625, 129540. (www)

  @article{ , abstract = {Remote sensing (RS) images have been used in a large amount of applications due to their wide coverage and good timeliness. For fire monitoring, smoke in RS images is regarded as the indicator of fire disasters and it is of great significance to recognize RS smoke. However, some RS scenes are visually similar to smoke, e.g, cloud, dust and haze. The high inter-class similarities not only make classification of these smoke-like scenes a tough task, but also will lead to the appearance of inaccurately labeled samples when collecting corresponding datasets. The supervised classification task relies highly on accurately labeled samples, and models will overfit to noisy labels when training on inaccurately labeled datasets. In this work, the optimal transport (OT) theory is utilized to deal with noisy labels in classification of RS smoke-like scenes. Served as a loss function, OT objective enables closer distance between samples which are mapped with each other in the transportation plan. In scene classification, each sample has a class label and the OT objective could be computed from both the feature space and the label space. Considering the existence of noisy labels, we use predicted label distributions to amplify the dissimilarities between feature representations, thereby reducing mis-mappings in the OT transportation plan. The proposed algorithm is named as predicted label distribution guided optimal transport (PLDGOT) and an extensive experiments show PLDGOT achieves remarkable performances in classification of RS smoke-like scenes with noisy labels.}, archive = {J_NEUCOM}, author = {Shikun Chen and Xin Lu and Chang Lu and Xiaobo Lu}, doi = {10.1016/j.neucom.2025.129540}, journal = {Neurocomputing}, month = {4}, pages = {129540}, shortjournal = {Neurocomputing}, title = {Predicted label distribution guided optimal transport for classification of remote sensing smoke-like scenes with noisy labels}, volume = {625}, year = {2025}, }
• (2025). When Brain–Computer interfaces meet the metaverse: Landscape, demonstrator, trends, challenges, and concerns. NEUCOM, 625, 129537. (www)

  @article{ , abstract = {The metaverse has gained tremendous popularity in recent years, allowing the interconnection of users worldwide. However, current systems in metaverse scenarios, such as virtual reality glasses, offer a partial immersive experience. In this context, Brain–Computer Interfaces (BCIs) can introduce a revolution in the metaverse, although a study of the applicability and implications of BCIs in these virtual scenarios is required. Based on a limited number of publications, this work reviews the applicability of BCIs in the metaverse, analyzing the current status of this integration based on different categories related to virtual worlds and the evolution of BCIs in these scenarios in the medium and long term. This work also proposes the design and implementation of a general framework that integrates BCIs with different data sources from sensors and actuators (e.g., VR glasses) based on a modular design to be easily extended. This manuscript also validates the framework in a demonstrator consisting of driving a car within a metaverse, using a BCI for neural data acquisition, a VR headset to provide realism, and a steering wheel and pedals. Four use cases (UCs) are selected, focusing on cognitive and emotional assessment of the driver, detection of drowsiness, and driver authentication while using the vehicle. The results demonstrate the applicability of BCIs to metaverse scenarios using the proposed framework, achieving over 80% F1-score for all UCs, with performance close to 100% for detecting emotions and authenticating users. Moreover, this manuscript offers an analysis of BCI trends in the metaverse, also identifying future challenges that the intersection of these technologies will face. Finally, it reviews the concerns that using BCIs in virtual world applications could generate according to different categories: accessibility, user inclusion, privacy, cybersecurity, physical safety, and ethics.}, archive = {J_NEUCOM}, author = {Sergio López Bernal and Mario Quiles Pérez and Enrique Tomás Martínez Beltrán and Gregorio Martínez Pérez and Alberto Huertas Celdrán}, doi = {10.1016/j.neucom.2025.129537}, journal = {Neurocomputing}, month = {4}, pages = {129537}, shortjournal = {Neurocomputing}, title = {When Brain–Computer interfaces meet the metaverse: Landscape, demonstrator, trends, challenges, and concerns}, volume = {625}, year = {2025}, }
• (2025). VCF: An effective vision-centric framework for visual question answering. NEUCOM, 625, 129536. (www)

  @article{ , abstract = {Recently, the wide application of large language models in the field of Visual Question Answering(VQA) has significantly boosted the progress in this field. Despite achieved advancements, LLMs cannot fully perceive and comprehend visual information well from image. Therefore, how to fully mine visual information is very important for language models to effectively deal with the VQA task. In response to this challenge, we propose a straightforward yet effective Vision-Centric Framework(VCF) for VQA, which mainly includes an adaptive visual perceptron module, a multi-source feature fusion module, and a large language model. The adaptive visual perceptron module effectively condenses and integrates the extensive visual information sequence from the visual encoder output using a fixed number of query embeddings. The multi-source feature fusion module is concentrated on extracting fine-grained visual perception information by fusing visual features of different scales. Finally, by channeling their outputs, the language model leverages its extensive implicit knowledge to produce a more nuanced and precise synthesis of visual information, ultimately delivering the answer. The synergy and complementarity of the two modules jointly enhance the robustness of the model. Through extensive experiments, VCF achieves nearly state-of-the-art experimental results on datasets such as VQAv2, OK-VQA, GQA, Text-VQA and others. At the same time, a series of ablation experiments have been conducted to demonstrate the efficacy of the proposed module. Additionally, VCF even achieves better or equivalent performance compared to some larger-scale models, such as LLaVa-1.5, Pink.}, archive = {J_NEUCOM}, author = {Fengjuan Wang and Longkun Peng and Shan Cao and Zhaoqilin Yang and Ruonan Zhang and Gaoyun An}, doi = {10.1016/j.neucom.2025.129536}, journal = {Neurocomputing}, month = {4}, pages = {129536}, shortjournal = {Neurocomputing}, title = {VCF: An effective vision-centric framework for visual question answering}, volume = {625}, year = {2025}, }
• (2025). ATPTrack: Visual tracking with alternating token pruning of dynamic templates and search region. NEUCOM, 625, 129534. (www)

  @article{ , abstract = {Constantly varying appearance of targets brings tremendous challenges for visual object tracking, especially in long-term tracking and background interference scenarios. Current leading trackers attempt to introduce multi-level fixed and dynamic templates to encode changing target information and abundant spatiotemporal contexts. These methods achieve astonishing performance improvements. However, dynamic templates are obtained from intermediate frames with high tracking confidence scores. These frames are not manually annotated. Therefore, the accuracy of dynamic templates relies entirely on tracking results. Dynamic templates may contain a large amount of uninformative and irrelevant background noise due to imprecise tracking. Additionally, multiple templates result in increased computational complexity as well. In order to tackle the above problems, a novel tracker dubbed ATPTrack is proposed for efficient end-to-end tracking. ATPTrack combines the initial target information of fixed templates and updated target representations of dynamic templates. Particularly, ATPTrack develops an alternating token trimming method that prunes dynamic templates and search region progressively. After token simplification, target-related information is highlighted in both dynamic templates and search region, which makes it more computationally efficient. Furthermore, a novel similarity ranking module is incorporated into the dynamic template pruning (DTP) block and the search region pruning (SRP) block for selecting the most discriminative target tokens. The DTP and SRP blocks are responsible for efficient trimming of dynamic templates and the search region, separately. The proposed ATPTrack achieves competitive performance on multiple tracking benchmarks. Compared to merely trimming the search region, ATPTrack further reduces the multiply-accumulate operations (MACs) by 11.5 % with negligible performance drop of 0.3 % by alternately pruning dynamic templates and search region.}, archive = {J_NEUCOM}, author = {Shuo Zhang and Dan Zhang and Qi Zou}, doi = {10.1016/j.neucom.2025.129534}, journal = {Neurocomputing}, month = {4}, pages = {129534}, shortjournal = {Neurocomputing}, title = {ATPTrack: Visual tracking with alternating token pruning of dynamic templates and search region}, volume = {625}, year = {2025}, }
• (2025). TSCnet: A text-driven semantic-level controllable framework for customized low-light image enhancement. NEUCOM, 625, 129509. (www)

  @article{ , abstract = {Deep learning-based image enhancement methods show significant advantages in reducing noise and improving visibility in low-light conditions. These methods are typically based on one-to-one mapping, where the model learns a direct transformation from low light to specific enhanced images. Therefore, these methods are inflexible as they do not allow highly personalized mapping, even though an individual’s lighting preferences are inherently personalized. To overcome these limitations, we propose a new light enhancement task and a new framework that provides customized lighting control through prompt-driven, semantic-level, and quantitative brightness adjustments. The framework begins by leveraging a Large Language Model (LLM) to understand natural language prompts, enabling it to identify target objects for brightness adjustments. To localize these target objects, the Retinex-based Reasoning Segment (RRS) module generates precise target localization masks using reflection images. Subsequently, the Text-based Brightness Controllable (TBC) module adjusts brightness levels based on the generated illumination map. Finally, an Adaptive Contextual Compensation (ACC) module integrates multi-modal inputs and controls a conditional diffusion model to adjust the lighting, ensuring seamless and precise enhancements accurately. Experimental results on benchmark datasets demonstrate our framework’s superior performance at increasing visibility, maintaining natural color balance, and amplifying fine details without creating artifacts. Furthermore, its robust generalization capabilities enable complex semantic-level lighting adjustments in diverse open-world environments through natural language interactions. Project page is https://miaorain.github.io/lowlight09.github.io/ .}, archive = {J_NEUCOM}, author = {Miao Zhang and Jun Yin and Pengyu Zeng and Yiqing Shen and Shuai Lu and Xueqian Wang}, doi = {10.1016/j.neucom.2025.129509}, journal = {Neurocomputing}, month = {4}, pages = {129509}, shortjournal = {Neurocomputing}, title = {TSCnet: A text-driven semantic-level controllable framework for customized low-light image enhancement}, volume = {625}, year = {2025}, }
• (2025). Attention-modulated frequency-aware pooling via spatial guidance. NEUCOM, 625, 129507. (www)

  @article{ , abstract = {Pooling is widely used in computer vision to expand the receptive field and enhance semantic understanding by reducing spatial resolution. However, current mainstream downsampling methods primarily rely on local spatial aggregation. While they effectively reduce the spatial resolution of feature maps and extract discriminative features, they are still limited by the constraints of the receptive field and the inadequacy of single-domain information, making it challenging to effectively capture fine details while suppressing noise. To address these limitations, we propose a Dual-Domain Downsampling (D3) method, which leverages the complementarity of spatial and frequency domains. We employ an invertible local two-dimensional Discrete Cosine Transform (2D DCT) transformation to construct a frequency domain pooling window. In the spatial domain, we design an Inverted Multiform Attention Modulator (IMAM) that expands the receptive field through multiform convolutions, while adaptively constructing dynamic frequency weights guided by rich spatial information. This allows for fine-grained modulation of different frequency components, either amplifying or attenuating them in different spatial regions, effectively reducing noise while preserving detail. Extensive experiments on ImageNet-1K, MSCOCO, and complex scene detection datasets across various benchmark models consistently validate the effectiveness of our approach. On the ImageNet-1K classification task, our method achieve up to a 1.95% accuracy improvement, with significant performance gains over state-of-the-art methods on MSCOCO and other challenging detection scenarios. The code will be made publicly available at: https://github.com/HZAI-ZJNU/D3 .}, archive = {J_NEUCOM}, author = {Yunzhong Si and Huiying Xu and Xinzhong Zhu and Rihao Liu and Hongbo Li}, doi = {10.1016/j.neucom.2025.129507}, journal = {Neurocomputing}, month = {4}, pages = {129507}, shortjournal = {Neurocomputing}, title = {Attention-modulated frequency-aware pooling via spatial guidance}, volume = {625}, year = {2025}, }
• (2025). Hybrid lyrebird red panda Optimization_Shepard convolutional neural network for recognition of speech emotion in audio signals. NEUCOM, 625, 129506. (www)

  @article{ , abstract = {Speech serves as the primary mode of human communication, where semantic meaning is conveyed through the combination and arrangement of words. Recent research in Speech Emotion Recognition (SER) has assisted in maintaining and improving social relationships and behaviors among individuals. In recent times, several advancements have been attained in SER systems due to the incorporation of Deep learning (DL) models. However, the conventional techniques often require large, well-annotated datasets for effective training, which was resource-intensive to collect and label. Moreover, models may struggle to generalize across diverse speakers, emotional expressions, and recording conditions, potentially limiting their real-world applicability. Therefore, this paper presents a Lyrebird Red Panda Optimization _Shepard Convolutional Neural Network (LRPO_ShCNN) for SER. Initially, the input speech signal is preprocessed by using Adaptive Gaussian filter. After that, the significant features from the preprocessed image are extracted. Further, data augmentation process is carried out and it generates new data points from existing data. After that, feature selection is done with LRPO. Finally, SER is accomplished by utilizing ShCNN, where the emotions are classified. Moreover, the hyperparameters of ShCNN are tuned with LRPO, which is developed by the integration of Lyrebird Optimization Algorithm (LOA) and Red Panda Optimization (RPO). The evaluation results shows that the LRPO_ShCNN obtained Accuracy, Positive Predictive Value (PPV), Negative Predictive Value (NPV), True Positive Rate (TPR), and True Negative Rate (TNR) as 91.092 %, 90.552 %, 90.876 %, 91.230 %, and 91.818 % respectively.}, archive = {J_NEUCOM}, author = {Kanimozhi N. and Devi Priya R.}, doi = {10.1016/j.neucom.2025.129506}, journal = {Neurocomputing}, month = {4}, pages = {129506}, shortjournal = {Neurocomputing}, title = {Hybrid lyrebird red panda Optimization_Shepard convolutional neural network for recognition of speech emotion in audio signals}, volume = {625}, year = {2025}, }
• (2025). A conversational ambiguity resolution model for contextual knowledge embedding. NEUCOM, 625, 129505. (www)

  @article{ , abstract = {In multi-turn dialogue, there are often ambiguous phenomena such as pronoun referencing and omission of language components, resulting in incomplete discourse information and ambiguous language understanding. This leads the dialogue system unable to produce accurate responses information. This situation usually requires more dialogue history information to understand the correct intention of the dialogue process. To address this issue, we propose a dialogue ambiguity resolution model based on contextual knowledge, which utilizes commonsense knowledge and domain knowledge to obtain relevant knowledge subgraphs through a dynamic weight parameter model. By utilizing historical dialogue information to evaluate the association level between knowledge triplets and current dialogue context, the association score is used as the basis for searching for context-associated relational knowledge. This kind of knowledge can be served as a reasoning path, thereby helping to better resolve ambiguity in conversation. In order to fully integrate the context-associated relational knowledge and the original dialogue information, while preserving the original dialogue feature information, we design a knowledge gate regulator that blends the last layer of hidden feature vectors of the original dialogue context with the feature vectors generated by the information aggregation layer. For the ambiguity resolution task, we achieve state-of-the-art accuracy on both the MuDoCo and SB-TOP datasets, with F1-scores of 96.07 (+2.95 higher than previous work) and 97.95 (+1.15 higher than previous work), respectively. For the coreference resolution task on the WSC dataset, our method improves the performance of the same base model by 4.8%, demonstrating a significant advantage when the number of parameters is comparable.}, archive = {J_NEUCOM}, author = {Kai Li and Jietao Liu and Yanshan Yang}, doi = {10.1016/j.neucom.2025.129505}, journal = {Neurocomputing}, month = {4}, pages = {129505}, shortjournal = {Neurocomputing}, title = {A conversational ambiguity resolution model for contextual knowledge embedding}, volume = {625}, year = {2025}, }
• (2025). Mean-square stability of stochastic clifford-valued cohen-grossberg neural networks with variable coefficients and time-varying delays. NEUCOM, 625, 129500. (www)

  @article{ , abstract = {In this paper, the mean-square stability of the stochastic Clifford-valued Cohen-Grossberg neural networks (SCVCGNNs) with variable coefficients and multiple time-varying delays is considered. Rather than disassembling the SCVCGNNs into real-valued or complex-valued counterparts, this paper treats the networks holistically. We propose a novel adequate condition via the fixed point theorem to achieve mean-square stability, and the obtained results also hold for real, complex and quaternion stochastic neural networks. Lastly, the effectiveness of the proposed criterion is illustrated with two numerical simulation examples.}, archive = {J_NEUCOM}, author = {Mingbing Luo and Zili Jiang and Manchun Tan}, doi = {10.1016/j.neucom.2025.129500}, journal = {Neurocomputing}, month = {4}, pages = {129500}, shortjournal = {Neurocomputing}, title = {Mean-square stability of stochastic clifford-valued cohen-grossberg neural networks with variable coefficients and time-varying delays}, volume = {625}, year = {2025}, }
• (2025). Global attention network with rain prior for real time single image deraining. NEUCOM, 625, 129490. (www)

  @article{ , abstract = {Poor visibility caused by rainy image can have a negative impact on the performance of computer vision applications. While several image deraining algorithms have been popularly adopted, most of them suffer from two main limitations: (1) they cannot well handle real and complex rain scenes by only focusing on one type of rain in images (e.g. raindrops or rain streaks) whereas the reality often coexists with both types, (2) they face significant difficulties in practical application because of ignoring the speed of inference. To address the above problems, we propose a global attention network (GANet) that can quickly and effectively separate rain streaks and raindrops. Inspired by the fact that rain in images often appears white, we leverage this prior to obtain an initial rain-free background image to guide neural network-based image deraining. Moreover, a new global attention block (GAB) is designed to simultaneously extract the rain features from spatial and channel dimensions. By cascading multiple GABs, the proposed method can effectively obtain the features of rain streaks and raindrops and progressively separates the rain-free image. Furthermore, owing to the nonlinear properties of GAB, the activation functions are omitted, which can speed up the inference time. And the depth-wise and point-wise convolutions are employed to promote computation efficiency as well. Extensive experiments on raindrop and rain streak datasets demonstrate that our method outperforms state-of-the-art methods, achieving up to 37.53 dB PSNR on Rain100L with an inference speed of 39 FPS, which is 2–30 times faster than competitors.}, archive = {J_NEUCOM}, author = {Yuan Yuan and Xuanbin Guo and Dandan Ma}, doi = {10.1016/j.neucom.2025.129490}, journal = {Neurocomputing}, month = {4}, pages = {129490}, shortjournal = {Neurocomputing}, title = {Global attention network with rain prior for real time single image deraining}, volume = {625}, year = {2025}, }
• (2025). RTLinearFormer: Semantic segmentation with lightweight linear attentions. NEUCOM, 625, 129489. (www)

  @article{ , abstract = {Real-time semantic segmentation aims to balance networks’ segmentation accuracy and inference speed. Recently, dual-resolution network shows its efficiency and effectiveness in this field. However, due to self-attention’s O( N 2 ) computational complexity, pure CNN-based networks still dominate in this architecture. In this paper, we propose RTLinearFormer, an efficient dual-resolution network with Transformer specially for real-time semantic segmentation, which achieves trade-off between segmentation accuracy and inference speed. To this end, our RTLinearFormer leverages lightweight channel-wise multi-scale attention with linear complexity, channel feature focusing and multi-scale features to extract global contextual information in low-resolution branch and uses cross-resolution ReLU-based linear attention with linear complexity to gather global contextual information derived from low-resolution branch with detailed information in high-resolution branch. Extensive experiments demonstrate the effectiveness and efficiency of our proposed RTLinearFormer in real-time semantic segmentation task. The model achieves 78.41% val mIoU and 66.7 FPS on Cityscapes using a single RTX 3090 GPU. When pretrained on Cityscapes, it achieves 77.4% test mIoU and 141.3 FPS on CamVid. It shows promising results on Cityscapes and CamVid, demonstrating superior performance over other state-of-the-art models. To further validate its generalization ability, we also test it on the ADE20K and COCOStuff datasets. It achieves 35.70% val mIoU and 107.3 FPS on ADE20K, and 31.15% test mIoU and 150.2 FPS on COCOStuff, setting new benchmarks for state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Yuhang Gu and Chong Fu and Wei Song and Xingwei Wang and Junxin Chen}, doi = {10.1016/j.neucom.2025.129489}, journal = {Neurocomputing}, month = {4}, pages = {129489}, shortjournal = {Neurocomputing}, title = {RTLinearFormer: Semantic segmentation with lightweight linear attentions}, volume = {625}, year = {2025}, }
• (2025). Learnable bi-directional data augmentation for few-shot cross-domain point cloud classification. NEUCOM, 625, 129486. (www)

  @article{ , abstract = {This paper studies the 3D Few-Shot Domain Adaptation (FSDA) task, which aims to adapt models from one domain to another using only a few labeled target point clouds. Compared with the widely adopted Unsupervised Domain Adaptation (UDA) problem setting, which assumes a large number of unlabeled target data is available, obtaining few-shot labeled target data costs less. For 3D point clouds, local geometric structure information is important in domain adaptation, making previous high-level global feature alignment-based methods used in 2D FSDA no longer suitable. Therefore, we propose to use point cloud augmentation to enhance geometric diversity and design a Learnable Bi-directional Data Augmentation (LBDA) approach to tackle the 3D FSDA task. Our proposed LBDA enhances source and target domain data, transforming them into easily transferable representations that bridge the domain gap by a proposed ellipsoid-based loss constraint. Additionally, we present a Sample-level Adaptive Weighting (SAW) method, allowing the DA model to selectively focus more on readily transferable samples and vice versa, thus mitigating the risk of negative transfer. Our experimental results on the PointDA-10 dataset and Sim-to-Real dataset showcase the effectiveness of our approach in the 5-shot scenario, outperforming state-of-the-art UDA methods. This paper contributes to the burgeoning field of point cloud analysis, offering a promising solution to the 3D Few-Shot Domain Adaptation problem.}, archive = {J_NEUCOM}, author = {Xu Sun and Lin Zhang and Jiahao Xu and Jiakang Yuan and Tao Chen}, doi = {10.1016/j.neucom.2025.129486}, journal = {Neurocomputing}, month = {4}, pages = {129486}, shortjournal = {Neurocomputing}, title = {Learnable bi-directional data augmentation for few-shot cross-domain point cloud classification}, volume = {625}, year = {2025}, }
• (2025). Micro-expression recognition in wild video environments: Latent feature-based ANN (LFANN) from 3D reconstructed faces. NEUCOM, 625, 129480. (www)

  @article{ , abstract = {Facial micro-expression (ME) recognition has garnered substantial interest in both computer vision and psychology for its ability to detect nonverbal emotions through subtle, involuntary facial muscle movements. Despite this potential, the practical deployment and commercialization of ME recognition systems have been hindered by the lack of comprehensive datasets, leaving these systems underdeveloped and limiting their application in real-world scenarios. This study addresses these challenges by introducing a novel ME recognition system specifically designed for natural environments and evaluated using the in-the-wild benchmark dataset, MEVIEW. The proposed system, namely the Latent Feature-based ANN (LFANN), begins with 3D facial reconstruction to align faces, effectively handling challenges such as head rotation, occlusions, and inconsistent lighting. The system further identifies multiple peak frames based on landmark coordinates in predefined facial regions, ensuring a discriminative representation of the video. The displacement of facial muscle movements across consecutive frames is captured via landmark coordinate shifts and transformed into a latent space for improved feature representation. This end-to-end proposed framework is capable of achieving notable performance metrics, including 70% accuracy, 70.83% F1-score, 75.56% UAR, and 67.07% UF1. Additionally, 75% of the detected peak frames fall within the onset and offset periods, validating the effectiveness of the peak frame detection method. Comprehensive evaluations and ablation studies highlight the significance of each system component, positioning this approach as a key advancement in ME recognition study.}, archive = {J_NEUCOM}, author = {Y.S. Gan and Kun-Hong Liu and Gen-Bing Liong and Sze-Teng Liong}, doi = {10.1016/j.neucom.2025.129480}, journal = {Neurocomputing}, month = {4}, pages = {129480}, shortjournal = {Neurocomputing}, title = {Micro-expression recognition in wild video environments: Latent feature-based ANN (LFANN) from 3D reconstructed faces}, volume = {625}, year = {2025}, }
• (2025). Mdatagen: A python library for the artificial generation of missing data. NEUCOM, 625, 129478. (www)

  @article{ , abstract = {Missing data is characterized by the presence of absent values in data (i.e., missing values) and it is currently categorized into three different mechanisms: Missing Completely at Random, Missing At Random, and Missing Not At Random. When performing missing data experiments and evaluating techniques to handle absent values, these mechanisms are often artificially generated (a process referred to as data amputation ) to assess the robustness and behavior of the used methods. Due to the lack of a standard benchmark for data amputation, different implementations of the mechanisms are used in related research (some are often not disclaimed), preventing the reproducibility of results and leading to an unfair or inaccurate comparison between existing and new methods. Moreover, for users outside the field, experimenting with missing data or simulating the appearance of missing values in real-world domains is unfeasible, impairing stress testing in machine learning systems. This work introduces mdatagen , an open source Python library for the generation of missing data mechanisms across 20 distinct scenarios, following different univariate and multivariate implementations of the established missing mechanisms. The package therefore fosters reproducible results across missing data experiments and enables the simulation of artificial missing data under flexible configurations, making it very versatile to mimic several real-world applications involving missing data. The source code and detailed documentation for mdatagen are available at https://github.com/ArthurMangussi/pymdatagen .}, archive = {J_NEUCOM}, author = {Arthur Dantas Mangussi and Miriam Seoane Santos and Filipe Loyola Lopes and Ricardo Cardoso Pereira and Ana Carolina Lorena and Pedro Henriques Abreu}, doi = {10.1016/j.neucom.2025.129478}, journal = {Neurocomputing}, month = {4}, pages = {129478}, shortjournal = {Neurocomputing}, title = {Mdatagen: A python library for the artificial generation of missing data}, volume = {625}, year = {2025}, }
• (2025). Large language model augmented syntax-aware domain adaptation method for aspect-based sentiment analysis. NEUCOM, 625, 129472. (www)

  @article{ , abstract = {Cross-domain aspect-based sentiment analysis aims to leverage knowledge from the source domain to identify the sentiment polarity towards a given aspect attribute in the text content from the target domain. Existing domain adaptation approaches either focus on acquiring domain-independent shared feature representations or adjusting the obtained feature distribution to the target domain, which fails to address critical domain-specific attributes, leading to misaligned feature representations. We propose a large language model augmented syntax-aware domain adaptation method that integrates advanced large language models with structured syntactic knowledge to recognize semantic attributes and address the lack of syntax sensitivity in large language models. A domain topic predictor based on adversarial training is developed to enhance the robustness and generalization of the framework across different domains. Additionally, automatic soft prompt learning is conducted based on analysed domain topics and task-relevant feature representations for domain-specific fine-tuning, aiding the architecture in conveying domain-specific semantic information in the cross-domain environment. The feature aggregation approach dynamically fuses six categories of analysed feature representations for fine-grained sentiment classification. To the best of our knowledge, this study represents the pioneering effort to systematically leverage syntactic and cross-domain characteristics to enhance pre-trained large language models in addressing cross-domain aspect-based sentiment analysis tasks. Experimental results on publicly available benchmark datasets validate the effectiveness of the proposed architecture.}, archive = {J_NEUCOM}, author = {Haochen Zou and Yongli Wang}, doi = {10.1016/j.neucom.2025.129472}, journal = {Neurocomputing}, month = {4}, pages = {129472}, shortjournal = {Neurocomputing}, title = {Large language model augmented syntax-aware domain adaptation method for aspect-based sentiment analysis}, volume = {625}, year = {2025}, }
• (2025). OTL-CE: Online transfer learning for data streams with class evolution. NEUCOM, 625, 129470. (www)

  @article{ , abstract = {Learning from data streams has become a crucial area of research due to the continuous evolution of data sources. One of the key challenges is concept drift, where data distributions shift over time. However, real-world data streams not only experience concept drift but also undergo class evolution, where new classes appear, and existing ones vanish. Most existing approaches focus on adapting to concept drift, overlooking the possibility of class evolution in real-world scenarios, which limits their performance in practical applications. To address this issue, a class-based ensemble method, namely online transfer learning for data streams with class evolution (OTL-CE), is proposed. Upon the arrival of a new data chunk, OTL-CE employs a One-vs-All strategy to build a binary classifier for each class within the data chunk, subsequently adding it to the respective target sub-ensemble. To effectively transfer knowledge from heterogeneous source domains, a density-based correlation alignment method is introduced to align the distribution of data with the same class label. Following this, a source sub-ensemble is constructed based on the aligned source domain instances. Finally, knowledge from source and target domains is integrated by weighted combining the corresponding class sub-ensembles, where the weight of each sub-ensemble is determined by its prediction accuracy on the latest data chunk. The experimental results demonstrate that OTL-CE achieves superior performance across most datasets under four evaluation metrics: Accuracy, Balanced Accuracy, Recall, and G-mean. In particular, it outperforms other algorithms on the CovTYPE dataset, delivering a notable improvement of at least 5.8% in G-mean. Additionally, the results confirm that the improved density-based correlation alignment method can effectively alleviates negative knowledge transfer.}, archive = {J_NEUCOM}, author = {Botao Jiao and Shihui Liu}, doi = {10.1016/j.neucom.2025.129470}, journal = {Neurocomputing}, month = {4}, pages = {129470}, shortjournal = {Neurocomputing}, title = {OTL-CE: Online transfer learning for data streams with class evolution}, volume = {625}, year = {2025}, }
• (2025). AMT-net: Attention-based multi-task network for scene depth and semantics prediction in assistive navigation. NEUCOM, 625, 129468. (www)

  @article{ , abstract = {Traveling safely and independently in unfamiliar environments remains a significant challenge for people with visual impairments. Conventional assistive navigation systems, while aiming to enhance spatial awareness, typically handle crucial tasks like semantic segmentation and depth estimation separately, resulting in high computational overhead and reduced inference speed. To address this limitation, we introduce AMT-Net, a novel multi-task deep neural network designed for joint semantic segmentation and monocular depth estimation. The AMT-Net is designed with a single unified decoder, which boosts not only the model’s efficiency but also its scalability on portable devices with limited computational resources. We propose two self-attention-based modules, CSAPP and RSAB, to leverage the strengths of convolutional neural networks for extracting robust local features and Transformers for capturing essential long-range dependencies. This design enhances the ability of our model to interpret complex scenes effectively. Furthermore, AMT-Net has low computational complexity and achieves real-time performance, making it suitable for assistive navigation applications. Extensive experiments on the public NYUD-v2 dataset and the TrueSight dataset demonstrated our model’s state-of-the-art performance and the effectiveness of the proposed components.}, archive = {J_NEUCOM}, author = {Yunjia Lei and Joshua Luke Thompson and Son Lam Phung and Abdesselam Bouzerdoum and Hoang Thanh Le}, doi = {10.1016/j.neucom.2025.129468}, journal = {Neurocomputing}, month = {4}, pages = {129468}, shortjournal = {Neurocomputing}, title = {AMT-net: Attention-based multi-task network for scene depth and semantics prediction in assistive navigation}, volume = {625}, year = {2025}, }
• (2025). Instruction-guided path planning with 3D semantic maps for vision-language navigation. NEUCOM, 625, 129457. (www)

  @article{ , abstract = {The goal of vision-language navigation (VLN) is to empower an agent to understand natural, human-like instructions and successfully navigate through an environment to execute them. This task requires the integration of language understanding, visual perception, and decision-making capabilities. In this paper, we propose a novel approach to VLN that incorporates an additional pre-exploration stage—a setting that, while uncommon in the literature, we believe is quite natural for many practical applications. Specifically, we propose an instruction-aware Path Proposal module and a Discrimination model (iPPD) to leverage pre-extracted 3D semantic information. iPPD generates instruction-aware path proposals , which are instruction-aligned candidate paths that help reduce the solution space. To better align modalities and represent map observations along a path, we propose a novel Path Feature Encoding scheme tailored for semantic maps. Furthermore, we design an attention-based Language Driven Discriminator to evaluate path candidates and select the best path as the final result. Compared to single-step greedy decision methods, our method reasons about the global path, thanks to the information extracted during pre-exploration, which naturally avoids error accumulation. Extensive experiments demonstrate that the effective use of global information from the pre-exploration stage can boost the performance of the language-guided navigation task with less training effort. Our code is available at https://github.com/zehao-wang/iPPD-sem .}, archive = {J_NEUCOM}, author = {Zehao Wang and Mingxiao Li and Minye Wu and Marie-Francine Moens and Tinne Tuytelaars}, doi = {10.1016/j.neucom.2025.129457}, journal = {Neurocomputing}, month = {4}, pages = {129457}, shortjournal = {Neurocomputing}, title = {Instruction-guided path planning with 3D semantic maps for vision-language navigation}, volume = {625}, year = {2025}, }
• (2025). Accelerating hyperparameter optimization with a secretary. NEUCOM, 625, 129455. (www)

  @article{ , abstract = {Hyperparameter optimization (HPO) is a vital step in machine learning (ML) for enhancing model performance. However, the vast and complex nature of the search space makes HPO both challenging and resource-intensive. Automatic HPO methods have demonstrated their ability to efficiently explore high-dimensional hyperparameter spaces and identify optimal solutions, but training and evaluating the model for each set of hyperparameters is still computationally expensive. To further reduce the computational cost, we propose a novel strategy that wraps the HPO process and terminates it based on the sequence of hyperparameters evaluated. The algorithm is inspired by the classic secretary problem, with two additional variations to better adjust to the HPO process. We evaluated the algorithm using popular HPO samplers, including Random Search (RS), Tree-structured Parzen Estimator (TPE), Bayesian Optimization with Gaussian Processes (BOGP), Genetic Algorithms (GA), and Particle Swarm Optimization (PSO). Results indicate that the proposed algorithm accelerates the HPO process by an average of 34%, with only a minimal trade-off in solution quality of 8%. The algorithm is straightforward to implement, compatible with any HPO setup, and particularly effective in the early stages of optimization. This makes it a valuable tool for practitioners aiming to quickly identify promising hyperparameters or reducing the search space, significantly cutting down the time and computational resources required.}, archive = {J_NEUCOM}, author = {Víctor Muñoz and Carmen Ballester and Dorin Copaci and Luis Moreno and Dolores Blanco}, doi = {10.1016/j.neucom.2025.129455}, journal = {Neurocomputing}, month = {4}, pages = {129455}, shortjournal = {Neurocomputing}, title = {Accelerating hyperparameter optimization with a secretary}, volume = {625}, year = {2025}, }
• (2025). DAD++: Improved data-free test time adversarial defense. NEUCOM, 625, 129453. (www)

  @article{ , abstract = {With the increasing deployment of deep neural networks in safety–critical applications such as self-driving cars, medical imaging, anomaly detection, etc., adversarial robustness has become a crucial concern in the reliability of these networks in real-world scenarios. A plethora of works based on adversarial training and regularization-based techniques have been proposed to make these deep networks robust against adversarial attacks. However, these methods require either retraining models or training them from scratch, making them infeasible to defend pre-trained models when access to training data is restricted. To address this problem, we propose a test time Data-free Adversarial Defense (DAD) containing detection and correction frameworks. Moreover, to further improve the efficacy of the correction framework in cases when the detector is under-confident, we propose a soft-detection scheme. We conduct a wide range of experiments and ablations on several datasets and network architectures to show the efficacy of our proposed approach. Furthermore, we demonstrate the applicability of our approach in imparting adversarial defense at test time under data-free (or data-efficient) applications/setups, such as Data-free Knowledge Distillation and Source-free Unsupervised Domain Adaptation, as well as Semi-supervised classification frameworks. We observe that in all the experiments and applications, our DAD++ gives an impressive performance against various adversarial attacks with a minimal drop in clean accuracy. The source code is available here .}, archive = {J_NEUCOM}, author = {Gaurav Kumar Nayak and Inder Khatri and Shubham Randive and Ruchit Rawal and Anirban Chakraborty}, doi = {10.1016/j.neucom.2025.129453}, journal = {Neurocomputing}, month = {4}, pages = {129453}, shortjournal = {Neurocomputing}, title = {DAD++: Improved data-free test time adversarial defense}, volume = {625}, year = {2025}, }
• (2025). Formal analysis on the DNN-kWTA model with non-ideal transfer function and noisy integrator. NEUCOM, 625, 129452. (www)

  @article{ , abstract = {The dual neural network-based (DNN)- k WTA model, based on a DNN architecture, serves as a continuous-time neural approach for identifying the k largest values from a set of n input values. This model possesses two notable characteristics. Firstly, it exhibits significantly fewer interconnections compared to other neural models. Secondly, it demonstrates the capability to accurately detect the k largest values within a finite time. However, the latter quality hinges on the assumption of perfect realization, which is often unattainable in practice. Imperfections such as the inability to achieve an ideal step function for neuron transfer functions and the presence of time-varying multiplicative noise, stemming from random fluctuations at the recurrent neuron’s output, can impact its accuracy, i.e., the ability of correctly finding out the k largest values. This paper analyzes the behaviors of the DNN- k WTA model under the aforementioned imperfections, including multiplicative noise and a non-ideal transfer function. We first derive the neural dynamics for the DNN- k WTA model under the mentioned imperfections. When the noise level is excessively high or the transfer function gain is too low, the neural dynamics fail to converge. Consequently, we theoretically demonstrate that under specific conditions, the neural dynamics converge to an equilibrium point. Leveraging the characteristics of the dynamics, we develop a practical method to assess the model’s ability to accurately identify the k largest numbers. This method enables us to effectively estimate the likelihood of the model correctly identifying these numbers. Additionally, when the input numbers adhere to a uniform distribution, we establish a lower bound expression for the probability of the model correctly identifying the k largest numbers. It is worth noting that, according to basic probability theory, any probability density function can be transformed into a uniform distribution without affecting the ordering of random variables. Simulations are conducted to validate our theoretical analysis.}, archive = {J_NEUCOM}, author = {Wenhao Lu and Yuanjin Zheng and Chi-Sing Leung}, doi = {10.1016/j.neucom.2025.129452}, journal = {Neurocomputing}, month = {4}, pages = {129452}, shortjournal = {Neurocomputing}, title = {Formal analysis on the DNN-kWTA model with non-ideal transfer function and noisy integrator}, volume = {625}, year = {2025}, }
• (2025). A hybrid mamba and sparse look-up table network for perceptual-friendly underwater image enhancement. NEUCOM, 625, 129451. (www)

  @article{ , abstract = {Red light attenuation, medium absorption differences, and suspended particles are the main factors causing color distortion and detail blurring in underwater images. However, most existing deep learning methods face challenges in the enhancement process, such as high computational costs, insufficient global modeling, imprecise or excessive local adjustments, and a lack of perceptual-friendly color distribution. This article proposes a hybrid network for underwater image enhancement that combines Mamba and 3D sparse Look-up Table (LUT) to address the aforementioned issues. Mamba achieves interaction between global context and local information through four-directional and bidirectional cross-scanning, global coarse and local fine-grained features are effectively aggregated within the corresponding stages of the Encoder–Decoder, enabling image global modeling and restoration of local details at a lower computational cost. 3D sparse LUT compensates for color degradation through the mapping of color change matrices, making the overall quality more consistent with visual perception. Noteworthy improvements are showcased across three full and non-reference underwater benchmarks, our method achieves gains of up to 28.6 dB and 0.98 on PSNR and SSIM compared to twelve state-of-the-art models, effectively correcting color distortion while improving texture details. The code is available at https://github.com/SUIEDDM/UIEMa .}, archive = {J_NEUCOM}, author = {Weiming Li and Xuelong Wu and Shuaishuai Fan and Glyn Gowing}, doi = {10.1016/j.neucom.2025.129451}, journal = {Neurocomputing}, month = {4}, pages = {129451}, shortjournal = {Neurocomputing}, title = {A hybrid mamba and sparse look-up table network for perceptual-friendly underwater image enhancement}, volume = {625}, year = {2025}, }
• (2025). Single-frame multi-exposure image fusion via narrowband filter decoupled imaging. NEUCOM, 625, 129441. (www)

  @article{ , abstract = {Multi-exposure image fusion (MEF) can efficiently enhance the dynamic range of image. It can break through the physical imaging limitations inherent in photoelectric sensors. However, the single-camera multi-exposure method requires a time investment, while the multi-camera single-exposure approach is subject to imaging conditions. In this paper, we propose a single-frame decoupled imaging method that enables acquiring multiple differently exposed images from a single exposure captured by one color camera. The method leverages the physical imaging process of a color camera, decoupling the narrowband filtered RAW data into multiple exposure images by exploiting the variations in quantum efficiency distributions. And based on this approach, we construct a decomposed single-frame (DSF) images dataset. The sequence of images within this dataset are naturally spatio-temporally consistent and no longer require registration. Furthermore, a decomposed single-frame MEF network is proposed, termed as DSF-MEF, which employs a hierarchical encoder-decoder structure to predict exposure weight mappings. Specifically, we design a residual mixed attention module (RMAM) for exposure weight prediction. It uses channel and spatial domain attention mechanisms and residual jump connections to perform feature extraction. Subsequently, to improve the overall spatial continuity representation of the exposure weight map sequence, we construct multiscale feature integration module (MFIM) to capture exposure information at different resolution scales. A loss function composed of image structural similarity, gradient texture similarity, and pixel intensity is designed to comprehensively optimize fusion performance. Experimental results show that our method not only achieves single-frame HDR fusion imaging, but also achieves better fusion visual effects compared to other advanced methods.}, archive = {J_NEUCOM}, author = {Zhuang Zhao and Xin Ke and Jing Han and Zijian Wu and Jun Lu and Lianfa Bai and Shuaifeng Gong and Yan Zhang and Yong Peng and Fengchao Xiong and Duan Wei}, doi = {10.1016/j.neucom.2025.129441}, journal = {Neurocomputing}, month = {4}, pages = {129441}, shortjournal = {Neurocomputing}, title = {Single-frame multi-exposure image fusion via narrowband filter decoupled imaging}, volume = {625}, year = {2025}, }
• (2025). Enhancing transformer-based models for long sequence time series forecasting via structured matrix. NEUCOM, 625, 129429. (www)

  @article{ , abstract = {Recently, Transformer-based models for long sequence time series forecasting have demonstrated promising results. The self-attention mechanism as the core component of these Transformer-based models exhibits great potential in capturing various dependencies among data points. Despite these advancements, it has been a subject of concern to improve the efficiency of the self-attention mechanism. Unfortunately, current specific optimization methods are facing the challenges in applicability and scalability for the future design of long sequence time series forecasting models. Hence, in this article, we propose a novel architectural framework that enhances Transformer-based models through the integration of Surrogate Attention Blocks (SAB) and Surrogate Feed-Forward Neural Network Blocks (SFB). The framework reduces both time and space complexity by the replacement of the self-attention and feed-forward layers with SAB and SFB while maintaining their expressive power and architectural advantages. The equivalence of this substitution is fully demonstrated. The extensive experiments on 10 Transformer-based models across five distinct time series tasks demonstrate an average performance improvement of 12.4%, alongside 61.3% reduction in parameter counts. 1}, archive = {J_NEUCOM}, author = {Zhicheng Zhang and Yong Wang and Shaoqi Tan and Bowei Xia and Yujie Luo}, doi = {10.1016/j.neucom.2025.129429}, journal = {Neurocomputing}, month = {4}, pages = {129429}, shortjournal = {Neurocomputing}, title = {Enhancing transformer-based models for long sequence time series forecasting via structured matrix}, volume = {625}, year = {2025}, }
• (2025). Sparse communication in multi-agent deep reinforcement learning. NEUCOM, 625, 129344. (www)

  @article{ , abstract = {Learning to communicate efficiently is central to multi-agent deep reinforcement learning (MADRL). Existing methods often require agents to exchange messages intensively, which abuses communication channels and leads to high communication overhead. Only a few methods target on learning sparse communication, but they allow limited information to be shared, which affects the efficiency of policy learning. In this work, we propose a multi-agent deep reinforcement learning framework with a decentralized communication scheduling process. The proposed framework, which we call Model-Based Communication (MBC), employs supervised learning to build a message estimation model. This model is used by individual agents to decide if they have to communicate their local information to other agents: agents do not communicate their local information if the intended messages can be properly estimated by others. The MBC framework enables multiple agents to make decisions with sparse communication. We evaluate our framework in a variety of mixed cooperative-competitive environments in both homogeneous and heterogeneous domains. The experimental results show that the MBC improves the performance the state-of-art baselines in both domains and leads to a lower communication overhead compared to the baselines.}, archive = {J_NEUCOM}, author = {Shuai Han and Mehdi Dastani and Shihan Wang}, doi = {10.1016/j.neucom.2025.129344}, journal = {Neurocomputing}, month = {4}, pages = {129344}, shortjournal = {Neurocomputing}, title = {Sparse communication in multi-agent deep reinforcement learning}, volume = {625}, year = {2025}, }
• (2025). Minimizing learning experiences in embodied agents language and action learning. NEUCOM, 624, 129510. (www)

  @article{ , abstract = {Language learning necessarily requires the ability to generalize knowledge extracted from a limited number of training examples to the unbounded set of meanings that can be expressed through language. For learning robots, the necessity to learn from limited experiences is further compounded by the high cost of collecting embodied training data. In this article we train a transformer neural network with behavioral cloning to control an embodied agent based on multimodal (language and vision) inputs. We analyze the role of several factors in determining the amount of learning experiences necessary to acquire integrated language and action skills. Our results indicate that the embodied and situated nature of the agents, the ability to reuse previous acquired knowledge to learn new related skills, and the capacity to extract knowledge from the distributional properties of data can greatly reduce the number of necessary examples. We arrive at these conclusions by comparing the results of a series of experiments in which the role of each considered factor is systematically varied.}, archive = {J_NEUCOM}, author = {Nicola Milano and Stefano Nolfi}, doi = {10.1016/j.neucom.2025.129510}, journal = {Neurocomputing}, month = {4}, pages = {129510}, shortjournal = {Neurocomputing}, title = {Minimizing learning experiences in embodied agents language and action learning}, volume = {624}, year = {2025}, }
• (2025). Counterfactual contrastive learning for weakly supervised temporal sentence grounding. NEUCOM, 624, 129508. (www)

  @article{ , abstract = {Temporal sentence grounding in videos (TSGV) aims to retrieve a temporal moment that corresponds to a language query from an untrimmed video. Fine-grained temporal location acquisition is labor-intensive, leading to increased attention on weakly supervised methods that establish video-language correspondence only at the video level. Existing methods mainly employ a contrastive learning approach: mining positive and negative temporal proposals and reconstructing masked queries for cross-modal similarity comparisons. However, the lack of fine-grained annotations in weakly supervised settings complicates the capture of foreground information in untrimmed videos, hindering accurate positive proposal mining and reducing contrastive learning effectiveness. To address this, we construct multiple counterfactual sequences from different perspectives to provide more supervisory information, enabling the model to learn robust visual/semantic representations that are less sensitive to background and more attuned to key foreground information, thereby enhancing accuracy. Specifically, we first create biased unimodal and unbiased multimodal counterfactual sequences at the modality level, proposing a training strategy based on a biased unimodal model to assist the unbiased multimodal model and reduce unimodal bias. Next, we develop intra-sequence and inter-sequence counterfactual sequences across different videos for enhanced video-level supervision, improving the model's ability to discriminate similar video-sentence samples. Finally, we construct complete and incomplete key concepts counterfactual sequences for fine-grained supervision, enhancing accuracy in locating foreground boundaries by identifying and replacing non-optional and indispensable concepts in the original video/query. Extensive experimental evaluations demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Yenan Xu and WanRu Xu and ZhenJiang Miao}, doi = {10.1016/j.neucom.2025.129508}, journal = {Neurocomputing}, month = {4}, pages = {129508}, shortjournal = {Neurocomputing}, title = {Counterfactual contrastive learning for weakly supervised temporal sentence grounding}, volume = {624}, year = {2025}, }
• (2025). BnRep: A repository of bayesian networks from the academic literature. NEUCOM, 624, 129502. (www)

  @article{ , abstract = {Bayesian networks (BNs) are widely used for modeling complex systems with uncertainty, yet repositories of pre-built BNs remain limited. This paper introduces bnRep , an open-source R package offering a comprehensive collection of documented BNs, facilitating benchmarking, replicability, and education. With over 200 networks from academic publications, bnRep integrates seamlessly with bnlearn and other R packages, providing users with interactive tools for network exploration. The package includes a Shiny app for intuitive and interactive filtering and visualization, further enhancing accessibility for users of all expertise levels.}, archive = {J_NEUCOM}, author = {Manuele Leonelli}, doi = {10.1016/j.neucom.2025.129502}, journal = {Neurocomputing}, month = {4}, pages = {129502}, shortjournal = {Neurocomputing}, title = {BnRep: A repository of bayesian networks from the academic literature}, volume = {624}, year = {2025}, }
• (2025). MHAVSR: A multi-layer hybrid alignment network for video super-resolution. NEUCOM, 624, 129501. (www)

  @article{ , abstract = {Video super-resolution (VSR) aims to restore high-resolution (HR) frames from low-resolution (LR) frames, the key to this task is to fully utilize the complementary information between frames to reconstruct high-resolution sequences. Current works tackle with this by exploiting a sliding window strategy or a recurrent architecture for single alignment, which either lacks long range modeling ability or is prone to frame-by-frame error accumulation. In this paper, we propose a Multi-layer Hybrid Alignment network for VSR (MHAVSR), which combines a sliding window with a recurrent structure and extends the number of propagation layers based on this hybrid structure. Repeatedly, at each propagation layer, alignment operations are performed simultaneously on bidirectional neighboring frames and hidden states from recursive propagation, which improves the alignment while fully utilizing both the short-term and long-term information in the video sequence. Next, we present a flow-enhanced dual-deformable alignment module, which improves the accuracy of deformable convolutional offsets by optical flow and fuses the separate alignment results of the hybrid alignment to reduce the artifacts caused by alignment errors. In addition, we introduce a spatial–temporal reconstruction module to compensate the representation capacity of model at different scales. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches. In particular, on the Vid4 test set, our model exceeds the IconVSR by 0.82 dB in terms of PSNR with a similar number of parameters. Codes are available at https://github.com/fzuqxt/MHAVSR .}, archive = {J_NEUCOM}, author = {Xintao Qiu and Yuanbo Zhou and Xinlin Zhang and Yuyang Xue and Xiaoyong Lin and Xinwei Dai and Hui Tang and Guoyang Liu and Rui Yang and Zhen Li and Xiaojing Wei and Junxiu Yang and Tong Tong and Qinquan Gao}, doi = {10.1016/j.neucom.2025.129501}, journal = {Neurocomputing}, month = {4}, pages = {129501}, shortjournal = {Neurocomputing}, title = {MHAVSR: A multi-layer hybrid alignment network for video super-resolution}, volume = {624}, year = {2025}, }
• (2025). Constrained deep reinforcement transfer learning for short-term forecasting of wind discrepancies at ocean stations. NEUCOM, 624, 129491. (www)

  @article{ , abstract = {Accurate ocean wind speed forecasting holds significant value in energy applications. Transfer learning takes advantage of the similarity in wind speeds between neighboring sea areas, reducing time and computational costs. However, coastal areas often experience significant environmental changes, causing uneven wind fields and noticeable wind speed differences between nearby locations. Traditional intelligent prediction methods suffer from low data utilization and struggle to distinguish the characteristics of source and target marine regions, reducing the effectiveness of transfer learning. To address this, we propose a deep reinforcement transfer learning model to improve multi-step and short-term ocean wind speed prediction, converting supervised regression tasks into decision-making processes. Specifically, by constraining both initial and fine-tuning training outputs to ensure physical consistency, the Twin Delayed Deep Deterministic Policy Gradient (TD3) model is guided to focus on distinguishing differences, while the attention mechanism amplifies local features across adjacent sea areas. The adjusted TD3 model is pre-trained at data-rich locations, transferred and fine-tuned at target locations with limited data for multi-step short-term wind speed predictions. To validate the model, nine sites along the coast are selected for training, and predictions are validated at five sites. Ten predictive models are established to compare the performance of the proposed model. Experimental results show that under limited conditions, the proposed model's accuracy improved by 46 % compared to the baseline model, demonstrating that the model, when combined with transfer learning, not only significantly saves data but also ensures accuracy, stability and generalization in predictions.}, archive = {J_NEUCOM}, author = {Jun Zhang and Yaoran Chen and Hang Pan and Liyuan Cao and Chunxiang Li}, doi = {10.1016/j.neucom.2025.129491}, journal = {Neurocomputing}, month = {4}, pages = {129491}, shortjournal = {Neurocomputing}, title = {Constrained deep reinforcement transfer learning for short-term forecasting of wind discrepancies at ocean stations}, volume = {624}, year = {2025}, }
• (2025). SRMamba-T: Exploring the hybrid mamba-transformer network for single image super-resolution. NEUCOM, 624, 129488. (www)

  @article{ , abstract = {Single Image Super-Resolution (SISR) has made significant advancements with both CNN-based and Transformer-based models. However, CNNs often struggle to capture long-range dependencies effectively, and Transformers, though powerful, are hindered by quadratic computational complexity. In recent years, state space models (SSMs), such as Mamba, have emerged as promising alternatives due to their ability to model long-range dependencies with linear time complexity. In light of this, we propose SRMamba-T, a hybrid model that strategically combines Mamba and Transformer architectures to balance computational efficiency with high performance. Specifically, we employ Transformer layers following Mamba layers to further enhance the model’s capability to handle long-range spatial information and expand its effective receptive fields. To design a lightweight network, we propose a multi-directional selective scanning module to reduce parameter count and improve computational efficiency. Additionally, a feature fusion module serves as a bottleneck to effectively integrate hierarchical features and enhance the model’s representational ability. Comprehensive experimental evaluations across five widely recognized benchmarks underscore our model’s effectiveness, demonstrating its superiority over existing state-of-the-art (SOTA) methods. For example, our model achieved a significant PSNR improvement of 0.28 dB for × 2 lightweight super-resolution on Urban100, with a reduction in MACs by 38.7% (from 198.1G to 121.5G) compared to the SOTA model, MambaIR.}, archive = {J_NEUCOM}, author = {Cencen Liu and Dongyang Zhang and Guoming Lu and Wen Yin and Jielei Wang and Guangchun Luo}, doi = {10.1016/j.neucom.2025.129488}, journal = {Neurocomputing}, month = {4}, pages = {129488}, shortjournal = {Neurocomputing}, title = {SRMamba-T: Exploring the hybrid mamba-transformer network for single image super-resolution}, volume = {624}, year = {2025}, }
• (2025). Dual-attention enhanced variational encoding for interpretable remaining useful life prediction. NEUCOM, 624, 129487. (www)

  @article{ , abstract = {In Prognostics Health Management (PHM), predicting Remaining Useful Life (RUL) is a key technique for equipment health evaluation. The utilization of deep learning methods has improved prediction accuracy. However, these approaches often fail to provide the transparency and interpretability that maintenance personnel require to diagnose equipment degradation effectively. To address this challenge, a Dual-Attention Enhanced Variational Encoding (DAEVE) approach based on Transformer is developed for more interpretable RUL prediction. This framework integrates both sensor and time step encoders, a latent space with inductive bias and a regression model: the fusion encoder compresses input data into a three-dimension(3-D) latent space, facilitating both the prediction and interpretation of the equipment degradation process. Four turbofan aircraft engine datasets are applied in extensive experiments to evaluate the efficacy of proposed method. The results demonstrate that DAEVE outperforms most state-of-the-art methods in prediction accuracy. Furthermore, the proposed method exhibits the latent degradation trajectories and more informative sensors in diverse stages. This research could enhance maintenance decision-making processes and reduce operational risks, contributing to the advancement of predictive maintenance in the aerospace and related industries.}, archive = {J_NEUCOM}, author = {Wen Liu and Jyun-You Chiang and Guojun Liu and Haobo Zhang}, doi = {10.1016/j.neucom.2025.129487}, journal = {Neurocomputing}, month = {4}, pages = {129487}, shortjournal = {Neurocomputing}, title = {Dual-attention enhanced variational encoding for interpretable remaining useful life prediction}, volume = {624}, year = {2025}, }
• (2025). YOLO-FAS: A lightweight model for detecting rebar intersections location and tying status. NEUCOM, 624, 129485. (www)

  @article{ , abstract = {Current deep learning (DL) algorithms for detecting complex rebar mesh intersection points rely heavily on large amounts of training data to ensure recognition accuracy and generalization capabilities. However, deploying such algorithms on small mobile robots is challenging due to limited computational resources, making it difficult to achieve both real-time performance and high precision. To address this issue, this paper proposes a lightweight YOLO-FAS model specifically designed for rebar intersections detection on tying robots. This model is based on YOLOv5s and optimizes the FastNet structure by introducing the PGConv module, enhances feature fusion with the AFPN (Asymptotic feature pyramid network) module, and improves localization accuracy using the EIOU loss function. Experimental results demonstrate that the model's parameters and computational load are reduced by 60.6 % and 62 %, respectively. Furthermore, through BN (Batch Normalization) channel pruning and QAT (Quantization Aware Training), the model is compressed, and inference is accelerated using TensorRT. The improved YOLO-FAS model achieves a FPS (frames per second) increase of 84.1 % in FP32 mode and 33.2 % in INT8 mode. Finally, after real-world deployment testing, the system's average memory usage is reduced by 0.77 GB, and the accuracy of recognition of intersection points reaches 98.21 %, representing an improvement of 3.04 % over YOLOv5s. The results indicate that this method effectively achieves model lightweighting while ensuring efficient and accurate detection of rebar intersection points, demonstrating robust performance and promising application prospects.}, archive = {J_NEUCOM}, author = {Hao Duan and Mingming Yu and Tengfeng Ai and Mengmeng Zhu and Haili Jiang and Shuai Guo}, doi = {10.1016/j.neucom.2025.129485}, journal = {Neurocomputing}, month = {4}, pages = {129485}, shortjournal = {Neurocomputing}, title = {YOLO-FAS: A lightweight model for detecting rebar intersections location and tying status}, volume = {624}, year = {2025}, }
• (2025). Statistic deviation mode balancer (SDMB): A novel sampling algorithm for imbalanced data. NEUCOM, 624, 129484. (www)

  @article{ , abstract = {In supervised learning, the efficacy of classifier algorithms is heavily dependent on the quality of data. Imbalanced datasets, where the class distribution is not uniform, pose a significant challenge, often leading to suboptimal classifier performance. Traditional approaches to rectifying this imbalance have relied on duplicating minority class instances or generating synthetic data, which can introduce bias or outliers. Our novel Statistic Deviation Mode Balancer (SDMB) algorithm addresses these issues by generating new instances that closely mirror the original data structure. Utilizing standard deviation and mode analysis, SDMB strategically synthesizes minority class data while avoiding the pitfalls of outlier generation. The result is a balanced dataset that facilitates more accurate learning by classifier algorithms. We have rigorously tested SDMB across various datasets and compared its performance against existing balancing methods. Our findings indicate that SDMB not only outperforms its counterparts but also significantly enhances the practical application of classifier algorithms in real-world datasets.}, archive = {J_NEUCOM}, author = {Mahmoud Alimoradi and Reza Sadeghi and Arman Daliri and Mahdieh Zabihimayvan}, doi = {10.1016/j.neucom.2025.129484}, journal = {Neurocomputing}, month = {4}, pages = {129484}, shortjournal = {Neurocomputing}, title = {Statistic deviation mode balancer (SDMB): A novel sampling algorithm for imbalanced data}, volume = {624}, year = {2025}, }
• (2025). Bipartite consensus for distributed networks with random time-delay in a additive noisy environment via ergodic backward products. NEUCOM, 624, 129483. (www)

  @article{ , abstract = {This article addresses the mean square bipartite consensus (MSBC) issue of stochastic distributed network with random time-delay in the presence of antagonistic information and additive noises by the ergodic technique of backward products (BP). First, a new distributed stochastic approximation controller is proposed for achieving the MSBC with random time-delay based on the information of antagonistic neighbor communication. Compared with traditional procedure when dealing with time-delay, where Lyapunov functional is utilized for system with time-delay, this paper adopted a new method of dimension extension method to handling random time-delay. By employing an ergodic BP and iterative algorithm, some sufficient criterions for realizing the MSBC of stochastic distributed network under a additive noisy and random time-delay setting, are obtained. Ultimately, our theories can be verified by some numerical simulations.}, archive = {J_NEUCOM}, author = {Jingxin Shang and Yingxue Du and Zhi Liu and Ancai Zhang and Yan Zhang}, doi = {10.1016/j.neucom.2025.129483}, journal = {Neurocomputing}, month = {4}, pages = {129483}, shortjournal = {Neurocomputing}, title = {Bipartite consensus for distributed networks with random time-delay in a additive noisy environment via ergodic backward products}, volume = {624}, year = {2025}, }
• (2025). Inferring single-cell trajectories via critical cell identification using graph centrality algorithm. NEUCOM, 624, 129482. (www)

  @article{ , abstract = {Trajectory inference (TI) aims to infer cell differentiation trajectories in biological processes. Numerous computational methods have been developed to infer cell lineages from single-cell gene expression data. However, cluster-based methods involve a discretization that fails to capture the continuous nature of differentiation processes, while graph-based methods directly estimate the differentiation process from gene expression profiles without detecting subpopulations, making them susceptible to noise. To address these limitations, we propose scTICG, a single-cell trajectory inference method through critical cell identification and a greedy strategy. scTICG integrates the strengths of cluster-based and graph-based methods. Initially, a cluster-based backbone structure is constructed to serve as a coarse-grained trajectory. Then, considering the dynamics of cell state transitions and the influence of certain critical cells, we identify these critical cells using the graph centrality algorithm. Subsequently, these critical cells are leveraged to refine the trajectory using a greedy strategy. We evaluate scTICG on five public datasets and compare its performance with eight state-of-the-art trajectory inference methods. The experimental results demonstrate that scTICG can infer more accurate and robust trajectories compared to competitive methods. The R code for scTICG is freely available at https://github.com/DHUDBlab/scTICG .}, archive = {J_NEUCOM}, author = {Yanglan Gan and Jiaqi Chu and Guangwei Xu and Cairong Yan and Guobing Zou}, doi = {10.1016/j.neucom.2025.129482}, journal = {Neurocomputing}, month = {4}, pages = {129482}, shortjournal = {Neurocomputing}, title = {Inferring single-cell trajectories via critical cell identification using graph centrality algorithm}, volume = {624}, year = {2025}, }
• (2025). Adversarially adaptive temperatures for decoupled knowledge distillation with applications to speaker verification. NEUCOM, 624, 129481. (www)

  @article{ , abstract = {Knowledge Distillation (KD) aims to transfer knowledge from a high-capacity teacher model to a lightweight student model, thereby enabling the student model to attain a level of performance that would be unattainable through conventional training methods. In conventional KD, the loss function’s temperature that controls the smoothness of class distributions is fixed. We argue that distribution smoothness is critical to the transfer of knowledge and propose an adversarial adaptive temperature module to set the temperature dynamically during training to enhance the student’s performance. Using the concept of decoupled knowledge distillation (DKD), we separate the Kullback–Leibler (KL) divergence into a target-class term and a non-target-class term. However, unlike DKD, we adversarially update the temperature coefficients of the target and non-target classes to maximize the distillation loss. We named our method A dversarially A daptive T emperature for DKD (AAT-DKD). Our approach demonstrates improvements over KD methods across three test sets of Voxceleb1 for two student models (x-vector and ECAPA-TDNN). Specifically, compared to the traditional KD and DKD, our method achieves a remarkable reduction of 17.78% and 11.90% in EER using ECAPA-TDNN speaker embedding. Moreover, our method performs well on CN-Celeb and VoxSRC21, further highlighting its robustness and effectiveness across different datasets.}, archive = {J_NEUCOM}, author = {Zezhong Jin and Youzhi Tu and Chong-Xin Gan and Man-Wai Mak and Kong-Aik Lee}, doi = {10.1016/j.neucom.2025.129481}, journal = {Neurocomputing}, month = {4}, pages = {129481}, shortjournal = {Neurocomputing}, title = {Adversarially adaptive temperatures for decoupled knowledge distillation with applications to speaker verification}, volume = {624}, year = {2025}, }
• (2025). A dual reverse distillation scheme for image anomaly detection. NEUCOM, 624, 129479. (www)

  @article{ , abstract = {Image anomaly detection and localization are challenging in industrial applications due to highly unbalanced distributed real-world data. Reverse Distillation (RD), utilizing a teacher–student model, has been employed for this task, yet it often suffers from over generalization issue, as the student network struggles to effectively recover abnormal features. To address this issue, this paper proposes a dual reverse distillation (DRD) scheme that combines pseudo abnormal and normal images and restores their features simultaneously with the reverse distillation method. Specifically, DRD incorporates a novel multi-feature cascade fusion (MFCF) block, a combined knowledge distillation loss function, and a detection refinement method. The MFCF block enhances feature integration across layers and captures more comprehensive information, while the knowledge distillation-loss function balances features from batch and individual images to better align output features with normal features. Furthermore, a refinement method at the inference stage further improves detection accuracy. Such strategies help improve the recovery of abnormal features while reducing over generalization. Experiments with many public datasets, including MVTec, BTAD and VisA, were conducted to verify effectiveness of the proposed scheme. Results show that, for MVTec, it achieved 99.46% AUROC at the image level and 98.39% AUROC at the pixel level, outperforming RD by 1.02% and 0.59%, respectively. Ablation experiments also confirmed that DRD can mitigate the over generalization issue effectively.}, archive = {J_NEUCOM}, author = {Chenkun Ge and Xiaojun Yu and Hao Zheng and Zeming Fan and Jinna Chen and Perry Ping Shum}, doi = {10.1016/j.neucom.2025.129479}, journal = {Neurocomputing}, month = {4}, pages = {129479}, shortjournal = {Neurocomputing}, title = {A dual reverse distillation scheme for image anomaly detection}, volume = {624}, year = {2025}, }
• (2025). Adaptive lightweight network construction method for self-knowledge distillation. NEUCOM, 624, 129477. (www)

  @article{ , abstract = {Self-Knowledge Distillation (self-KD) has become a promising method for neural network compression due to its advances in computational efficiency. Nevertheless, its applicability is constrained by the inherent inflexibility of the network architecture and the absence of quantitative metrics to evaluate the distillability of the architecture. To address these problems, a two-stage adaptive dynamic distillation network framework (ADDN) is proposed to adapt the architecture based on the distillability of the current architecture, containing a hypernetwork topology construction stage and a subnetwork training stage. To evaluate the distillability of current architectures without necessitating extensive training, we propose a set of low-cost distillability metrics that evaluate architectures from the perspective of architectural similarity and clustering ability. Furthermore, to simplify the hypernetwork structure and reduce the complexity of the construction process, a hierarchical filtering module is introduced to dynamically refine and remove candidate operations within the architecture incrementally, contingent upon the distillability of the current architecture. To validate the effectiveness of our approach, we conduct extensive experiments on various image classification datasets and compare with current works. Experimental results demonstrate that the self-knowledge distillation network architecture obtained by our proposed methodology simultaneously attains superior distillability and efficiency while significantly curtailing construction expenses.}, archive = {J_NEUCOM}, author = {Siyuan Lu and Weiliang Zeng and Xueshi Li and Jiajun Ou}, doi = {10.1016/j.neucom.2025.129477}, journal = {Neurocomputing}, month = {4}, pages = {129477}, shortjournal = {Neurocomputing}, title = {Adaptive lightweight network construction method for self-knowledge distillation}, volume = {624}, year = {2025}, }
• (2025). A defense mechanism against label inference attacks in vertical federated learning. NEUCOM, 624, 129476. (www)

  @article{ , abstract = {Vertical Federated Learning (VFL, for short) is a category of Federated Learning that is gaining increasing attention in the context of Artificial Intelligence. According to this paradigm, machine/deep learning models are trained collaboratively among parties with vertically partitioned data. Typically, in a VFL scenario, the labels of the samples are kept private from all parties except the aggregating server, that is, the label owner. However, recent work discovered that by exploiting the gradient information returned by the server to bottom models, with the knowledge of only a small set of auxiliary labels on a very limited subset of training data points, an adversary could infer the private labels. These attacks are known as label inference attacks in VFL. In our work, we propose a novel framework called KD k (knowledge distillation with k -anonymity) that combines knowledge distillation and k -anonymity to provide a defense mechanism against potential label inference attacks in a VFL scenario. Through an exhaustive experimental campaign, we demonstrate that by applying our approach, the performance of the analyzed label inference attacks decreases consistently, even by more than 60%, maintaining the accuracy of the whole VFL almost unaltered.}, archive = {J_NEUCOM}, author = {Marco Arazzi and Serena Nicolazzo and Antonino Nocera}, doi = {10.1016/j.neucom.2025.129476}, journal = {Neurocomputing}, month = {4}, pages = {129476}, shortjournal = {Neurocomputing}, title = {A defense mechanism against label inference attacks in vertical federated learning}, volume = {624}, year = {2025}, }
• (2025). Subspace screening rule for low-rank tensor completion with tensor nuclear norm. NEUCOM, 624, 129474. (www)

  @article{ , abstract = {The tensor completion problem involves recovering a tensor from partially observed entries. Tensor completion methods based on the tensor nuclear norm have garnered significant attention due to their remarkable completion performance. However, the inability to rapidly solve it has become a limiting factor. To address the limitation, we propose a subspace screening rule that accelerates the tensor nuclear norm approach. Subspace screening rule reduces the scale of matrix variables by exploiting the representation of low-rank matrices through a limited number of subspaces. At each iteration, we utilize optimality conditions to selectively capture and compute only the subspaces with non-zero coefficients, thereby accelerating the computational process. Extensive experiments on image and video datasets validate our method. The results demonstrate its ability to efficiently solve the tensor completion problem while maintaining high completion quality, thereby highlighting its practical effectiveness.}, archive = {J_NEUCOM}, author = {Haotian Chen and Yitian Xu}, doi = {10.1016/j.neucom.2025.129474}, journal = {Neurocomputing}, month = {4}, pages = {129474}, shortjournal = {Neurocomputing}, title = {Subspace screening rule for low-rank tensor completion with tensor nuclear norm}, volume = {624}, year = {2025}, }
• (2025). Sparse temporal knowledge graph completion based on path imitation. NEUCOM, 624, 129473. (www)

  @article{ , abstract = {Sparse temporal knowledge graph completion is a particular task in temporal knowledge graph completion, which aims to use the sparse knowledge in the sparse temporal knowledge graph to complete the missing facts. How to reduce the influence of sparsity in sparse temporal knowledge graphs is the critical issue to improve the completion effect. Existing completion models for sparse temporal knowledge graphs reduce the sparsity of temporal knowledge graphs by introducing additional action space and adaptive rewards, but these models still have limitations. On the one hand, the inference process of these models does not take the influence of noise information into account. On the other hand, the one-sidedness of the reasoning path of model training is also an important factor restricting the completion effect. In order to solve the above problems, we propose a Sparse Temporal knowledge graph completion model based on Path Imitation (STPI). In order to solve the noise problem, we propose a dynamic completion strategy based on path imitation. To solve the problem of one-sided paths, we quantify the similarity of paths and introduce them into soft rewards for reward shaping. Taking into account the limited data present in the sparse temporal knowledge graph, we gather details from neighboring entities to create updated entity embeddings. This method aids in enhancing the precision of our inference outcomes. Experimental results show that STPI outperforms state-of-the-art models on sparse public datasets while ensuring interpretability.}, archive = {J_NEUCOM}, author = {Jiahui Hu and Luyi Bai and Xuemeng Shang and Guangchen Feng and Yang Gao}, doi = {10.1016/j.neucom.2025.129473}, journal = {Neurocomputing}, month = {4}, pages = {129473}, shortjournal = {Neurocomputing}, title = {Sparse temporal knowledge graph completion based on path imitation}, volume = {624}, year = {2025}, }
• (2025). UFG-net: Uncertainty and frequency guided network for image forgery localization. NEUCOM, 624, 129471. (www)

  @article{ , abstract = {In Image Forgery Localization (IFL) tasks, post-processing of manipulated images often leaves subtle traces, increasing model uncertainty in predicting tampered areas. Current methods fail to leverage these uncertainties to extract key features effectively. Additionally, important information in the frequency domain is often overlooked. However, existing methods that merely combine frequency and spatial domain features can lead to data redundancy and feature misalignment across domains. To address these challenges, we propose the Uncertainty and Frequency Guided Network (UFG-Net), which integrates uncertainty information during training and utilizes frequency domain insights to enhance feature extraction in the spatial domain. In addition, the Uncertainty and Frequency Guidance (UFG) module is introduced to optimize feature extraction using a multi-scale approach with dual attention mechanisms, while the Forgery Inference (FI) module is designed to enhance final localization precision. Extensive experiments across five diverse datasets demonstrate UFG-Net's exceptional capabilities in IFL. Notably, on the IMD dataset, UFG-Net achieved improvements of 1.8 % in AUC, 1.7 % in F1-score, and 2 % in IoU, outperforming the second-best models across all metrics. Moreover, cross-dataset experiments and robustness tests further underscore UFG-Net's outstanding generalization and robustness.}, archive = {J_NEUCOM}, author = {Kai Wang and Qixian Hao and Shaozhang Niu and Jiwei Zhang and Wen Zhang}, doi = {10.1016/j.neucom.2025.129471}, journal = {Neurocomputing}, month = {4}, pages = {129471}, shortjournal = {Neurocomputing}, title = {UFG-net: Uncertainty and frequency guided network for image forgery localization}, volume = {624}, year = {2025}, }
• (2025). Aggregated center-based clustering algorithm based on principal component radius. NEUCOM, 624, 129469. (www)

  @article{ , abstract = {Center-based clustering algorithms such as k -means have strict requirements on data distribution. As an important initialization improvement method, k -means++ can effectively overcome the problem of local convergence, but it still need to accurately specify the number of clusters and not applicable to non-convex clusters. In order to overcome the above limitations, this paper proposes an aggregated center-based clustering algorithm based on the principle component radius (PCR). First, k -means++ is used as the initialization method to generate an excess of clusters. Subsequently, a cluster characterization method, PCR, with shrinking effect is proposed to accurately characterize the core region of clusters, guiding cluster merging. On this foundation, a graph based aggregation process can efficiently complete cluster identification and adaptively determine the number of clusters. Experimental results on all synthetic and real-world datasets verify the superiority of the proposed algorithm and its high tolerance to parameters and initialization.}, archive = {J_NEUCOM}, author = {Mingchang Cheng and Liu Liu and Tiefeng Ma and Lin Ma and Qijing Yan}, doi = {10.1016/j.neucom.2025.129469}, journal = {Neurocomputing}, month = {4}, pages = {129469}, shortjournal = {Neurocomputing}, title = {Aggregated center-based clustering algorithm based on principal component radius}, volume = {624}, year = {2025}, }
• (2025). Hierarchical task-aware temporal modeling and matching for few-shot action recognition. NEUCOM, 624, 129467. (www)

  @article{ , abstract = {Few-shot action recognition seeks to classify new action categories using only a few labeled video samples as reference. Due to the lack of sufficient training samples and the complex structure of video data, it is difficult to extract global features that can be directly used for classification. Therefore, most previous works adopt image encoders to extract the features of each frame individually, and then perform temporal fusion and alignment for the query and support features. However, they neglect the importance of sufficient spatiotemporal modeling and relationships with other categories in the few-shot task when extracting features, rendering them less effective at distinguishing between the given action classes, especially those that require the perception of local motion. In this paper, we present Hierarchical Task-aware Temporal Modeling and Matching (HTTMM) to better perceive critical motion patterns and extract task-relevant discriminative features for matching. Specifically, we propose a task guidance generator and a hierarchical task-aware temporal module. The former collects samples across all categories to get task contextual information and generate task guidance. The latter performs hierarchical spatiotemporal modeling in a task-aware manner by incorporating the task guidance and leveraging the multi-stage visual features of the image encoder. This design makes the extracted features rich in motion cues and more discriminative among the new classes in the task. Based on the global and local features obtained, we further propose a simple matching strategy that takes multi-level relationships into consideration to improve the robustness of matching. To demonstrate the effectiveness of our method, we evaluate it on four commonly used datasets, i.e. , Kinetics, UCF101, HMDB51, and Something–Something v2. The experimental results show that our method outperforms the existing state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Yucheng Zhan and Yijun Pan and Siying Wu and Yueyi Zhang and Xiaoyan Sun}, doi = {10.1016/j.neucom.2025.129467}, journal = {Neurocomputing}, month = {4}, pages = {129467}, shortjournal = {Neurocomputing}, title = {Hierarchical task-aware temporal modeling and matching for few-shot action recognition}, volume = {624}, year = {2025}, }
• (2025). Neighbor structure aware based cross-view consensus graph learning for incomplete multi-view clustering. NEUCOM, 624, 129466. (www)

  @article{ , abstract = {Multi-view clustering has gained significant attention due to its ability to fully leverage the relationships between different views. In the face of the incomplete observability of complex systems, incomplete multi-view clustering (IMVC) has become an important research direction in machine learning. Currently, graph-based methods have shown excellent performance in clustering research. However, there are still some shortcomings: first, some methods only focus on non-missing sample information; second, most only explore pairwise relationships between samples while neglecting connections between neighboring clusters; finally, many methods fail to simultaneously utilize global structural information and potential information between views. To address these issues, this paper proposes a neighbor-group structure-aware cross-view incomplete multi-view clustering method (NACG_IMVC), aimed at exploring the relationships between sample neighbor groups. This method constructs a high-quality consensus graph by leveraging intra-view local information and inter-view global structural information. Additionally, a novel indicator matrix is designed to account for the structure of missing data, thereby enhancing the quality of the consensus graph. Experimental results on multiple datasets indicate that the clustering performance of NACG_IMVC surpasses that of several advanced clustering methods.}, archive = {J_NEUCOM}, author = {Haifeng Yang and Xingyi Li and Yijing Chen}, doi = {10.1016/j.neucom.2025.129466}, journal = {Neurocomputing}, month = {4}, pages = {129466}, shortjournal = {Neurocomputing}, title = {Neighbor structure aware based cross-view consensus graph learning for incomplete multi-view clustering}, volume = {624}, year = {2025}, }
• (2025). Distributed nash equilibrium seeking for constrained games over jointly strongly connected switching networks under dynamic event-triggered mechanism. NEUCOM, 624, 129465. (www)

  @article{ , abstract = {This paper studies the distributed Nash equilibrium seeking problems for constrained games with players’ dynamics governed by first-order integrator and high-order integrator, respectively. In the considered problems, the switching communication networks are supposed to be jointly strongly connected, and each player’s action set is supposed to be compact and convex. By utilizing the projected gradient play and the consensus protocol, distributed Nash equilibrium seeking strategies are designed. Moreover, in order to save communication resources, dynamic event-triggered mechanisms are designed and introduced into the proposed strategies. Further, we show that there is no Zeno behavior, and the unique Nash equilibrium is globally exponentially stable under the proposed strategies. Finally, the effectiveness of the theoretical results is verified by two examples.}, archive = {J_NEUCOM}, author = {Shuoye Dai and Yu Zhang}, doi = {10.1016/j.neucom.2025.129465}, journal = {Neurocomputing}, month = {4}, pages = {129465}, shortjournal = {Neurocomputing}, title = {Distributed nash equilibrium seeking for constrained games over jointly strongly connected switching networks under dynamic event-triggered mechanism}, volume = {624}, year = {2025}, }
• (2025). Explainable robo-advisor: An online learning framework for new investors without trading records. NEUCOM, 624, 129463. (www)

  @article{ , abstract = {Intelligent investment advisors, commonly known as robo-advisors, are extensively utilized in the wealth management industry for the precise design and marketing of financial products. The primary function of robo-advisors is to accurately extract an individual investor’s multi-dimensional preferences and provide personalized financial product recommendations. However, traditional robo-advisors often suffer from a lack of explainability, dependence on static historical data, and an inability to offer effective recommendations in cold-start scenarios. In this study, we propose an online learning framework named CatBoostNN, which simulates the tree structures of Categorical Boosting model based on the Neural Networks. The proposed CatBoostNN can achieve online learning and update model parameters in a timely manner based on new samples. It can also automatically process categorical variables through Order TS. The user preferences measured by feature importance further make the recommendation process more transparent and explainable. For new investors, our approach takes into account both direct relationships and multiple layers of indirect relationships within the social network. Different nodes in this social relationship network carry varying influences on the target investor, thereby improving recommendation accuracy. We evaluate our approach using a large-scale, real-world dataset comprising more than 700,000 clients of a virtual bank. The results demonstrate the superior effectiveness of the proposed framework compared to other well-recognized benchmark models and conventional cold-start methods.}, archive = {J_NEUCOM}, author = {Xiao Hu and Siqin Kang and Shuyu Xue and Shaokeng Zhu}, doi = {10.1016/j.neucom.2025.129463}, journal = {Neurocomputing}, month = {4}, pages = {129463}, shortjournal = {Neurocomputing}, title = {Explainable robo-advisor: An online learning framework for new investors without trading records}, volume = {624}, year = {2025}, }
• (2025). Passivity for coupled quaternion-valued neural networks with multiple derivative couplings. NEUCOM, 624, 129462. (www)

  @article{ , abstract = {In this paper, a coupled quaternion-valued neural networks (CQVNN) with multiple derivative couplings is introduced, and the passivity problem of such CQVNN is addressed by utilizing the non-separation method. Some quaternion properties are exploited to analyze the passivity for CQVNN, and a quaternion-valued adaptive controller is also devised to guarantee the passivity for such network. Moreover, in light of the obtained passivity results, several synchronization criteria for CQVNN are derived. Finally, two numerical examples exhibit the effectiveness of the obtained synchronization and passivity criteria.}, archive = {J_NEUCOM}, author = {Jin-Liang Wang and Wanlu Wei and Shun-Yan Ren}, doi = {10.1016/j.neucom.2025.129462}, journal = {Neurocomputing}, month = {4}, pages = {129462}, shortjournal = {Neurocomputing}, title = {Passivity for coupled quaternion-valued neural networks with multiple derivative couplings}, volume = {624}, year = {2025}, }
• (2025). EXtreme gradient boosting mix-fit-standalone: A high-performance gradient boosting tree framework for vertical federated learning. NEUCOM, 624, 129461. (www)

  @article{ , abstract = {Vertical federated learning (VFL) is a crucial technology for addressing the issues of data silos. Among its applications, the eXtreme Gradient Boosting (XGBoost) technique based on VFL has been widely adopted because of its excellent versatility and high modeling accuracy. However, with the development of computer technology, the dimensions of the data available for modeling are increasingly rich, and the volume of samples is increasing. However, the existing XGBoost technology based on VFL demonstrates slow modeling speeds when handling large datasets with high dimensions. This slow pace limits model developers to testing only a few sets of model parameters within a given time frame, thereby significantly limiting the potential to extract values from the data. To address this issue, this study proposes the eXtreme Gradient Boosting Mix-Fit-Standalone (XGBoost-MFS) algorithm, a gradient boosting tree framework based on perturbation factor technology. XGBoost-MFS significantly improves modeling speed and significantly reduces network overhead by combining the advantages of existing technologies. Experimental comparisons across various datasets proved the effectiveness of the XGBoost-MFS algorithm, demonstrating its significant advantages in improving model speed. This offers a new and effective solution in the field of VFL.}, archive = {J_NEUCOM}, author = {Shimao Xie and Yun Che and Niannian Chen and Hang Mao and SiCheng Chen}, doi = {10.1016/j.neucom.2025.129461}, journal = {Neurocomputing}, month = {4}, pages = {129461}, shortjournal = {Neurocomputing}, title = {EXtreme gradient boosting mix-fit-standalone: A high-performance gradient boosting tree framework for vertical federated learning}, volume = {624}, year = {2025}, }
• (2025). Multi actors-critic based particle swarm optimization algorithm. NEUCOM, 624, 129460. (www)

  @article{ , abstract = {This paper introduces the Multi-Actor–Critic-based Particle Swarm Optimization Algorithm (MACPSO), an innovative approach to optimizing particle performance through dynamic parameter control. MACPSO applies a multi-actor–critic framework to enhance exploration and maintain population diversity. It features multiple actor networks for adaptive parameter adjustment, coupled with a single critic network to guide the shared optimization goal. The algorithm incorporates particle grouping for intra-group updates and an inter-group updating scheme to facilitate the exchange of optimal information. Furthermore, MACPSO integrates a mutation mechanism aimed at improving the performance of the least effective particles, thereby ensuring sustained diversity within the population. To validate the superior performance of the proposed algorithm, a comparative study was conducted with nine advanced PSO variants using the CEC2017 benchmark suite. The experimental results demonstrate that MACPSO achieved the highest overall ranking in terms of average error across twenty-nine different benchmark functions. Additionally, numerical, graphical, and statistical analyses confirm that MACPSO outperforms the other nine PSO variants in terms of both effectiveness and robustness.}, archive = {J_NEUCOM}, author = {Hui Chen and Li-Yong Shen and Chunmei Wang and Lei Tian and Sanguo Zhang}, doi = {10.1016/j.neucom.2025.129460}, journal = {Neurocomputing}, month = {4}, pages = {129460}, shortjournal = {Neurocomputing}, title = {Multi actors-critic based particle swarm optimization algorithm}, volume = {624}, year = {2025}, }
• (2025). PGFC-net: Parallel-encoding gaussian feature coordination-enhanced network for accurate 3D hepatic vessel and inferior vena cava segmentation. NEUCOM, 624, 129459. (www)

  @article{ , abstract = {Segmentation and visualisation of the hepatic vessels and inferior vena cava in three-dimensional computed tomography (CT) images are crucial for computer-aided diagnosis and preoperative planning. However, the multilayer distribution, complex branching, and low resolution of the hepatic vessels make image segmentation extremely challenging. This study aims to accurately segment the hepatic vessels and the inferior vena cava to address the difficulty in identifying hepatic vessels at low resolution and the issue of visual discontinuity caused by small vessels appearing fragmented in the segmentation results. This study proposes the Parallel-Encoding Gaussian Feature Coordination-Enhanced Network (PGFC-Net), which uses parallel encoders to accurately identify low-resolution hepatic vessels and the inferior vena cava by modelling global contextual relationships and extracting local features. To address the fragmented or disconnected vessel issue in the segmentation results, we introduce a novel feature fusion block called Gaussian Coordination-Enhanced Multi-Head Attention (GCEMA). This mechanism facilitates a more effective fusion and coordination of two features, thereby enhancing the robustness of the network. In a set of experiments using the public dataset 3DIRCADb01, the Dice similarity coefficient for hepatic vessels reached 82.16 %. In another set of experiments aimed at verifying the generalisation of the method by extending the dataset, the Dice similarity coefficients for the hepatic vessels and inferior vena cava were 81.31 % and 93.45 %, respectively. PGFC-Net effectively enhances the accuracy of the hepatic vessel and inferior vena cava segmentation tasks compare to the other nine advanced networks and notably reduces instances of vessel terminal fragmentation or disintegration in the outcomes, offering valuable insights for further research and potential clinical applications.}, archive = {J_NEUCOM}, author = {ShuYi Jiang and JiaYin Bao and Ming Yue and Kai Chen and Jian Wang}, doi = {10.1016/j.neucom.2025.129459}, journal = {Neurocomputing}, month = {4}, pages = {129459}, shortjournal = {Neurocomputing}, title = {PGFC-net: Parallel-encoding gaussian feature coordination-enhanced network for accurate 3D hepatic vessel and inferior vena cava segmentation}, volume = {624}, year = {2025}, }
• (2025). A dynamic decoder with speculative termination for low latency inference in spiking neural networks. NEUCOM, 624, 129458. (www)

  @article{ , abstract = {Spiking Neural Networks (SNNs) represent a novel class of biologically inspired neural networks. Their distinctive event-driven feature provides superior energy efficiency compared to conventional artificial neural networks, enabling operation on neuromorphic hardware with minimal power consumption. Rate coding stands as the prevailing coding schemes, characterized by robustness and good performance in classification tasks. However, it requires continuous inference over a specified number of time steps for each images, which limits the reduction of inference latency. This paper introduces a novel dynamic decoding method which select the final winner at the output stage from either the dominated winner at first time step (Win-First) or the output that wins continuous several time steps (Win-Continuously). The proposed decoder is capable of dynamically determining time when inference can be terminated prematurely to reduce inference latency. Compared to existing SOTA decoding methods, output selection by Win-First or Win-Continuously (WFWC) does not involve complex exponential or logarithmic operations, resulting in significantly low energy consumption and hardware overhead. Experimental evaluations are conducted on both directly trained SNNs and SNNs obtained through ANN-to-SNN conversion. On CIFAR-10 dataset, SNNs with WFWC decoder only need average 1.69 time steps to achieve an accuracy of 94.37%, which is equivalent to 2.36 × acceleration compared to baseline directly trained SNNs with fixed 4 time steps. Compared to SNNs obtained through ANN-to-SNN conversion with fixed 30 time steps, by using WFWC decoder, SNNs only require average 5.76 time steps to achieve an accuracy of 95.44%, equivalent to 5.20 × acceleration.}, archive = {J_NEUCOM}, author = {Yuxuan Yang and Zihao Xuan and Song Chen and Yi Kang}, doi = {10.1016/j.neucom.2025.129458}, journal = {Neurocomputing}, month = {4}, pages = {129458}, shortjournal = {Neurocomputing}, title = {A dynamic decoder with speculative termination for low latency inference in spiking neural networks}, volume = {624}, year = {2025}, }
• (2025). RPN: A region-to-pixel-mask-based convolutional network for lesion segmentation of fundus images. NEUCOM, 624, 129456. (www)

  @article{ , abstract = {Lesions segmentation of fundus images play a crucial role in the field of diagnosing diabetic retinopath. However, their performance has consistently been hindered by challenges arising from the small size of lesions, which leads to class imbalance issues. This paper proposes a two-branch rebalancing image segmentation model, inspired by the physiological structures of human peripheral and central vision, to jointly address the small lesion size and the resulting class imbalance problem. The peripheral visual branch predicts the categories of regions that aggregate multiple pixels. Any region containing a portion of the lesion will be categorized as a lesion region (positive samples). This significantly enhances the model’s sensitivity to lesions and increases the number of positive samples. However, it also prevents the model from accurately identifying lesion boundaries. Since the high sensitivity of peripheral vision comes at the cost of reduced resolution, the human visual system relies on central vision for detailed observation of objects of interest. Similarly, in the central vision branch, we focus only on the regions identified by the peripheral vision branch and carefully assess the boundaries of the lesions. This further balances the distribution of positive and negative samples. We evaluated the performance of our model using two publicly available datasets, namely IDRiD and DDR. Our proposed method achieved competitive results compared to state-of-the-art methods. The source code of RPN is available at https://github.com/zh-hongda/RPN .}, archive = {J_NEUCOM}, author = {Hongda Zhang and Chun Ouyang and Wei Lin and Zhonghong Shen and Bowen Liu and Yi Liu and Zhongxue Gan}, doi = {10.1016/j.neucom.2025.129456}, journal = {Neurocomputing}, month = {4}, pages = {129456}, shortjournal = {Neurocomputing}, title = {RPN: A region-to-pixel-mask-based convolutional network for lesion segmentation of fundus images}, volume = {624}, year = {2025}, }
• (2025). Bio-inspired neuron based on threshold selector and tunnel diode capable of excitability modulation. NEUCOM, 624, 129454. (www)

  @article{ , abstract = {Understanding the neuron behavior is essential for advancing power-efficient neuromorphic computing. This paper represents a novel model of the spiking neuron obtained by combining a threshold-switching selector with a tunnel diode. The study investigates the circuitry, identification of models for the basic nonlinear elements, and dynamical behavior of the proposed bio-inspired neuron. The proposed neuron circuit comprises only five elements forming three parallel branches: a volatile memristor in series with a tuning DC voltage source, a tunnel diode in series with a modulatory DC voltage source, and a capacitor. The bidirectional threshold switching selector device AND-TS serves as an example of a volatile memristor. In its empirical compact model, selected parameters enable replication of the main kinetic and dynamic characteristics exhibited by the AND-TS device. Models of four various germanium backward tunnel diodes (GI401A, GI403A, BD4, and BD5) are also identified, allowing accurate representation of their electrical characteristics. The study explores the dynamical behavior of the proposed neuron using high-resolution two-parametric diagrams. These diagrams reveal different operating modes, providing insight into the neuron’s response to different input stimuli. An assessment of the energy characteristics of the proposed neuron design is given. Additionally, we provide a SPICE model of the developed neuron for its rapid deployment in field solutions.}, archive = {J_NEUCOM}, author = {Valerii Ostrovskii and Timur Karimov and Vyacheslav Rybin and Yulia Bobrova and Viacheslav Arlyapov and Denis Butusov}, doi = {10.1016/j.neucom.2025.129454}, journal = {Neurocomputing}, month = {4}, pages = {129454}, shortjournal = {Neurocomputing}, title = {Bio-inspired neuron based on threshold selector and tunnel diode capable of excitability modulation}, volume = {624}, year = {2025}, }
• (2025). ChiGa-net: A genetically optimized neural network with refined deeply extracted features using χ2 statistical score for trustworthy parkinson’s disease detection. NEUCOM, 624, 129450. (www)

  @article{ , abstract = {Automated neurodegenerative disease detection systems that integrate computer vision and machine learning (ML) provide non-invasive, cost-effective tools to assist neurologists in diagnosing and monitoring disorders like Parkinson’s disease (PD). Handwriting analysis plays a crucial role in the automated assessment of PD. Although many ML methods utilizing either automated or hand-crafted features – or a combination of both – have shown promising results in PD detection, they often suffer from overestimation bias and struggle to produce consistent results when tested on new, unseen datasets. To address these challenges, we propose a method that extracts more representative features, ensuring both trustworthiness and generalizability for real-time applications. We introduce ChiGa-Net, a novel framework that integrates the χ 2 statistical score-based refined deep feature extraction with a genetically optimized neural network. To provide more practical results, this framework was trained and validated on a specific dataset, namely NewHandPD, and then rigorously tested on a new, unseen dataset, HandPD. Numerical experiments confirm the trustworthiness and generalizability of ChiGa-Net. For spiral images, the proposed approach achieves 92.56% balanced accuracy, 85.13% sensitivity, and 100% specificity during external testing. These results demonstrate superior performance compared to previously published methods and ML systems based on either deep learning or hand-crafted features. These findings suggest that the developed method yields reliable and generalizable results across diverse datasets, making it a robust and trustworthy tool for real-world clinical applications.}, archive = {J_NEUCOM}, author = {Liaqat Ali and Man-Fai Leung and Muhammad Asghar Khan and Redhwan Nour and Yakubu Imrana and Athanasios V. Vasilakos}, doi = {10.1016/j.neucom.2025.129450}, journal = {Neurocomputing}, month = {4}, pages = {129450}, shortjournal = {Neurocomputing}, title = {ChiGa-net: A genetically optimized neural network with refined deeply extracted features using χ2 statistical score for trustworthy parkinson’s disease detection}, volume = {624}, year = {2025}, }
• (2025). Lightweight visual backbone network with enhanced comprehensive strength through context-aware dual attention mechanism. NEUCOM, 624, 129449. (www)

  @article{ , abstract = {Recent studies have increasingly concentrated on integrating visual neural networks into edge devices, especially for robots and drones that require effective real-time decision-making and autonomous path planning. Achieving this necessitates lightweight visual networks with robust overall performance. In this study, we adopt a small-scale architecture based on convolutional neural networks (CNNs) and propose a Context-Aware Decoupled Fully Connected (CADFC) attention mechanism to enhance the performance of existing CNN-based visual networks. The core concept of the CADFC attention mechanism is to incorporate both remote and local contexts, enabling the network to capture local features and their relevant remote features. Moreover, we design a novel bottleneck that integrates CADFC attention with depthwise convolution. This design thinking is to accumulate focused regions, learning from depthwise convolution, alongside their surrounding features based on CADFC attention. Experimental results demonstrate that CADFC MobileNet outperforms the recent SOTA network GhostNetV2, achieving a top-1 accuracy of 76.8% with 15% fewer parameters, surpassing GhostNetV2’s 75.3% accuracy by 1.5%, even in resource-constrained environments.}, archive = {J_NEUCOM}, author = {Jianxin Xue and Yaohua Hu and Sicheng Hua and Minyu Chen and Ling-I Wu and Xi Chang and Guoqiang Li}, doi = {10.1016/j.neucom.2025.129449}, journal = {Neurocomputing}, month = {4}, pages = {129449}, shortjournal = {Neurocomputing}, title = {Lightweight visual backbone network with enhanced comprehensive strength through context-aware dual attention mechanism}, volume = {624}, year = {2025}, }
• (2025). Multi-relation graph contrastive learning with adaptive strategy for social recommendation. NEUCOM, 624, 129448. (www)

  @article{ , abstract = {Numerous social recommendation models leverage user–item rating data (i.e., collaborative domain) and social connections (i.e., social domain) to obtain users’ preferences for delivering personalized recommendations. However, existing algorithms for Top-N recommendation fail to fully exploit the diversity of the rating data, thereby overlooking certain implicit user preferences. Additionally, it is imperative to acknowledge the unique characteristics inherent in each object, necessitating careful consideration of such distinctions when designing recommendation algorithms. In this paper, we propose m ulti-relation graph c ontrastive l earning with an a daptive strategy (MCLA) for social recommendation. Specifically, we divide the user–item rating data and the multiple social data into distinct relations, treating each as an independent graph. Graph Neural Networks (GNNs) are then employed to aggregate messages from each relation separately. Furthermore, we use an adaptive fusion strategy to fuse user embeddings in the collaborative domain and social domain by utilizing the user’s characteristics. To alleviate the data sparsity problem after relation splitting, we present a simple yet effective adaptive noise contrastive learning method. Experiments on several real-world datasets demonstrate the effectiveness of the proposed MCLA. Both source code and benchmark data are available via https://github.com/chenai1018/MCLA .}, archive = {J_NEUCOM}, author = {Yuhan Xia and Yufei Tang and Bohang Yang and Chenghao Liu and Qian Tao}, doi = {10.1016/j.neucom.2025.129448}, journal = {Neurocomputing}, month = {4}, pages = {129448}, shortjournal = {Neurocomputing}, title = {Multi-relation graph contrastive learning with adaptive strategy for social recommendation}, volume = {624}, year = {2025}, }
• (2025). H-vmunet: High-order vision mamba UNet for medical image segmentation. NEUCOM, 624, 129447. (www)

  @article{ , abstract = {In the field of medical image segmentation, variant models based on Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have been extensively developed and applied. However, CNNs often struggle with processing long-sequence information, while Vision Transformers exhibit low sensitivity to local feature extraction and face challenges with computational complexity. Recently, the emergence of state-space models (SSMs), particularly 2D-selective-scan (SS2D), has challenged the longtime dominance of traditional CNNs and ViTs as foundational modules in visual neural networks. In this paper, we extend the adaptation of SS2D by proposing a High-order Vision Mamba UNet (H-vmunet) model for medical image segmentation. Among them, the H-vmunet model includes the proposed novel High-order 2D-selective-scan (H-SS2D) and Local-SS2D module. The H-SS2D is used to reduce the introduction of redundant information when SS2D features are feature-learning in the global receptive field. The Local-SS2D module is used to improve the learning ability of local features in SS2D. We conducted comprehensive comparison and ablation experiments on three publicly available medical image datasets (ISIC2017, Spleen, and CVC-ClinicDB), and the results consistently demonstrate the strong performance of H-vmunet in medical image segmentation tasks. The code is available from https://github.com/wurenkai/H-vmunet .}, archive = {J_NEUCOM}, author = {Renkai Wu and Yinghao Liu and Pengchen Liang and Qing Chang}, doi = {10.1016/j.neucom.2025.129447}, journal = {Neurocomputing}, month = {4}, pages = {129447}, shortjournal = {Neurocomputing}, title = {H-vmunet: High-order vision mamba UNet for medical image segmentation}, volume = {624}, year = {2025}, }
• (2025). From theory to practice: The evolution and comparative analysis of homogeneous vs. heterogeneous graph neural networks in recommender systems. NEUCOM, 624, 129446. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) have emerged as powerful tools for recommendation systems, addressing challenges such as complex user-item relationships and dynamic behaviors. This paper provides a concise review of GNN applications, focusing on embedding techniques and state-of-the-art algorithms. We evaluate three GNN architectures: Homogeneous GNN, Heterogeneous GNN, and a Heterogeneous GNN enhanced with Skip-Gram node embeddings. These models are assessed on subsets of the Amazon 2023 dataset, covering Fashion, Beauty, and Musical Instruments, using metrics such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Our findings highlight the superior performance of GNNs in capturing nuanced user-item interactions, improving recommendation accuracy, scalability, and adaptability. The integration of Skip-Gram embeddings further enhances item similarity modeling, enabling more personalized recommendations. We also analyze the computational efficiency of these models, offering insights for their deployment in large-scale systems. This study bridges the gap between theoretical advancements and practical applications of GNNs in recommendation systems. By synthesizing recent trends, identifying research gaps, and presenting actionable insights, it serves as a foundational reference for researchers and practitioners aiming to optimize GNN-based models for diverse scenarios. The results underscore the transformative potential of GNNs in delivering accurate, scalable, and real-time recommendations, setting the stage for future innovations in this domain.}, archive = {J_NEUCOM}, author = {Maryam Khanian Najafabadi and Rei-An Chen and Javad Rezazadeh and Amin Beheshti and Nasrin Shabani}, doi = {10.1016/j.neucom.2025.129446}, journal = {Neurocomputing}, month = {4}, pages = {129446}, shortjournal = {Neurocomputing}, title = {From theory to practice: The evolution and comparative analysis of homogeneous vs. heterogeneous graph neural networks in recommender systems}, volume = {624}, year = {2025}, }
• (2025). Local and global structure-aware contrastive framework for entity alignment. NEUCOM, 624, 129445. (www)

  @article{ , abstract = {Entity alignment (EA) seeks to identify equivalent real-world entities across different knowledge graphs. Recently, integrating graph neural networks (GNNs) with graph augmentation techniques to aggregate local structural information of entities has been proven effective for EA. However, stacking multiple GNN layers to capture higher-order neighbors often leads to oversmoothing of entity embeddings and the introduction of noise from irrelevant neighbors. In this paper, we propose a novel approach, Local and Global Structure-Aware Contrastive Framework (LGEA), to effectively learn the mutual information between the local and global structures of entities. Specifically, we propose a graph augmentation method using Singular Value Decomposition to capture the global structure. In the Global Structure-Aware Encoder module, we design a Residual Gated Unit to reduce noise and mitigate oversmoothing. LGEA incorporates contrastive learning to maximize the consistency between local and global embeddings. Additionally, we introduce a Degree-Aware Relation Encoder to integrate relational semantic information, enriching the entity embeddings. Extensive experiments on established EA benchmarks demonstrate that our method significantly outperforms previous approaches.}, archive = {J_NEUCOM}, author = {Cunda Wang and Weihua Wang and Qiuyu Liang and Guanglai Gao}, doi = {10.1016/j.neucom.2025.129445}, journal = {Neurocomputing}, month = {4}, pages = {129445}, shortjournal = {Neurocomputing}, title = {Local and global structure-aware contrastive framework for entity alignment}, volume = {624}, year = {2025}, }
• (2025). FSANet: Feature shuffle and adaptive channel attention network for arbitrary shape scene text detection. NEUCOM, 624, 129443. (www)

  @article{ , abstract = {Natural scene text detection has made significant progress in the era of deep learning. However, existing methods still exhibit deficiencies when faced with challenges such as complex backgrounds, extreme aspect ratios, and arbitrary-shaped text. To address these issues, we propose a segmentation-based Feature Shuffle Attention Network (FSANet) designed to enhance high-resolution feature extraction and multi-scale feature enhancement for robust detection of arbitrary-shaped text. FSANet is composed of two principal modules: (1) the High-Resolution Feature Extraction Network (FEN), which employs two Group Shuffle Blocks (GSBs) to maintain high-resolution details and promote feature interaction and information flow, and (2) the Adaptive Channel Attention Module (ACAM), which reduces background noise and redundant features by adaptively learning inter-feature correlations across scales, assigning weights to prioritize local features within a global context. Extensive experiments conducted on four public benchmark datasets show that, compared to the baseline method, the proposed method demonstrates an improved F-measure on the ICDAR2015, Total-Text, MSRA-TD500, and ICDAR2017-MLT datasets, with an average increase of 1.68%. The recall metric also consistently improves, with an average increase of 2.0%. Notably, the proposed method achieves the highest F-measure of 85.4%, 75.9% on the ICDAR2015, ICDAR2017-MLT datasets, respectively. Furthermore, on the other two datasets, the performance of the proposed method surpasses that of most existing methods, indicating that FSANet outperforms the majority of state-of-the-art approaches. The code will be publicly released at https://github.com/runminwang/FSANet .}, archive = {J_NEUCOM}, author = {Juan Xu and Runmin Wang and Jielei Hei and Xiaofei Cao and Zukun Wan and Congzhen Yu and Yajun Ding and Changxin Gao and Shengyou Qian}, doi = {10.1016/j.neucom.2025.129443}, journal = {Neurocomputing}, month = {4}, pages = {129443}, shortjournal = {Neurocomputing}, title = {FSANet: Feature shuffle and adaptive channel attention network for arbitrary shape scene text detection}, volume = {624}, year = {2025}, }
• (2025). A doubly decoupled network for edge detection. NEUCOM, 624, 129442. (www)

  @article{ , abstract = {Deep learning-based edge detectors are increasingly scaled up to achieve higher performance. However, this growth comes at the cost of higher computational demands and a greater risk of model overfitting. In this paper, a Doubly Decoupled Network (DDN) is proposed to alleviate these problems by decoupling the data and features separately. Firstly, we decouple image-edge pairs and build learnable Gaussian distributions. The images are mapped to learnable Gaussian distributions, and the corresponding edges are mapped to samples of the distributions. More robust features are generated by normalizing the latent variables and sampling from the Gaussian distributions. Secondly, we compress the redundant calculations by decoupling image features. Shallow features need not only high resolution to provide accurate spatial information, but also diversity to provide sufficient cues for deep features. We decouple image features into spatial and semantic features and encode them separately. Shallow features no longer provide cues for semantic features and their diversity can be drastically compressed. The proposed DDN achieves the state-of-the-art accuracy across multiple edge detection benchmarks, while the computational cost is similar to that of VGG-based methods. The code is available at https://github.com/Li-yachuan/DDN .}, archive = {J_NEUCOM}, author = {Yachuan Li and Xavier Soria Poma and Yongke Xi and Guanlin Li and Chaozhi Yang and Qian Xiao and Yun Bai and Zongmin Li}, doi = {10.1016/j.neucom.2025.129442}, journal = {Neurocomputing}, month = {4}, pages = {129442}, shortjournal = {Neurocomputing}, title = {A doubly decoupled network for edge detection}, volume = {624}, year = {2025}, }
• (2025). Temporal multivariate-factors independence convolution network for anomaly detection in dynamic networks. NEUCOM, 624, 129439. (www)

  @article{ , abstract = {Anomaly detection in dynamic networks is an important research task in many application domains, such as social media, financial fraud, and e-commerce. It aims to identify the anomalous objects (such as nodes or edges) that are different from the normal objects in dynamic networks with the evolution of graph structure or node features. Network embedding has attracted considerable interest in anomaly detection, which maps each node into a low-dimensional embedding space vector and preserves the topological structure and features to perform the anomaly detection task. However, existing network embedding works treat the neighbors in the same view to learn the node representations for anomaly detection, which may not be perfectly suitable for the evolving attributes of neighbors acting different effects on the node representations. In this paper, we propose a novel approach, Temporal Multivariate-factors Independence Convolution Network (TMICN), which learns the node representations and detects anomalies in dynamic networks. Specifically, we first propose a Features Evolving module to grasp dynamics patterns from the temporal features within the time window of the observation nodes. Afterwards, we design a Multivariate-factors Learning module treating the outputting of the Features Evolving module as different views, which utilizes the neighborhood routing to learn evolving relationships between nodes and adopt the embedding dissemination to consider the influence of high-order neighbors for updating the node representations constantly. In addition, we adopt an Independence Modeling module to achieve independence between the different factor representations. Extensive experiments on real-world datasets show that our model effectively improves the performance of anomaly detection in dynamic networks.}, archive = {J_NEUCOM}, author = {Yang Yu and Minglai Shao and Xin Li and Wenjun Wang}, doi = {10.1016/j.neucom.2025.129439}, journal = {Neurocomputing}, month = {4}, pages = {129439}, shortjournal = {Neurocomputing}, title = {Temporal multivariate-factors independence convolution network for anomaly detection in dynamic networks}, volume = {624}, year = {2025}, }
• (2025). Bridging the metrics gap in image style transfer: A comprehensive survey of models and criteria. NEUCOM, 624, 129430. (www)

  @article{ , abstract = {Image style transfer is a technique that combines the content of a real photograph with the artistic style of another image to create a new and stylized image. In this paper, we aim to provide a comprehensive review of the field, tracing its development from traditional methods to modern neural network-based approaches. We focus on recent advancements utilizing Transformers and diffusion models, which represent significant progress in generative AI. Furthermore, we systematically analyze the objective evaluation metrics employed in style transfer, offering valuable insights for researchers. Finally,this paper highlights unresolved issues and suggests future directions to foster innovation in style transfer techniques.}, archive = {J_NEUCOM}, author = {Xiaotong Zhou and Yuhui Zheng and Junming Yang}, doi = {10.1016/j.neucom.2025.129430}, journal = {Neurocomputing}, month = {4}, pages = {129430}, shortjournal = {Neurocomputing}, title = {Bridging the metrics gap in image style transfer: A comprehensive survey of models and criteria}, volume = {624}, year = {2025}, }
• (2025). REHRSeg: Unleashing the power of self-supervised super-resolution for resource-efficient 3D MRI segmentation. NEUCOM, 624, 129425. (www)

  @article{ , abstract = {High-resolution (HR) 3D magnetic resonance imaging (MRI) can provide detailed anatomical structural information, enabling precise segmentation of regions of interest for various medical image analysis tasks. Due to the high demands of acquisition device, collection of HR images with their annotations is always impractical in clinical scenarios. Consequently, segmentation results based on low-resolution (LR) images with large slice thickness are often unsatisfactory for subsequent tasks. In this paper, we propose a novel Resource-Efficient High-Resolution Segmentation framework (REHRSeg) to address the above-mentioned challenges in real-world applications, which can achieve HR segmentation while only employing the LR images as input. REHRSeg is designed to leverage self-supervised super-resolution (self-SR) to provide pseudo supervision, therefore the relatively easier-to-acquire LR annotated images generated by 2D scanning protocols can be directly used for model training. The main contribution to ensure the effectiveness in self-SR for enhancing segmentation is three-fold: (1) We mitigate the data scarcity problem in the medical field by using pseudo-data for training the segmentation model. (2) We design an uncertainty-aware super-resolution (UASR) head in self-SR to raise the awareness of segmentation uncertainty as commonly appeared on the ROI boundaries. (3) We align the spatial features for self-SR and segmentation through structural knowledge distillation to enable a better capture of region correlations. Experimental results demonstrate that REHRSeg achieves high-quality HR segmentation without intensive supervision, while also significantly improving the baseline performance for LR segmentation. Our codes are publicly available at https://github.com/zhiyuns/REHRSeg .}, archive = {J_NEUCOM}, author = {Zhiyun Song and Yinjie Zhao and Xiaomin Li and Manman Fei and Xiangyu Zhao and Mengjun Liu and Cunjian Chen and Chung-Hsing Yeh and Qian Wang and Guoyan Zheng and Songtao Ai and Lichi Zhang}, doi = {10.1016/j.neucom.2025.129425}, journal = {Neurocomputing}, month = {4}, pages = {129425}, shortjournal = {Neurocomputing}, title = {REHRSeg: Unleashing the power of self-supervised super-resolution for resource-efficient 3D MRI segmentation}, volume = {624}, year = {2025}, }
• (2025). Seeking the optimal accuracy-rate equilibrium in face recognition. NEUCOM, 624, 129423. (www)

  @article{ , abstract = {Considerable facial data is often compressed prior to analysis to accommodate limitations in transmission or storage capacities. However, this compression may lead to the loss of crucial identity details, thereby diminishing the effectiveness of facial recognition (FR) systems. In this study, we aim to establish an optimal accuracy-rate equilibrium (ARE) to maximize the compression ratio without substantially compromising the performance of the FR system. We first investigate the effect of image compression in deep FR. Based on the definition of ARE in FR, we proposed a method to locate the ARE values of face images with true acceptance rate and false accept rate. Subsequently, we develop an ARE prediction method for the FR system (ARE-FR), which automatically infers ARE images of face images. The goal of the proposed ARE-FR is to maximize redundancy removal without impairment of robust identity information. Considering that high-level semantic features effectively capture crucial identity information, we force the proposed ARE-FR to focus only on the features in identity-related regions when predicting the ARE images. These features are derived from the interactive relationships between deep and shallow features. Experimental results have demonstrated that combining our proposed ARE-FR with the image coding algorithm is capable of saving more bits while maintaining the performance of the FR system.}, archive = {J_NEUCOM}, author = {Yu Tian and Fu-Zhao Ou and Shiqi Wang and Baoliang Chen and Sam Kwong}, doi = {10.1016/j.neucom.2025.129423}, journal = {Neurocomputing}, month = {4}, pages = {129423}, shortjournal = {Neurocomputing}, title = {Seeking the optimal accuracy-rate equilibrium in face recognition}, volume = {624}, year = {2025}, }
• (2025). IMedSeg: Towards efficient interactive medical segmentation. NEUCOM, 624, 129419. (www)

  @article{ , abstract = {Click-based Interactive Medical Segmentation (IMS) aims to segment target objects under limited user interactions. Although existing deep learning-based methods achieve promising performance, they still exhibit deficiencies: (1) they fail to fully and explicitly utilize click information and (2) when used to refine preexisting masks, they cannot avoid destroying correctly segmented regions. Against these issues, we propose IMedSeg to formulate the IMS task into a progressive loop process, where each click by the user results in a precise and efficient modification. In particular, we design a CNN/Transformer-based model to perform the initial segmentation on the target image. To enhance segmentation accuracy, we especially propose a memory-efficient interactive medical segmentation backbone called ICS-Unet by incorporating CSWin Transformer with Unet. Then, a Similarity Mask Correction Module (SMCM) is proposed to refine segmentation only on the local region with user clicks to avoid corrupting the correct mask. Moreover, a Compressed Similarity Aggregator (CSA) is constructed for computing the correlations between the deep features of clicked and unclicked points, making full use of click information for pixel-wise classification of the local region. Note that our proposed method is a generic framework that can be readily combined with existing interactive segmentation backbones. Extensive experiments show that our method attains superior performance over state-of-the-arts in terms of accuracy, memory and inference time. The results on ISIC-2017 and OAI-ZIB datasets reveal average Dice coefficients of 96.18% and 87.44%, and IoUs of 82.97% and 78.19%, respectively.}, archive = {J_NEUCOM}, author = {Zidi Shi and Ting Chen and Qian Zhou and Hua Zou and Xuan Xiao}, doi = {10.1016/j.neucom.2025.129419}, journal = {Neurocomputing}, month = {4}, pages = {129419}, shortjournal = {Neurocomputing}, title = {IMedSeg: Towards efficient interactive medical segmentation}, volume = {624}, year = {2025}, }
• (2025). A novel dynamic nonlinear non-gaussian approach for fault detection and diagnosis. NEUCOM, 624, 129403. (www)

  @article{ , abstract = {Fault detection and diagnosis are crucial for the safe operation of modern industrial systems. Nevertheless, the inherent dynamics, nonlinearities, and non-Gaussian disturbances of complex systems pose great challenges. To enhance fault detection and diagnosis in those environments, a novel approach that combines local-linearized gate recursive unit (LGRU) with high-order principal component analysis (HPCA) is proposed. The LGRU modeling methodology, which has similar nonlinear dynamics modeling capabilities as GRU, is introduced to improve the temporal separability of system dynamics errors caused by modeling inaccuracies or faults. Furthermore, to more effectively identify faults with both Gaussian and non-Gaussian spatial distribution, high-order statistics are utilized to establish HPCA latent variable extraction criteria. Based on these criteria, a relatively high-order decomposition contribution (rHDC) index is constructed to identify the fault-causative variables. Comparative simulations of numerical examples and TE processes demonstrate the efficacy and sophistication of the proposed methodology.}, archive = {J_NEUCOM}, author = {Yihan Ma and Fei Ye and Dazi Li and Jinglin Zhou}, doi = {10.1016/j.neucom.2025.129403}, journal = {Neurocomputing}, month = {4}, pages = {129403}, shortjournal = {Neurocomputing}, title = {A novel dynamic nonlinear non-gaussian approach for fault detection and diagnosis}, volume = {624}, year = {2025}, }
• (2025). DSVTN-ASD: Detection of stereotypical behaviors in individuals with autism spectrum disorder using a dual self-supervised video transformer network. NEUCOM, 624, 129397. (www)

  @article{ , abstract = {Early detection of Autism Spectrum Disorder (ASD) is essential for improving the standard of life for those with autism. Traditional diagnostic approaches in hospitals are challenging and expensive, due to time constraints, difficulties in contextual understanding, the intermittent nature of behaviors, and clinicians' subjective observations. Recognizing the limitations of traditional methods, there is a rising interest in using technology, such as video analysis, to determine stereotypical behaviors associated with autism. Existing activity recognition techniques using supervised learning are based on the datasets with labels. However, this is practically impossible due to the difficulty of precisely labeling the recurring behaviors and the size of the dataset. To alleviate this, a self-supervised model called the Dual Self-Supervised Video Transformer Network (DSVTN-ASD) is proposed with three proxy tasks to enhance its learning capabilities. This model is trained using a few samples of labeled autism videos, and unlabeled video samples of normal and ASD behavior. During inference, pseudo-labels are generated for the unlabeled samples that help in anomaly detection. The pose key point extraction phase and repeated behavior detection are incorporated to predict each pose's key points and identify abnormal behavior. The experimentation is validated on the extended Self-Stimulatory Behavior Dataset (SSBD) and achieves Micro and Macro Area Under the Receiver Operating Characteristic Curve (AUROC) of 95.01 % and 93.13 %, respectively. This AUROC value is better than the other unsupervised strategies in the literature, and thus it can serve as a benchmark for carrying out work based on self-supervised learning.}, archive = {J_NEUCOM}, author = {Asmetha Jeyarani R. and Radha Senthilkumar}, doi = {10.1016/j.neucom.2025.129397}, journal = {Neurocomputing}, month = {4}, pages = {129397}, shortjournal = {Neurocomputing}, title = {DSVTN-ASD: Detection of stereotypical behaviors in individuals with autism spectrum disorder using a dual self-supervised video transformer network}, volume = {624}, year = {2025}, }
• (2025). Residual-time gated recurrent unit. NEUCOM, 624, 129396. (www)

  @article{ , abstract = {Recurrent neural networks (RNNs) are well-suited for sequential data processing, which have been widely used in natural language processing, speech recognition, and other sequential signal recognition tasks. However, for long sequential data processing, as the number of network layers increases, conventional RNNs are prone to problems such as vanishing gradients and weak capacity of memorizing long-term information, leading to reduced accuracy. In this article, we propose a residual-time gated recurrent unit (RT-GRU) based deep learning model to address these problems. The RT-GRU model introduces residual information into the candidate hidden state representation of the GRU in the backpropagation direction, making the network more sensitive to gradient changes. Additionally, it establishes a residual connection in the backpropagation through time direction, allowing the network to capture long-term dependencies. Moreover, the gating mechanism of the highway networks is utilized to control the residual information flow. The results of the adding problem, language modeling, text classification, and sentiment analysis experiments demonstrate that RT-GRU achieves a superior performance over several state-of-the-art baseline methods on some well-known datasets.}, archive = {J_NEUCOM}, author = {Yirong Wu and Feng Hu and Chonghao Yue and Shuifa Sun}, doi = {10.1016/j.neucom.2025.129396}, journal = {Neurocomputing}, month = {4}, pages = {129396}, shortjournal = {Neurocomputing}, title = {Residual-time gated recurrent unit}, volume = {624}, year = {2025}, }
• (2025). Fair streaming feature selection. NEUCOM, 624, 129394. (www)

  @article{ , abstract = {Streaming feature selection techniques have become essential in processing real-time data streams, as they facilitate the identification of the most relevant attributes from continuously updating information. Despite their performance, current algorithms for streaming feature selection frequently fall short in managing biases and avoiding discrimination that could be perpetuated by sensitive attributes, potentially leading to unfair outcomes in the resulting models. To address this issue, we propose FairSFS, a novel algorithm for Fair S treaming F eature S election, to uphold fairness in the feature selection process without compromising the ability to handle data in an online manner. FairSFS adapts to incoming feature vectors by dynamically adjusting the feature set and discerns the correlations between classification attributes and sensitive attributes from this revised set, thereby forestalling the propagation of sensitive data. Empirical evaluations show that FairSFS not only maintains accuracy that is on par with leading streaming feature selection methods and existing fair feature techniques but also significantly improves fairness metrics.}, archive = {J_NEUCOM}, author = {Zhangling Duan and Tianci Li and Zhaolong Ling and Xingyu Wu and Jingye Yang and Zhaohong Jia}, doi = {10.1016/j.neucom.2025.129394}, journal = {Neurocomputing}, month = {4}, pages = {129394}, shortjournal = {Neurocomputing}, title = {Fair streaming feature selection}, volume = {624}, year = {2025}, }
• (2025). A generalized spiking locally competitive algorithm for multiple optimization problems. NEUCOM, 624, 129392. (www)

  @article{ , abstract = {We introduce a generalized Spiking Locally Competitive Algorithm (LCA) that is biologically plausible and exhibits adaptability to a large variety of neuron models and network connectivity structures. In addition, we provide theoretical evidence demonstrating the algorithm’s convergence in optimization problems of signal recovery. Furthermore, our algorithm demonstrates superior performance over traditional optimization methods, such as FISTA, particularly by achieving faster early convergence in practical scenarios including signal denoising, seismic wave detection, and computed tomography reconstruction. Notably, our algorithm is compatible with neuromorphic chips, such as Loihi, facilitating efficient multitasking within the same chip architecture—a capability not present in existing algorithms. These advancements make our generalized Spiking LCA a promising solution for real-world applications, offering significant improvements in execution speed and flexibility for neuromorphic computing systems.}, archive = {J_NEUCOM}, author = {Xuexing Du and Zhong-qi K. Tian and Songting Li and Douglas Zhou}, doi = {10.1016/j.neucom.2025.129392}, journal = {Neurocomputing}, month = {4}, pages = {129392}, shortjournal = {Neurocomputing}, title = {A generalized spiking locally competitive algorithm for multiple optimization problems}, volume = {624}, year = {2025}, }
• (2025). Hypnos: A domain-specific large language model for anesthesiology. NEUCOM, 624, 129389. (www)

  @article{ , abstract = {The recent success of large language models (LLMs) has sparked a growing interest in their domain-specific tuning for medical applications. However, sufficient data collection in highly specialized domains such as anesthesiology poses significant challenges. In this work, we explore training a domain-specific LLM for anesthesiology, referred to as Hypnos, with limited real-world data, employing a progressive general-to-specific strategy. We incorporate medical question answering (QA) data collected at three different scales and specific levels: (1) 8 million pieces of general medical QA data from public websites; (2) approximately 180k pieces of synthetic anesthesia QA data generated by popular medical LLMs; and (3) approximately 35k pieces of real-world anesthesia QA data collected from anesthesia websites or books. The training process consists of two stages: fine-tuning a pretrained general LLM using general medical data and real-world anesthesia data, followed by further fine-tuning using synthetic and real-world anesthesia data. To evaluate the proposed Hypnos model, we conduct intensive experiments on a newly introduced anesthesia QA benchmark. The experimental results demonstrate that the proposed Hypnos outperforms recent medical LLMs in terms of widely used metrics (such as BLEU, ROUGE, and GLEU) as well as GPT-4 and human evaluations.}, archive = {J_NEUCOM}, author = {Zhonghai Wang and Jie Jiang and Yibing Zhan and Bohao Zhou and Yanhong Li and Chong Zhang and Baosheng Yu and Liang Ding and Hua Jin and Jun Peng and Xu Lin and Weifeng Liu}, doi = {10.1016/j.neucom.2025.129389}, journal = {Neurocomputing}, month = {4}, pages = {129389}, shortjournal = {Neurocomputing}, title = {Hypnos: A domain-specific large language model for anesthesiology}, volume = {624}, year = {2025}, }
• (2025). Adaptively forget with crossmodal and textual distillation for class-incremental video captioning. NEUCOM, 624, 129388. (www)

  @article{ , abstract = {With the increasing volume of video data, it is critical that intelligent systems have the ability to continuously analyze the increasing number of videos while avoiding the usage of data from early tasks. These video data are difficult to store, and training models from scratch face privacy and high energy consumption pitfalls. However, existing research has primarily focused on incremental learning for image and video classification, with limited attention given to analyzing multimodal inputs and leveraging the spatio-temporal contextual information embedded in videos to generate meaningful captions. In this paper, we first present the C lass- I ncremental V ideo C aptioning (CI-VC) task. There are two difficulties in this task, which are the stability–plasticity dilemma faced in processing textual information with spatio-temporal states and correlating video-text inter-modal information, respectively. To fill this unexplored area, we primitively propose a method to A daptively F orget with crossmodal and textual D istillation for class-incremental video captioning (AFD). Considering the unique encoder–decoder networks of captioning and the privacy of input data, we refer to the three stages without playback of old samples: visual encoding, textual decoding and network supervision. In the phase of learning new knowledge, we use F ine- g rained S ensitivity S election (FgSS) to refine the selection of stored knowledge from the old model, forgetting some irrelevant information to make room for the memory of new classes. In order to memorize the text information during decoding, we design G rowing Word Embedding ( G ) with elastic capacity to store the embedding tokens of text that have already been seen as textual meta-signal pairs to be retrieved the lexical information during further caption generation. Meanwhile, the D ual- l evel K nowledge D istillation (DlKD) is performed at both cross-modal semantic information and the textual final output to constrain from the perspective of the extracted level specific features, integrating the newly learned knowledge and consolidating the inter-model and intra-model knowledge of the old classes. To illustrate the ability of our model to resist forgetting, we designed a metric CIDER ˜ t to detect the stage forgetting rate. Experiments on the public datasets MSR-VTT, MSVD and VATEX show that the proposed method significantly resists the forgetting of previous tasks without replaying old samples, and performs well on the new task.}, archive = {J_NEUCOM}, author = {Huiyu Xiong and Lanxiao Wang and Heqian Qiu and Taijin Zhao and Benliu Qiu and Hongliang Li}, doi = {10.1016/j.neucom.2025.129388}, journal = {Neurocomputing}, month = {4}, pages = {129388}, shortjournal = {Neurocomputing}, title = {Adaptively forget with crossmodal and textual distillation for class-incremental video captioning}, volume = {624}, year = {2025}, }
• (2025). SwiftPCN: Fast, implementation-efficient, and accurate point cloud completion network with flexible output resolution. NEUCOM, 624, 129387. (www)

  @article{ , abstract = {Existing point cloud completion networks (PCCNs) have achieved remarkable completion accuracy, but often at the cost of inference speed and memory consumption. Furthermore, the networks’ output resolution is often fixed, which could lead to a mismatched point density when the input scans are of different resolutions. In this work, we propose SwiftPCN, a pair of novel PCCNs that can recover the missing points accurately and efficiently. SwiftPCN consists of two novel networks: (1) SwiftPCN-fixed that achieves superior trade-off between completion accuracy, inference latency, and number of parameters when compared with existing networks, and (2) SwiftPCN-flex that can generate point clouds of various resolution. In SwiftPCN-fixed, we utilize the prototypical encoder, a novel encoder that enables efficient modeling of binary interactions between input points. Meanwhile, in SwiftPCN-flex, we propose an implementation-efficient occupancy network that extracts the implicit representation of point clouds using a simplified prototypical encoder and 2D convolutions. We evaluate SwiftPCN networks on synthetic and real-world datasets, in which they successfully maintain competitive completion accuracy while reducing inference latency and number of parameters by 10-20 times compared with previous state-of-the-art networks, demonstrating its suitability for real-world applications where resources are limited and real-time inference is necessary. The codes are available at https://github.com/kaist-avelab/swift-pcn .}, archive = {J_NEUCOM}, author = {Kevin Tirta Wijaya and Dong-Hee Paek and Seung-Hyun Kong}, doi = {10.1016/j.neucom.2025.129387}, journal = {Neurocomputing}, month = {4}, pages = {129387}, shortjournal = {Neurocomputing}, title = {SwiftPCN: Fast, implementation-efficient, and accurate point cloud completion network with flexible output resolution}, volume = {624}, year = {2025}, }
• (2025). Classification and regression task integration in distillation for object detectors. NEUCOM, 624, 129386. (www)

  @article{ , abstract = {Knowledge distillation is a popular technique for model compression. However, since knowledge distillation originated in image classification, it primarily concentrates on classification tasks in object detection and often overlooks regression tasks. Emphasizing the classification task while neglecting the regression task can lead to skewed predictions about the model’s learning condition in knowledge distillation methods. To address this, we propose Task Integration Distillation (TID), a method that integrates both classification and regression tasks, enhancing the model’s ability to accurately capture the learning condition. Inspired by real-world teaching strategies and the concept of learning conditions, TID emphasizes the importance of features derived from both tasks, ensuring a balanced consideration of key and weak areas. We quantify and map the outputs of the two tasks from the object detector onto the feature map. Through this output mapping, we identify key areas and weak areas of the features to assess the model’s current learning state. Extensive experiments demonstrate that TID consistently outperforms existing methods, with a notable increase of about 2.0% in mean Average Precision (mAP) over recent feature decoupling and distillation approaches.}, archive = {J_NEUCOM}, author = {Hai Su and Zhenwen Jian and Yanghui Wei and Songsen Yu}, doi = {10.1016/j.neucom.2025.129386}, journal = {Neurocomputing}, month = {4}, pages = {129386}, shortjournal = {Neurocomputing}, title = {Classification and regression task integration in distillation for object detectors}, volume = {624}, year = {2025}, }
• (2025). Pseudo contrastive learning for graph-based semi-supervised learning. NEUCOM, 624, 129375. (www)

  @article{ , abstract = {Pseudo Labeling is a technique used to improve the performance of semi-supervised Graph Neural Networks (GNNs) by generating additional pseudo-labels based on confident predictions. However, the quality of generated pseudo-labels has been a longstanding concern due to the sensitivity of the classification objective with respect to the given labels. To avoid the untrustworthy classification supervision indicating “a node belongs to a specific class”, we favor the fault-tolerant contrasting supervision demonstrating “two nodes do not belong to the same class”. Thus, the problem of generating high-quality pseudo-labels is then transformed into a relaxed version, i.e ., identifying reliable negative pairs. To achieve this, we propose a general framework for GNNs, termed Pseudo Contrastive Learning (PCL). It separates two nodes whose positive and negative pseudo-labels target the same class. To incorporate topological knowledge into learning, we devise a topologically weighted contrastive loss that spends more effort separating negative pairs with smaller topological distances. Experimentally, we apply PCL to various GNNs, which consistently outperform their counterparts using other popular general techniques on five real-world graphs.}, archive = {J_NEUCOM}, author = {Weigang Lu and Ziyu Guan and Wei Zhao and Yaming Yang and Yuanhai Lv and Lining Xing and Baosheng Yu and Dacheng Tao}, doi = {10.1016/j.neucom.2025.129375}, journal = {Neurocomputing}, month = {4}, pages = {129375}, shortjournal = {Neurocomputing}, title = {Pseudo contrastive learning for graph-based semi-supervised learning}, volume = {624}, year = {2025}, }
• (2025). DIRel: Joint relational triple extraction through dual implicit relation. NEUCOM, 624, 129374. (www)

  @article{ , abstract = {Joint extraction of entities and relations from unstructured texts is essential in information extraction and knowledge graph construction. Existing approaches usually extract entities in one identical relation space to construct relational triples, but ignore the difference between the relation of entity pairs and the relation of token pairs in entity pairs. Therefore, previous joint methods suffer from the problem of coarse-grained relations and redundant information. To address this issue, we propose a novel joint entity and relation extraction model based on dual implicit relation, named DIRel. Specifically, our model consists of an implicit relation predictor and a dual implicit relation tagging scheme. The former predicts all possible implicit relations belonging to each token pair, respectively. The latter ensures an effective decoding process that extracts relational triples according to the implicit relations predicted. With comprehensive experiments on three widely used datasets, we demonstrate that DIRel is more effective and computationally efficient.}, archive = {J_NEUCOM}, author = {Liang Zhang and Nan Zheng}, doi = {10.1016/j.neucom.2025.129374}, journal = {Neurocomputing}, month = {4}, pages = {129374}, shortjournal = {Neurocomputing}, title = {DIRel: Joint relational triple extraction through dual implicit relation}, volume = {624}, year = {2025}, }
• (2025). Evolving ensemble model based on hilbert schmidt independence criterion for task-free continual learning. NEUCOM, 624, 129370. (www)

  @article{ , abstract = {Continual Learning (CL) aims to extend the abilities of deep learning models for continuously acquiring new knowledge without forgetting. However, most CL studies assume that task identities and boundaries are known, which is not a realistic assumption in a real scenario. In this work, we address a more challenging and realistic situation in CL, namely the Task-Free Continuous Learning (TFCL), where an ensemble of experts is trained on non-stationary data streams without having any task labels. To deal with TFCL, we introduce the Evolving Ensemble Model (EEM), which can dynamically build new experts into a mixture, thus adapting to the changing data distributions while continuously learning new data sets. To ensure a compact network architecture for EEM during training, we propose a novel expansion mechanism that considers the Hilbert–Schmidt Independence Criterion (HSIC) for evaluating the statistical consistency between the knowledge learned by each expert and that corresponding to the given data. This expansion mechanism does not require storing all previous samples and is more efficient as it performs statistical evaluations in a low-dimensional feature space inferred by a deep network. We also propose a new dropout mechanism for selectively removing unimportant stored samples from the memory buffer used for storing the continuously incoming data before they are used for training. The proposed dropout mechanism ensures the diversity of information being learnt by the experts of our model. We perform extensive TFCL tests which show that the proposed approach achieves the state of the art. The source code is available at https://github.com/dtuzi123/HSCI-DEM .}, archive = {J_NEUCOM}, author = {Fei Ye and Adrian G. Bors}, doi = {10.1016/j.neucom.2025.129370}, journal = {Neurocomputing}, month = {4}, pages = {129370}, shortjournal = {Neurocomputing}, title = {Evolving ensemble model based on hilbert schmidt independence criterion for task-free continual learning}, volume = {624}, year = {2025}, }
• (2025). MHGTN: Multi-interaction heterogeneous graph transformer network for knowledge graph completion. NEUCOM, 624, 129368. (www)

  @article{ , abstract = {Knowledge graph embedding has been a popular research topic for knowledge graph completion (KGC). However, due to the complexity and heterogeneity inherent in a graph, including the structure of the graph itself and the diverse roles played by entities and relations within it, the accuracy of embeddings can be compromised. In this paper, to mitigate the performance impact of graph heterogeneity in knowledge graph completion, we propose a Multi-Interaction Heterogeneous Graph Transformer Network (MHGTN) with a two-stage-four-view interaction strategy. First, the proposed model jointly considers the diversity of entity and relation types during semantic interaction to represent the heterogeneity of knowledge graphs more effectively. Moreover, to improve the accuracy of the semantic representation of entities and relations, a new training method based on the contrastive learning mechanism is introduced for network training. Furthermore, we build a new dataset (named DB8k-28) containing information on entity and relation types to verify the performance of the proposed model. Experiments are conducted on three benchmark datasets with types and compared with sixteen different baselines. The results show that our method has better improvement compared with state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Fan Lili and Jiabin Yuan and Zhiguang Fan}, doi = {10.1016/j.neucom.2025.129368}, journal = {Neurocomputing}, month = {4}, pages = {129368}, shortjournal = {Neurocomputing}, title = {MHGTN: Multi-interaction heterogeneous graph transformer network for knowledge graph completion}, volume = {624}, year = {2025}, }
• (2025). Emotion categorization from facial expressions: A review of datasets, methods, and research directions. NEUCOM, 624, 129367. (www)

  @article{ , abstract = {Emotion classification plays a crucial role in the domain of human–computer interaction, as it holds substantial significance in effective communication. However, the task of emotion classification presents notable challenges primarily due to the subjective and multifaceted nature of emotions, encompassing varying cultural and individual interpretations. Numerous methodologies have been employed to address the complexities associated with facial emotion classification. Notably, contemporary techniques such as those that consider emotion-relevant features (e.g. mouth, eyes) have demonstrated a certain degree of success in comparison to alternative approaches (e.g. the widely employed end-to-end deep networks), which consider all pixel information in an image with equal importance. Nonetheless, a comprehensive set of guidelines evaluating not only the strengths, but also the weaknesses of deep networks remains absent. These approaches are vulnerable to even minor alterations in images, regardless of their relevance to emotion classification, e.g. a change in the foreground color of the image can reduce the accuracy of these methods by a significant amount. This results in a fragmented field, ultimately missing opportunities to enhance performance for emerging real-world applications. Therefore to address this issue, this paper offers a comprehensive review of state-of-the-art datasets and research, providing insights into different techniques and their unique contributions to emotion categorization. We also thoroughly analyze current challenges and issues, laying the groundwork for promising future research directions. Our review reveals the significance of considering contemporary methods, particularly in contrast to widely used deep networks, and their potential to enhance performance and application success.}, archive = {J_NEUCOM}, author = {Harisu Abdullahi Shehu and Will N. Browne and Hedwig Eisenbarth}, doi = {10.1016/j.neucom.2025.129367}, journal = {Neurocomputing}, month = {4}, pages = {129367}, shortjournal = {Neurocomputing}, title = {Emotion categorization from facial expressions: A review of datasets, methods, and research directions}, volume = {624}, year = {2025}, }
• (2025). UP-cycle-SENet: Unpaired phase-aware speech enhancement using deep complex cycle adversarial networks. NEUCOM, 624, 129365. (www)

  @article{ , abstract = {Speech enhancement (SE), which reconstructs intelligible speech by removing noise or interference from noisy speech, plays an important role in many speech applications. Due to the successful introduction of deep learning in SE, a significant performance improvement was recorded compared to the traditional methods. Most spectrogram-based deep SE networks have two main issues. First, many existing methods only focus on estimating the magnitude of the spectrogram while reusing the phase information. Reusing the phase part of a noisy spectrogram results in clear performance limitations, which can become more pronounced when the distortion caused by noise is severe. Second, most deep SE models adopt supervised learning, which requires a large number of paired datasets. Constructing a large dataset that includes clean speech is highly impractical due to the significant effort and cost involved. To address this issue, we propose UP-Cycle-SENet, an end-to-end complex SE network capable of estimating both the magnitude and phase parts of the speech spectrogram under unpaired dataset conditions. The proposed network leverages complex convolutional neural networks and extended modules to efficiently extract features in the complex domain without losing information. Additionally, the introduction of a CNN-based discriminator with non-autoregressive properties makes it suitable for fast training and inference. To effectively validate the benefits of the proposed network, comparative experiments were conducted using public datasets that mix Voice Bank and DEMAND. The experimental results demonstrated that the proposed framework outperforms previous methods in both parallel and non-parallel strategies.}, archive = {J_NEUCOM}, author = {Cheol-Hoon Park and Hyun-Duck Choi}, doi = {10.1016/j.neucom.2025.129365}, journal = {Neurocomputing}, month = {4}, pages = {129365}, shortjournal = {Neurocomputing}, title = {UP-cycle-SENet: Unpaired phase-aware speech enhancement using deep complex cycle adversarial networks}, volume = {624}, year = {2025}, }
• (2025). IMCGNN: Information maximization based continual graph neural networks for inductive node classification. NEUCOM, 624, 129362. (www)

  @article{ , abstract = {Continual graph learning is an emerging topic that enables models to incrementally acquire new knowledge while retaining prior experiences. It efficiently adapts to model evolving dynamic graphs, avoiding the computational burden of training from scratch. The key distinction of CGL from conventional continual learning is the interdependence of samples in graph-structured data versus the independence in conventional learning. Consequently, continual graph learning techniques should emphasize consolidating and leveraging the topological information in graph-structured data. Current methods inadequately address this need. Some approaches ignore topological information, resulting in significant information loss. Others attempt to preserve all learned information, leading to overly conservative models. Moreover, most of these methods employ graph neural networks (GNNs) as the base model, yet they fail to fully utilize the topological information learned by GNNs. Additionally, the majority of existing works focus on transductive setting, with inductive continual graph learning problems being scarcely explored. Our proposed Information Maximization based Continual Graph Neural Network (IMCGNN) focuses on inductive task-incremental node classification problems. This proposed work involves a replay module and a regularization module. The former extracts representative subgraphs from previous data, training them jointly with new data to retain historical experiences, whereas the latter preserves topological information and loss-related information with encoded knowledge by imposing elastic penalties on network parameters. Unlike heuristic node selection, our approach utilizes the information theory to guide node selection in forming a subgraph, aiming to preserve information better. Comparative experiments with nine baselines using two graph learning models on five benchmark datasets demonstrate the effectiveness and efficiency of our method.}, archive = {J_NEUCOM}, author = {QiAo Yuan and Sheng-Uei Guan and Tianlun Luo and Ka Lok Man and Eng Gee Lim}, doi = {10.1016/j.neucom.2025.129362}, journal = {Neurocomputing}, month = {4}, pages = {129362}, shortjournal = {Neurocomputing}, title = {IMCGNN: Information maximization based continual graph neural networks for inductive node classification}, volume = {624}, year = {2025}, }
• (2025). Wilcoxon-type multivariate cluster elastic net. NEUCOM, 624, 129358. (www)

  @article{ , abstract = {We propose a method for high dimensional multivariate regression that is robust to heavy-tailed distributions or outliers, while preserving estimation accuracy in normal random error distributions. We extend the Wilcoxon-type regression to a multivariate regression model as a tuning-free approach to robustness. Furthermore, the proposed method regularizes the L 1 and L 2 terms of the clustering based on k -means, which is extended from the multivariate cluster elastic net. The estimation of the regression coefficient and variable selection are produced simultaneously. Moreover, considering the relationship among the correlation of response variables through the clustering is expected to improve the estimation performance. The numerical simulation demonstrates that our proposed method overperforms the multivariate cluster method and other multiple regression methods in the case of heavy-tailed error distribution and outliers. The proposed method also indicates stability in normal error distribution. Finally, we confirm the efficacy of our proposed method using gene data.}, archive = {J_NEUCOM}, author = {Mayu Hiraishi and Kensuke Tanioka and Hiroshi Yadohisa}, doi = {10.1016/j.neucom.2025.129358}, journal = {Neurocomputing}, month = {4}, pages = {129358}, shortjournal = {Neurocomputing}, title = {Wilcoxon-type multivariate cluster elastic net}, volume = {624}, year = {2025}, }
• (2025). CCAF-net: Cascade complementarity-aware fusion network for traffic accident prediction in dashcam videos. NEUCOM, 624, 129285. (www)

  @article{ , abstract = {In the field of traffic safety, predicting traffic accidents is crucial for ensuring traffic safety. The widespread use of dashboard cameras makes the prediction of traffic accidents based on in-vehicle video particularly significant for research purposes. Existing vehicle-mounted video traffic accident prediction models have not fully and effectively utilized implicit information within traffic scenes, leading to a suboptimal understanding of traffic scenarios and consequently diminishing the timeliness and accuracy of accident predictions. To address this limitation, this paper presents a method leverages multimodal information from in-car videos, capturing both global and local features derived from RGB imagery, complemented by analogous features extracted from depth maps. Inspired by the ability of drivers to integrate various information in the scene comprehensively and complementarily for accident prediction, we innovatively propose a Cascade Complementarity-Aware Fusion Network (CCAF-Net) based on Collaborative Attention mechanisms to achieve effective integration of multimodal information. Consequently, the model can perceive more comprehensive and accurate representations of traffic scene features, thereby enhancing the ability to predict traffic accidents. Notably, our depth information is directly generated from RGB images, thereby eliminating the need for additional sensors. Extensive experiments conducted on two real datasets, DAD and CCD, demonstrate that CCAF-Net can derive multimodal information from a single source and that its performance surpasses existing state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Wei Liu and Yafei Li and Tao Zhang and Yixiang Gao and Longsheng Wei and Jun Chen}, doi = {10.1016/j.neucom.2024.129285}, journal = {Neurocomputing}, month = {4}, pages = {129285}, shortjournal = {Neurocomputing}, title = {CCAF-net: Cascade complementarity-aware fusion network for traffic accident prediction in dashcam videos}, volume = {624}, year = {2025}, }
• (2025). Low-rank high-order tensor recovery via joint transformed tensor nuclear norm and total variation regularization. NEUCOM, 624, 129266. (www)

  @article{ , abstract = {The problem of low-rank tensor recovery has garnered significant attention, particularly in applications involving visual data, within the tensor singular value decomposition (t-SVD) framework. However, most existing approaches are limited to third-order tensors, restricting their applicability to higher-order data such as fourth-order color videos or fifth-order light field images. In this work, we address the challenge of high-order tensor recovery by leveraging the recently proposed high-order transformed t-SVD framework. Specifically, we utilize the high-order transformed tensor nuclear norm to model the global low-rank structure of these visual tensor datasets, while incorporating high-order total variation (TV) regularization to preserve local piecewise smoothness. To solve the resulting models for both tensor completion and robust principal component analysis (RPCA), we design efficient algorithms based on the alternating direction method of multipliers (ADMM) framework. We also rigorously establish the global convergence of these algorithms. Extensive experiments on a variety of visual datasets demonstrate the superiority of our proposed methods when compared to numerous state-of-the-art techniques.}, archive = {J_NEUCOM}, author = {Xiaohu Luo and Weijun Ma and Wendong Wang and Yuanshi Zheng and Jianjun Wang}, doi = {10.1016/j.neucom.2024.129266}, journal = {Neurocomputing}, month = {4}, pages = {129266}, shortjournal = {Neurocomputing}, title = {Low-rank high-order tensor recovery via joint transformed tensor nuclear norm and total variation regularization}, volume = {624}, year = {2025}, }
• (2025). EHAN: An explicitly high-order attention network for accurate camouflaged object detection. NEUCOM, 624, 129249. (www)

  @article{ , abstract = {Wild animals often change their appearance such as color and textures to seamlessly blend into the surrounding environments to avoid predators and enemies, which poses extreme challenges to the task of Camouflaged Object Detection (COD). Recently, most COD methods have struggled to improve the detection performance using attention mechanisms to guide deep networks to focus on the subtle differences between camouflaged objects and backgrounds. However, these approaches tend to suffer from the lower orders in attention, which have limited abilities to model the subtle differences, as observed by applying the Taylor expansion to attention mappings. To solve this problem, we propose Explicitly High-order Attention (EHA) based on the tensor decomposition theory to effectively detect hidden objects, especially at edges. Furthermore, we use EHA as a basic block to design a new COD network EHAN, which consists of four modules: Cross-layer High-order Fusion (CHF), High-order Semantic Attention (HSA), High-order Edge Attention (HEA), and High-order Reverse Attention (HRA). Specifically, CHF progressively improves the features of a backbone by merging three of them at adjacent levels at a time. HSA roughly locates the disguised objects using the features of CHF at the three highest levels, whereas HEA integrates the features of CHF at the two lowest levels to predict edges accurately. HRA aims to discover distractions and incorporate the predicted edges for edge refinement. Extensive experiments on four public datasets demonstrate the superiority of our EHAN over state-of-the-art methods in quality and quantity.}, archive = {J_NEUCOM}, author = {Qingbo Wu and Guanxing Wu and Shengyong Chen}, doi = {10.1016/j.neucom.2024.129249}, journal = {Neurocomputing}, month = {4}, pages = {129249}, shortjournal = {Neurocomputing}, title = {EHAN: An explicitly high-order attention network for accurate camouflaged object detection}, volume = {624}, year = {2025}, }
• (2025). Binding double-head predictor with key context transformer for high performance tracking. NEUCOM, 623, 129444. (www)

  @article{ , abstract = {Single object tracking (SOT) is now facing a series of challenges such as frequent object occlusion, multifarious background clutter, changeful tracking angles, fluctuant lighting conditions and so on. In this paper, by binding a double-head predictor embedded with spatiotemporal module to a key context Transformer, a high-performance tracker is proposed for single object tracking. Firstly, a novel feature integration module (FIM) with a large receptive field is designed for capturing richer feature information before the feature fusion. More importantly, a double-head predictor embedded with a modified spatiotemporal module is used to improve the ability to extract more detailed spatiotemporal information. To further enhance the feature fusion, an improved Transformer with more key context information is constructed. Finally, a key context transformer tracking (KCTrack) model based on FIM, double-head predictor and improved Transformer is realized not only for extracting more feature information but also for classifying the information more deeply. The experimental results on TrackingNet, LaSOT, OTB2015, and GOT10k datasets both show the effectiveness of this improved model proposed in this paper. In particular, the success of 70.9 % and the precision of 92.7 % on the challenging OTB2015 dataset indicate that the proposed model exhibits high performance, which can be on a par with those state-of-the-art (SOTA) methods in the field of SOT. Undoubtedly, the KCTrack model would be very helpful for many potential applications based on computer vision in the future, especially for tracking single object}, archive = {J_NEUCOM}, author = {Bingbing Wang and Haoyang Yu and Wenqian Wan and Yongfan Duan and Hongtao Gong and Daoxun Jin and Xiu Zhu and Yi Song and Hui Wang and Gang Liu}, doi = {10.1016/j.neucom.2025.129444}, journal = {Neurocomputing}, month = {3}, pages = {129444}, shortjournal = {Neurocomputing}, title = {Binding double-head predictor with key context transformer for high performance tracking}, volume = {623}, year = {2025}, }
• (2025). MaxMin-l2-SVC-NCH: A novel approach for support vector classifier training and parameter selection. NEUCOM, 623, 129438. (www)

  @article{ , abstract = {The selection of Gaussian kernel parameters plays an important role in the applications of support vector classification (SVC). A commonly used method is the k-fold cross validation with grid search (CV), which is extremely time-consuming because it needs to train a large number of SVC models. In this paper, a new approach is proposed to train SVC and optimize the selection of Gaussian kernel parameters. We first formulate the training and the parameter selection of SVC as a minimax optimization problem named as MaxMin-L2-SVC-NCH, in which the minimization problem is an optimization problem of finding the closest points between two normal convex hulls (L2-SVC-NCH) while the maximization problem is an optimization problem of finding the optimal Gaussian kernel parameters. A lower time complexity can be expected in MaxMin-L2-SVC-NCH because CV is not needed. We then propose a projected gradient algorithm (PGA) for the training of L2-SVC-NCH. It is revealed that the famous sequential minimal optimization (SMO) algorithm is a special case of the PGA. Thus, the PGA can provide more flexibility than the SMO. Furthermore, the solution of the maximization problem is done by a gradient ascent algorithm with dynamic learning rate. The comparative experiments between MaxMin-L2-SVC-NCH and the previous best approaches on public datasets show that MaxMin-L2-SVC-NCH greatly reduces the number of models to be trained while maintaining competitive test accuracy. These findings indicate that MaxMin-L2-SVC-NCH is a better choice for SVC tasks.}, archive = {J_NEUCOM}, author = {Linkai Luo and Qiaoling Yang and Hong Peng and Yiding Wang and Ziyang Chen}, doi = {10.1016/j.neucom.2025.129438}, journal = {Neurocomputing}, month = {3}, pages = {129438}, shortjournal = {Neurocomputing}, title = {MaxMin-l2-SVC-NCH: A novel approach for support vector classifier training and parameter selection}, volume = {623}, year = {2025}, }
• (2025). DiffSteISR: Harnessing diffusion prior for superior real-world stereo image super-resolution. NEUCOM, 623, 129437. (www)

  @article{ , abstract = {Although diffusion prior-based single-image super-resolution has demonstrated remarkable reconstruction capabilities, its potential in the domain of stereo image super-resolution remains underexplored. One significant challenge lies in the inherent stochasticity of diffusion models, which makes it difficult to ensure that the generated left and right images exhibit high semantic and texture consistency. This poses a considerable obstacle to advancing research in this field. Therefore, We introduce DiffSteISR, a pioneering framework for reconstructing real-world stereo images. DiffSteISR utilizes the powerful prior knowledge embedded in pre-trained text-to-image model to efficiently recover the lost texture details in low-resolution stereo images. Specifically, DiffSteISR implements a time-aware stereo cross attention with temperature adapter (TASCATA) to guide the diffusion process, ensuring that the generated left and right views exhibit high texture consistency thereby reducing disparity error between the super-resolved images and the ground truth (GT) images. Additionally, a stereo omni attention control network (SOA ControlNet) is proposed to enhance the consistency of super-resolved images with GT images in the pixel, perceptual, and distribution space. Finally, DiffSteISR incorporates a stereo semantic extractor (SSE) to capture unique viewpoint soft semantic information and shared hard tag semantic information, thereby effectively improving the semantic accuracy and consistency of the generated left and right images. Extensive experimental results demonstrate that DiffSteISR accurately reconstructs natural and precise textures from low-resolution stereo images while maintaining a high consistency of semantic and texture between the left and right views.}, archive = {J_NEUCOM}, author = {Yuanbo Zhou and Xinlin Zhang and Wei Deng and Tao Wang and Tao Tan and Qinquan Gao and Tong Tong}, doi = {10.1016/j.neucom.2025.129437}, journal = {Neurocomputing}, month = {3}, pages = {129437}, shortjournal = {Neurocomputing}, title = {DiffSteISR: Harnessing diffusion prior for superior real-world stereo image super-resolution}, volume = {623}, year = {2025}, }
• (2025). Space-constrained random sparse adversarial attack. NEUCOM, 623, 129436. (www)

  @article{ , abstract = {Adversarial attacks aim to deceive deep neural networks (DNNs) by introducing carefully designed perturbations. Traditional black-box methods often produce large-scale perturbations, reducing their practicality in real-world scenarios. As a result, sparse adversarial attacks emerged to improve invisibility and effectiveness by minimizing the number of disturbed pixels, but a major challenge lies in balancing sparsity with attack success. To address this challenge, we propose a novel Space-Constrained Random Sparse Adversarial Attack (SRSA), which focuses on space most influential to the model’s decision-making process. SRSA employs a weighted sampling strategy to dynamically update perturbation scores for each space, prioritizing areas with higher impact. A heuristic search algorithm is then applied to precisely target pixels within the selected space, achieving both sparsity and effectiveness. Experimental results demonstrate that SRSA outperforms state-of-the-art black-box sparse adversarial attacks, such as Sparse-RS, SA-MOO, SAPF, and Homotopy Attack, on DNNs trained on ImageNet. Specifically, SRSA reduces the perturbation range and magnitude by approximately 20% compared to baseline methods. Moreover, compared to the optimal query-based black-box attack method, Square Attack, SRSA achieves a higher attack success rate while modifying 30% fewer pixels. These results demonstrate that SRSA enhances attack efficiency and effectiveness by incorporating local information into the perturbation search process.}, archive = {J_NEUCOM}, author = {Yueyuan Qin and Gang Hou and Tingting Yu and Bing Gao and Weiqiang Kong and Xiaoshan Liu}, doi = {10.1016/j.neucom.2025.129436}, journal = {Neurocomputing}, month = {3}, pages = {129436}, shortjournal = {Neurocomputing}, title = {Space-constrained random sparse adversarial attack}, volume = {623}, year = {2025}, }
• (2025). LGAST: Towards high-quality arbitrary style transfer with local–global style learning. NEUCOM, 623, 129434. (www)

  @article{ , abstract = {Arbitrary style transfer has become a research hotspot in academia, industry, and the arts. While current methods have made great development, but there are still three challenges: (1) They do not capture the local detailed textures of the style image; (2) They introduce undesired semantic information from the style image; (3) They fail to balance local–global style patterns. To this end, we introduce Local–global Arbitrary Style Transfer (LGAST), which is designed to learn style patterns from local perspective and style distributions from global perspective. Specifically, a Local Channel-Wise Attention (LCWA) is introduced to learn local detailed textures of the style image. Then, a Global Style Fusion Transformer (GSFT) is employed to learn the global style information further and to remove undesired information from the style image. Finally, a novel patch-wise Gram loss is proposed to further balance local style patterns and global style distribution. Our method can not only be used for arbitrary style transfer of images, but also do well in video style transfer with slight improvement. Extensive experiments demonstrated our proposed LGAST outperforms with SOTA style transfer methods.}, archive = {J_NEUCOM}, author = {Zhanjie Zhang and Yuxiang Li and Ruichen Xia and Mengyuan Yang and Yun Wang and Lei Zhao and Wei Xing}, doi = {10.1016/j.neucom.2025.129434}, journal = {Neurocomputing}, month = {3}, pages = {129434}, shortjournal = {Neurocomputing}, title = {LGAST: Towards high-quality arbitrary style transfer with local–global style learning}, volume = {623}, year = {2025}, }
• (2025). A robust two-stage framework for human skeleton action recognition with GAIN and masked autoencoder. NEUCOM, 623, 129433. (www)

  @article{ , abstract = {Human skeleton-based action recognition is crucial for human-robot interaction but faces challenges from target occlusion and similarity between actions. To address these issues, we propose a novel two-stage network. The first stage employs spatiotemporal generative adversarial interpolation network (STGAIN) to interpolate the missing skeleton key points. The second stage introduces a frame-joint motion masked autoencoder (FJMAE) structure, designing the frame-joint motion masking strategy and cross-attention decoder, enabling rapid differentiation between similar actions. The effectiveness of the two-stage network is validated by constructing six common human-robot interaction datasets. Experimental results on six self-constructed human-robot interaction datasets demonstrate the effectiveness of our two-stage network method, outperforming most state-of-the-art models in Top-1 accuracy. In addition, we evaluate the FJMAE on the NTU-RGB-D 60 dataset and NTU-RGB-D 120 dataset, exhibiting strong robustness. Our code is available at https://github.com/GD1201/A-two-stage-network-for-action-recognition .}, archive = {J_NEUCOM}, author = {Shiqing Wu and Guodong Lu and Zongwang Han and Long Chen}, doi = {10.1016/j.neucom.2025.129433}, journal = {Neurocomputing}, month = {3}, pages = {129433}, shortjournal = {Neurocomputing}, title = {A robust two-stage framework for human skeleton action recognition with GAIN and masked autoencoder}, volume = {623}, year = {2025}, }
• (2025). Enhancing rotated object detection via anisotropic gaussian bounding box and bhattacharyya distance. NEUCOM, 623, 129432. (www)

  @article{ , abstract = {Detecting rotated objects accurately and efficiently is a significant challenge in computer vision, particularly in applications such as aerial imagery, remote sensing, and autonomous driving. Although traditional object detection frameworks are effective for axis-aligned objects, they often underperform in scenarios involving rotated objects due to their limitations in capturing orientation variations. This paper introduces an improved loss function aimed at enhancing detection accuracy and robustness by leveraging the Gaussian bounding box representation and Bhattacharyya distance. In addition, we advocate for the use of an anisotropic Gaussian representation to address the issues associated with isotropic variance in square-like objects. Our proposed method addresses these challenges by incorporating a rotation-invariant loss function that effectively captures the geometric properties of rotated objects. We integrate this proposed loss function into state-of-the-art deep learning-based rotated object detection detectors, and extensive experiments demonstrated significant improvements in mean Average Precision metrics compared to existing methods. The results highlight the potential of our approach to establish a new benchmark in rotated object detection, with implications for a wide range of applications that require precise and reliable object localization irrespective of orientation.}, archive = {J_NEUCOM}, author = {Chien Thai and Mai Xuan Trang and Huong Ninh and Hoang Hiep Ly and Anh Son Le}, doi = {10.1016/j.neucom.2025.129432}, journal = {Neurocomputing}, month = {3}, pages = {129432}, shortjournal = {Neurocomputing}, title = {Enhancing rotated object detection via anisotropic gaussian bounding box and bhattacharyya distance}, volume = {623}, year = {2025}, }
• (2025). Revisiting intelligent audit from a data science perspective. NEUCOM, 623, 129431. (www)

  @article{ , abstract = {The contemporary auditing landscape faces challenges stemming from the proliferation of varied data and the intricacy of audit tasks. In response, auditors are progressively embracing data intelligence (DI) technologies to refine procedures and elevate audit quality, leading to the emergence of “intelligent audit”. Nevertheless, the prevailing focus of research on financial statement audits tends to overshadow the importance of broader contexts and the inherent complexity of audit tasks, particularly in government auditing. The research to date has primarily examined the impact of advanced technologies on auditing from technical and application perspectives, neglecting the essential steps and critical technological processes necessary for addressing audit challenges. Additionally, much of this discourse, including conclusions and proposed frameworks, lacks empirical and practical validation. This gap hinders engineers’ comprehension of intelligent auditing and impedes the development and implementation of effective intelligent audit systems. This paper provides a comprehensive review of intelligent auditing from the perspective of data science. Initially, we aim to formulate a formal definition of “intelligent audit”, a concept that currently lacks consensus in academia, and introduce three real-world case studies that highlight the diversity of data and the complexity of audit tasks. Subsequently, through bibliometric analysis, we provide an overview of the annual advancements and literature distribution pertinent to intelligent auditing. Furthermore, we investigate the paradigm shift and illustrative frameworks to shed light on the impacts of DI technologies. Additionally, we provide an in-depth summary of key technological applications in intelligent auditing, encompassing data modeling, anomaly detection, and advanced artificial intelligence techniques. The primary objective of this paper is to assist accountants and auditors in acquiring a profound comprehension of the current research landscape concerning the integration of novel technologies into auditing practices. Moreover, this survey seeks to attract the interest of computer scientists and artificial intelligence researchers in the auditing field, underscoring its potential as a promising area for application.}, archive = {J_NEUCOM}, author = {Hongru Lu and Zhiang Wu}, doi = {10.1016/j.neucom.2025.129431}, journal = {Neurocomputing}, month = {3}, pages = {129431}, shortjournal = {Neurocomputing}, title = {Revisiting intelligent audit from a data science perspective}, volume = {623}, year = {2025}, }
• (2025). Coupling importance sampling neural network for imbalanced data classification with multi-level learning bias. NEUCOM, 623, 129427. (www)

  @article{ , abstract = {Imbalanced data classification is a classic challenge in neural network learning. Current rebalancing methods for neural networks mainly rely on resampling and reweighting to alleviate the learning bias caused by imbalanced data learning. It is difficult to properly balance the learning and resampling and rarely dives into the problem of multi-level learning bias. In this paper, we propose a coupling importance sampling to couple the resampling and neural networks learning, while handling both class-level and cluster-level learning biases that occur between and within classes, respectively. Specifically, in the coupling of resampling and learning, as for the class-level learning bias, we extend the resampling from sample to cluster. A composite importance factor is developed to sample important clusters to balance samples between classes. Here, a distribution preservation strategy is additionally maintained to reduce the loss of important samples from resampled clusters. Regarding cluster-level learning bias, a learning regulatory factor is designed to highlight the importance of sampled clusters and avoid the recurrence of imbalance within classes. The proposed method is validated on 34 imbalanced datasets with imbalance ratios ranging from 16.90 to 100.14. The tested results show promising classification performance and prove the advantages of considering the multi-level learning bias in imbalanced data classification.}, archive = {J_NEUCOM}, author = {Zhan ao Huang and Wei Xiao and Zhipeng Yang and Xiaojie Li and Xi Wu}, doi = {10.1016/j.neucom.2025.129427}, journal = {Neurocomputing}, month = {3}, pages = {129427}, shortjournal = {Neurocomputing}, title = {Coupling importance sampling neural network for imbalanced data classification with multi-level learning bias}, volume = {623}, year = {2025}, }
• (2025). FIPNet: Self-supervised low-light image enhancement combining feature and illumination priors. NEUCOM, 623, 129426. (www)

  @article{ , abstract = {Most existing low-light image enhancement techniques face significant challenges in terms of their limited adaptability to varying illumination conditions and the introduction of noise during the enhancement process. Additionally, these methods typically rely on paired data, which may not always be available. To address these issues, we propose a novel self-supervised low-light image enhancement network based on the Retinex theory with combined priors. Specifically, we integrate illumination priors with histogram equalization-based feature priors to provide a richer set of texture and color features for model training. Furthermore, the proposed wave noise suppression module(WNSM) employs a frequency-domain basis to preserve local detail features while effectively suppressing noise amplification during the image decomposition process, eliminating the need for additional post-processing steps. Simultaneously, the illumination adaptive module(IAM) accounts for reflection characteristics under various low-light conditions, ensuring that the decomposed reflectance map accurately represents the original image with minimal illumination artifacts. Extensive experiments and ablation studies demonstrate the effectiveness of our method in handling different exposure levels and significantly reducing noise. Compared to existing low-light image enhancement techniques, our approach achieves superior performance on image quality evaluation metrics. The code repository is located at https://github.com/Mac-jie/FIPNet .}, archive = {J_NEUCOM}, author = {Jie Cao and Qijie Zou and Xin Ning and Zhen Wang and Gang Wang}, doi = {10.1016/j.neucom.2025.129426}, journal = {Neurocomputing}, month = {3}, pages = {129426}, shortjournal = {Neurocomputing}, title = {FIPNet: Self-supervised low-light image enhancement combining feature and illumination priors}, volume = {623}, year = {2025}, }
• (2025). Dual-domain based backdoor attack against federated learning. NEUCOM, 623, 129424. (www)

  @article{ , abstract = {The distributed training feature and data heterogeneity in federated learning (FL) render it susceptible to various threats, in which the backdoor attack stands out as the most destructive one. By injecting malicious functionality into the global model through poisoned updates, backdoor attacks can generate attacker-desired inference results on the trigger-embedded inputs while behaving normally on other data instances. The current backdoor triggers are of significant visual features that can be easily identified by humans or computers. Meanwhile, the common model update clipping mechanism is too simple and straightforward to be recognized by various defense methods with ease. Aiming at the above shortcomings, we proposed a dual-domain based backdoor attack (DDBA) against FL in this paper. On the one hand, DDBA generates an imperceptible dual-domain trigger for any image by superimposing in its low-frequency region of the amplitude spectrum and then applying a slight spatial distortion subsequently. On the other hand, DDBA truncates the model update dynamically based on a newly designed adaptive clipping mechanism to enhance its stealthiness. Finally, we carried out extensive experiments to evaluate the attack performance and stealth performance of DDBA on four publicly available datasets. The experiment results show that DDBA has excellent attack performance in both single-shot and multiple-shot attack scenarios as well as robust stealth performance under the existing defense methods against backdoor attacks.}, archive = {J_NEUCOM}, author = {Guorui Li and Runxing Chang and Ying Wang and Cong Wang}, doi = {10.1016/j.neucom.2025.129424}, journal = {Neurocomputing}, month = {3}, pages = {129424}, shortjournal = {Neurocomputing}, title = {Dual-domain based backdoor attack against federated learning}, volume = {623}, year = {2025}, }
• (2025). Towards improved evaluation of generative neural networks: The fréchet coefficient. NEUCOM, 623, 129422. (www)

  @article{ , abstract = {Generative adversarial networks (GANs) have shown remarkable capabilities for synthesizing realistic images and movies. However, evaluating the performance of GANs remains a challenging task. Specifically, existing metrics dedicated to this task, such as the Fréchet Inception Distance (FID), lack interpretability since they provide scores that are not bound to any range. This paper introduces the Fréchet Coefficient (FC), the novel metric that addresses the challenge by providing a clear performance score between 0 and 1, thus making interpreting and comparing results easier. Also, FC can use any convolutional neural network as a feature extractor, offering flexibility and potential for customization. We evaluate the performance of FC and benchmark it against FID on five diverse image datasets within the image-to-image translation framework. These datasets include medical, natural scene, and face images where GANs are tasked with synthesizing images from semantic segmentation maps. We also test FC’s performance under various image distortions. Experimental results demonstrate that FC is a reliable metric for evaluating GAN performance. It consistently outperforms FID regarding interpretability, making it a valuable tool for researchers and practitioners working with GANs.}, archive = {J_NEUCOM}, author = {Adrian Kucharski and Anna Fabijańska}, doi = {10.1016/j.neucom.2025.129422}, journal = {Neurocomputing}, month = {3}, pages = {129422}, shortjournal = {Neurocomputing}, title = {Towards improved evaluation of generative neural networks: The fréchet coefficient}, volume = {623}, year = {2025}, }
• (2025). Tremor suppression for master-slave teleoperated robot based on machine learning: A review. NEUCOM, 623, 129421. (www)

  @article{ , abstract = {Effective tremor suppression is an indispensable prerequisite for achieving precise control of master-slave teleoperated robots. Tremor suppression strategies in master-slave robots have been under exploration by researchers since the 1980s. In recent years, particularly since 2010, machine learning (ML) based tremor suppression methods have emerged as a prominent research area due to their high robustness and predictive accuracy. This paper presents a systematic review of the research progress in tremor suppression of master-slave robots using ML approaches in recent years. The study begins by examining the historical development of tremor suppression and then provides a detailed introduction to current ML-based tremor suppression methods, analyzing their strengths and weaknesses. Furthermore, the paper delves into aspects such as data collection and preprocessing, model training and testing processes, and the challenges encountered. Moreover, the discussion extends to future research directions in this field, aiming to provide valuable insights for advancing the development and application of tremor suppression technology in master-slave robots.}, archive = {J_NEUCOM}, author = {Ke Liang and Yue Su and Gang Du and Chun Ma and Mantian Li and Mingzhang Pan}, doi = {10.1016/j.neucom.2025.129421}, journal = {Neurocomputing}, month = {3}, pages = {129421}, shortjournal = {Neurocomputing}, title = {Tremor suppression for master-slave teleoperated robot based on machine learning: A review}, volume = {623}, year = {2025}, }
• (2025). SA3D-L: A lightweight model for 3D object segmentation using neural radiance fields. NEUCOM, 623, 129420. (www)

  @article{ , abstract = {The Segment Anything Model (SAM) has recently made significant progress in object segmentation within 2D images. However, the task of segmenting objects in a 3D space remains a primary hurdle in computer vision. The neural radiance field (NeRF) utilizes a multilayer perceptron (MLP) to effectively learn the continuous representation of a 3D scene. Due to its consistent 3D perspectives from various angles, SAM, originally designed for 2D segmentation, can be extended for 3D object segmentation by incorporating NeRF. However, a limitation of NeRF is that the MLP encapsulates the whole scene as a single representation, without distinguishing individual objects. This study introduces a lightweight 3D Segment Anything Model ( SA3D-L ), which separately represents each segmented object within a scene by modifying the MLP output. Experimental results on established benchmarks revealed that the 3D segmentation representations of objects can be derived from their 2D masks, allowing the independent manipulation of segmented objects and the reconstruction of a new scene. The code is available at: https://github.com/liujian0819/SA3D-L .}, archive = {J_NEUCOM}, author = {Jian Liu and Zhen Yu}, doi = {10.1016/j.neucom.2025.129420}, journal = {Neurocomputing}, month = {3}, pages = {129420}, shortjournal = {Neurocomputing}, title = {SA3D-L: A lightweight model for 3D object segmentation using neural radiance fields}, volume = {623}, year = {2025}, }
• (2025). Enhancing visual tracking with multi-scale cross-correlation and progressive attention. NEUCOM, 623, 129417. (www)

  @article{ , abstract = {Recent advancements in attention-based Siamese trackers have led to significant improvements in performance. However, tracking efficiency remains a crucial metric that cannot be overlooked in practical applications. This paper analyses the characteristics of current attention mechanisms and Siamese frameworks. Based on these analyses, we propose a more efficient attention-based tracking algorithm, named MCAT. First, the construction of global feature dependencies, which is crucial for the functionality of attention mechanisms, always involves extensive computation of numerous insignificant dependencies. To address this issue, we propose a progressive attention mechanism LGPA. This method can effectively reduce unnecessary computations by progressively building feature dependencies from local to global levels under the guidance of spatial and channel distribution information. Additionally, despite keypoint-based tracking strategy has demonstrated significant potential in the research of Siamese frameworks, there is still room for improvement in the efficiency of multi-scale information fusion and utilization, which is crucial for improving the robustness of tracking models under complex scenarios. For this issue, we propose a multi-scale pixelwise cross-correlation method and a novel corner head to better mine and integrate multi-scale information. By embedding the proposed methods, we achieve substantial accuracy improvements with minimal computational overhead. With different backbones, we propose two end-to-end Siamese trackers: a fully convolutional tiny version (MCAT) and a transformer-based version (MCAT-swin). Extensive experiments across multiple benchmarks show that our trackers achieve state-of-the-art performance with high efficiency. On the long-term tracking benchmark LaSOT, MCAT-swin achieves an AUC score of 68.1% and a precision score of 72.5%, while running at 140FPS. The running speed of MCAT even exceeds 220FPS.}, archive = {J_NEUCOM}, author = {Qinghui Wang and Peng Yang and Lei Dou}, doi = {10.1016/j.neucom.2025.129417}, journal = {Neurocomputing}, month = {3}, pages = {129417}, shortjournal = {Neurocomputing}, title = {Enhancing visual tracking with multi-scale cross-correlation and progressive attention}, volume = {623}, year = {2025}, }
• (2025). CMFF6D: Cross-modality multiscale feature fusion network for 6D pose estimation. NEUCOM, 623, 129416. (www)

  @article{ , abstract = {6D pose estimation is applied in many fields including robotic grasping tasks. This task becomes particularly challenging in poor lighting conditions or when dealing with texture-less metal objects. Single-modality approaches are insufficient for precise 6D pose estimation. Consequently, it is imperative to exploit the complementary nature of point cloud and RGB information to enhance the accuracy and robustness of 6D pose estimation. To tackle this challenge, we propose a novel framework called CMFF6D, that achieves precise 6D pose estimation by learning the complementary features from the texture information of images, the geometric information of depth images, and the frequency information between the two modalities. Specifically, we propose a new network architecture called the GL-Swin Transformer to extract the image features and use RandLA-Net to learn the point cloud features. Then, these two kinds of features are adaptively merged with their spatial domain and frequency domain information during the bidirectional fusion process to obtain richer feature representation. In order to improve the pose estimation accuracy of small objects by better capturing the differences between stacked objects, we design a Residual Multiplicative Connection (RMC) block to reduce noise interference in shallow features, thereby obtaining shallow information with more boundary details. Finally, we utilize the Feature Pyramid Aggregation Network for multi-scale feature fusion, achieving a deeper integration of the complementary data. Extensive experiments show that the proposed approach achieves competitive performances on MP6D, LM, and LM-O datasets.}, archive = {J_NEUCOM}, author = {Zongwang Han and Long Chen and Shiqing Wu}, doi = {10.1016/j.neucom.2025.129416}, journal = {Neurocomputing}, month = {3}, pages = {129416}, shortjournal = {Neurocomputing}, title = {CMFF6D: Cross-modality multiscale feature fusion network for 6D pose estimation}, volume = {623}, year = {2025}, }
• (2025). Dual representations: A novel variant of self-supervised audio spectrogram transformer with multi-layer feature fusion and pooling combinations for sound classification. NEUCOM, 623, 129415. (www)

  @article{ , abstract = {The Self-Supervised Audio Spectrogram Transformer (SSAST) has recently been verified as the state-of-the-art model in various audio and speech command classification tasks. SSAST uses self-supervised learning to reduce the need of substantial data to pre-train transformers, removing the disadvantage of its supervised learning counterpart, the Audio Spectrogram Transformer (AST) model. Owing to the fact that transformers such as SSAST use only feature representations from the last layer for downstream classification tasks, we believe that such a process will lose some important information from middle layers during training. Therefore, in this research, we propose a novel variant of the SSAST model using a dual representation generated using fusion of the outputs from multi-layers (i.e. both middle and last layers) for audio classification. Specifically, we apply all-patch-wise pooling combinations to all patches from both a middle layer and the last layer of a pre-trained patch-based self-supervised learning model. As such, it generates two individual sequences of the output patches based on a variety of m e a n , m a x , and m i n pooling combinations to make the final double-sized representation. This dual representation includes more discriminative information and better knowledge, providing the linear multi-layer perceptron head layers with more useful information for audio classification. In comparison with existing state-of-the-art studies, the proposed model using the dual representations yielded by multi-layer feature fusion and pooling combinations significantly boosts performance on all tasks. The resulting accuracy rates are 93.67%, 100%, 79.59%, 79.59%, 91.22%, and 85.90% for CREMA-D, TESS, RAVDESS, Speech Emotion Classification, Isolated Urban Events, and CornellBirdCall, respectively.}, archive = {J_NEUCOM}, author = {Hyosun Choi and Li Zhang and Chris Watkins}, doi = {10.1016/j.neucom.2025.129415}, journal = {Neurocomputing}, month = {3}, pages = {129415}, shortjournal = {Neurocomputing}, title = {Dual representations: A novel variant of self-supervised audio spectrogram transformer with multi-layer feature fusion and pooling combinations for sound classification}, volume = {623}, year = {2025}, }
• (2025). FKAN: Fractional Kolmogorov–Arnold networks with trainable jacobi basis functions. NEUCOM, 623, 129414. (www)

  @article{ , abstract = {Recent advancements in neural network design have given rise to the development of Kolmogorov–Arnold Networks (KANs), which enhance interpretability and precision of these systems. This paper presents the Fractional Kolmogorov–Arnold Network (fKAN), a novel neural network architecture that incorporates the distinctive attributes of KANs with a trainable adaptive fractional-orthogonal Jacobi function as its basis function. By leveraging the unique mathematical properties of fractional Jacobi functions, including simple derivative formulas, non-polynomial behavior, and activity for both positive and negative input values, this approach ensures efficient learning and enhanced accuracy. The proposed architecture is evaluated across a range of tasks in deep learning and physics-informed deep learning. Precision is tested on synthetic regression data, image classification, transfer learning, image denoising, and sentiment analysis. Additionally, the performance is measured on various differential equations, including ordinary, partial, and fractional delay differential equations. The results demonstrate that integrating fractional Jacobi functions into KANs significantly improves training speed and performance across diverse fields and applications.}, archive = {J_NEUCOM}, author = {Alireza Afzal Aghaei}, doi = {10.1016/j.neucom.2025.129414}, journal = {Neurocomputing}, month = {3}, pages = {129414}, shortjournal = {Neurocomputing}, title = {FKAN: Fractional Kolmogorov–Arnold networks with trainable jacobi basis functions}, volume = {623}, year = {2025}, }
• (2025). Online censoring-based learning algorithms for fully complex-valued neural networks. NEUCOM, 623, 129412. (www)

  @article{ , abstract = {In the era of big data, the exponential increase in complex-valued data generation requires overwhelming resources such as storage capacity, device-to-device communications, data processing costs, and power consumption. Hence, in this study, we raise the question of whether all training data is beneficial to the training of fully complex-valued neural networks (FCVNNs). To this end, exploiting the online censoring (OC) strategy, we introduce novel cost-effective learning algorithms for training the FCVNNs, called the OC-based fully complex nonlinear gradient descent (OC-FCNGD) for the FCVNN and the OC-based augmented FCNGD (OC-AFCNGD) for the augmented FCVNN (AFCVNN). These algorithms are derived using the Wirtinger calculus to simplify their derivations, define them in more compact forms, and thus remove Schwarz symmetry restrictions on their complex activation functions. The proposed OC-FCNGD and OC-AFCNGD algorithms considerably reduce the training times of the FCVNN and AFCVNN by censoring less informative training data, that is, by avoiding unnecessary weight updating. In this way, the OC-FCNGD and OC-AFCNGD also achieve superior testing results than their standard counterparts. Importantly, the proposed algorithms are not limited to the training of shallow architectures such as the FCVNN and AFCVNN, they are also extended and applied to the fully complex-valued deep neural network (FCVDNN) and augmented FCVDNN (AFCVDNN) in this paper, maintaining the efficiency and performance advantages arising from the OC strategy even in deep neural network models. Furthermore, we rigorously analyze the deterministic convergence of the OC-FCNGD and OC-AFCNGD, including both weak and strong convergence, by considering the unified mean value theorem. Finally, comprehensive experiments validate the theoretical and practical advantages of the proposed algorithms.}, archive = {J_NEUCOM}, author = {Engin Cemal Mengüç and Danilo P. Mandic}, doi = {10.1016/j.neucom.2025.129412}, journal = {Neurocomputing}, month = {3}, pages = {129412}, shortjournal = {Neurocomputing}, title = {Online censoring-based learning algorithms for fully complex-valued neural networks}, volume = {623}, year = {2025}, }
• (2025). Federated training of GNNs with similarity graph reasoning for text–image retrieval. NEUCOM, 623, 129411. (www)

  @article{ , abstract = {Text–image retrieval is a key task in computer vision and natural language processing, focused on identifying the most relevant text or image for a given query. Recently, graph-based approaches have been introduced to model graph-structured data and capture semantic relationships for text–image retrieval. However, these methods face challenges such as high storage and computational costs, along with concerns regarding data privacy and security. To address these issues, we propose a federated training framework for graph neural networks with similarity graph reasoning (FGNN-SGR), where similarity graph reasoning is employed within each client to construct and refine graph-based representations of text–image data, thereby enhancing semantic alignment. A federated modularized weighted aggregation mechanism on the server is introduced to improve model convergence and global performance while ensuring privacy and scalability within the federated learning framework. Experimental results demonstrate that FGNN-SGR achieves performance competitive with state-of-the-art text–image retrieval methods while maintaining data privacy across multiple benchmark datasets.}, archive = {J_NEUCOM}, author = {Xueming Yan and Chuyue Wang and Yaochu Jin}, doi = {10.1016/j.neucom.2025.129411}, journal = {Neurocomputing}, month = {3}, pages = {129411}, shortjournal = {Neurocomputing}, title = {Federated training of GNNs with similarity graph reasoning for text–image retrieval}, volume = {623}, year = {2025}, }
• (2025). Multi-domain feature analysis of MI-EEG signals using tensor train decomposition and projected gradient non-negative matrix factorization. NEUCOM, 623, 129410. (www)

  @article{ , abstract = {Tensor decomposition is an effective method for extracting multi-domain features from signals and has been widely applied in the field of EEG. Compared to traditional tensor decomposition methods, the Sparse Regularity Tensor Train (SR-TT) decomposition algorithm offers advantages such as lower algorithmic time complexity and superior feature extraction capabilities. However, the SR-TT decomposition algorithm does not incorporate the inherent high-dimensional nonlinear characteristics of EEG signals, resulting in poor interpretability and lack of flexibility in selecting the TT-rank. To address these issues, this paper improves upon the SR-TT decomposition by integrating Projected Gradient Non-negative Matrix Factorization (PGDNMF), which allows for free selection of decomposition ranks and better aligns with the physical properties of motor-imagery EEG (MI-EEG) signals. The proposed Tensor Train Projected Gradient Non-negative Matrix Factorization (TT-PGDNMF) algorithm extracts multi-domain features from MI-EEG signals, and effectively handles high-dimensional nonlinear MI-EEG signals while extracting feature representations with superior classification performance. Experimental validation on three MI-EEG datasets shows that the TT-PGDNMF algorithm achieves classification accuracies of 92.10%, 94.60%, and 92.30%, respectively. Overall, it outperforms all currently known state-of-the-art algorithms. Furthermore, this method effectively extracts spatial features of MI-EEG signals and enables analysis through brain topographic visualization. The proposed method provides a novel feature extraction approach for MI-EEG signal research that aligns with the inherent physical characteristics of EEG signals.}, archive = {J_NEUCOM}, author = {Yunyuan Gao and Wang Xie and Zhizeng Luo and Michael Houston and Yingchun Zhang}, doi = {10.1016/j.neucom.2025.129410}, journal = {Neurocomputing}, month = {3}, pages = {129410}, shortjournal = {Neurocomputing}, title = {Multi-domain feature analysis of MI-EEG signals using tensor train decomposition and projected gradient non-negative matrix factorization}, volume = {623}, year = {2025}, }
• (2025). Learning of fuzzy cognitive map models without training data. NEUCOM, 623, 129409. (www)

  @article{ , abstract = {This paper introduces a novel zero-data learning algorithm tailored for Fuzzy Cognitive Map (FCM) models utilized in control applications where we must maintain concepts’ activation values within predefined intervals. Our approach allows domain experts to specify these intervals and optionally impose weight constraints, ensuring the algorithm produces feasible models. At the core of our approach lies a mathematical formalism that approximates the smallest feasible activation space for each neural concept, which translates into lower and upper bounds for concepts’ activation values. Moreover, a parameterized quasi-nonlinear reasoning rule allows controlling whether or not the network converges to a unique fixed point. The learning goal of our algorithm narrows down to computing a weight matrix minimizing the error between the analytical bounds and the target intervals specified by domain experts. To address such a constrained minimization problem, we employ numerical methods operating with approximate gradients, which provide highly accurate solutions with short execution times. The main contribution of our learning algorithm is that it does not require any training data to compute the network structure. Therefore, by accurately approximating the specified activation intervals, our learning algorithm guarantees that the outputs produced by the FCM model will remain within these intervals regardless of the initial conditions used to start the recurrent reasoning process.}, archive = {J_NEUCOM}, author = {Gonzalo Nápoles and Isel Grau and Leonardo Concepción and Yamisleydi Salgueiro and Koen Vanhoof}, doi = {10.1016/j.neucom.2025.129409}, journal = {Neurocomputing}, month = {3}, pages = {129409}, shortjournal = {Neurocomputing}, title = {Learning of fuzzy cognitive map models without training data}, volume = {623}, year = {2025}, }
• (2025). LPFSTNet: A lightweight and parameter-free head attention-based student–teacher network for fast 3D industrial anomaly detection. NEUCOM, 623, 129408. (www)

  @article{ , abstract = {The emerging three-dimensional (3D) anomaly detection networks have achieved remarkable detection accuracy in industrial applications. However, they always occupy excessive computing resources and exhibit bad real-time performance. Herein, a lightweight and parameter-free head attention-based student–teacher network (LPFSTNet) is demonstrated for unsupervised 3D anomaly detection. Firstly, a lightweight student network mainly containing simplified convolution operations is constructed, helping to reduce the parameter numbers while increasing the distinction between the student and teacher networks, thus enhancing generalization capability of the model. Then, a novel parameter-free head attention mechanism is proposed to regulate the output of the model, contributing to a higher detection accuracy. This head attention module can be directly integrated into the model without any training and adds no additional inference time. The experimental results validate that LPFSTNet achieves 94.7% image-level area under the receiver operating characteristic curve (I-AUROC) and 38.5 frames per second (FPS) on MVTec3D dataset, surpassing other mainstream 3D anomaly detection counterparts. Importantly, the model size is decreased by 74% compared to the baseline. Moreover, the experiment implemented on our own dataset verifies that our LPFSTNet model is robust to other dataset. The code and dataset are available at https://github.com/YahuiCheng/LPFSTNet .}, archive = {J_NEUCOM}, author = {Yahui Cheng and Junchao Chen and Guojun Wen and Xianhua Tan and Xingyue Liu}, doi = {10.1016/j.neucom.2025.129408}, journal = {Neurocomputing}, month = {3}, pages = {129408}, shortjournal = {Neurocomputing}, title = {LPFSTNet: A lightweight and parameter-free head attention-based student–teacher network for fast 3D industrial anomaly detection}, volume = {623}, year = {2025}, }
• (2025). Event-triggered control for projective synchronization fractional-order switched coupled neural networks with higher-order interactions. NEUCOM, 623, 129407. (www)

  @article{ , abstract = {This paper investigates the projective synchronization problem for fractional-order switched delayed coupled neural networks (FSDCNNs) with high-order interactions. Firstly, high-order interactions are incorporated into the FSDCNN model to more accurately simulate the complex interactions in actual neural networks. Secondly, a novel Halanay inequality, suitable for complex systems with time-varying disturbances, is proposed. This inequality provides an upper bound for the system’s solution, enabling a more precise estimation of the convergence rate. Thirdly, an event-triggered controller is introduced to achieve convergence in fractional-order switched delayed systems, effectively addressing memory effect issues. Based on this, projective synchronization criteria for FSDCNNs are established. Finally, the effectiveness of the event-triggered control strategy is validated through numerical simulation.}, archive = {J_NEUCOM}, author = {Dongsheng Yang and Hu Wang and Guojian Ren and Yongguang Yu and Xiao-Li Zhang}, doi = {10.1016/j.neucom.2025.129407}, journal = {Neurocomputing}, month = {3}, pages = {129407}, shortjournal = {Neurocomputing}, title = {Event-triggered control for projective synchronization fractional-order switched coupled neural networks with higher-order interactions}, volume = {623}, year = {2025}, }
• (2025). The good, the bad, and the algorithm: The impact of generative AI on cybersecurity. NEUCOM, 623, 129406. (www)

  @article{ , abstract = {Generative Adversarial Networks (GANs) are emerging as a transformative technology in cybersecurity, presenting both opportunities and challenges in enhancing defensive and offensive strategies. This paper explores the impact that Generative Artificial Intelligence (AI) has on cybersecurity, focusing on its application in the field of network and web security. Current research reveals robust defensive approaches; however, there remains a significant gap in the application of Generative AI to develop advanced attack scenarios capable of bypassing existing defense mechanisms. Our work attempts to fill this gap and spreads awareness regarding a potential exposure of Neural Network (NN)-based Intrusion Detection Systems (IDSs) against AI-enhanced attacks. Unlike conventional approaches that focus on Input Perturbation, Data Poisoning, or Spoofing, we propose a novel offensive strategy called Attack Obfuscation. This strategy leverages Conditional GANs (CGANs) to conceal genuine attacks by injecting synthetic traffic designed to deceive NN-based IDS. The experimental investigation validates the proposed approach against three distinct datasets and different typologies of attacks, managing to successfully deceive the IDS.}, archive = {J_NEUCOM}, author = {Luigi Coppolino and Salvatore D’Antonio and Giovanni Mazzeo and Federica Uccello}, doi = {10.1016/j.neucom.2025.129406}, journal = {Neurocomputing}, month = {3}, pages = {129406}, shortjournal = {Neurocomputing}, title = {The good, the bad, and the algorithm: The impact of generative AI on cybersecurity}, volume = {623}, year = {2025}, }
• (2025). Unlocking the high dimensional’ potential: Comparative analysis of qubits and qutrits in variational quantum neural networks. NEUCOM, 623, 129404. (www)

  @article{ , abstract = {Quantum machine learning is a promising research area with great potential. In particular, Variational quantum neural networks (VQNN) have shown high performance in many applications. However, while qubits, which are 2-level quantum systems, are the standard building blocks of quantum computing, the development of qudits, i.e. d-level quantum systems, has opened up new opportunities in VQNNs thanks to many properties such as robustness to noise and more quantum information processing with fewer quantum resources. In this study, we present a comparative analysis of qubits and qutrits (3-level quantum systems) systems performance in VQNNs while also exploring the effect of encoding strategies and entanglement on classifier performance. Our findings contribute to a better understanding the benefits and limitations of using qutrits in VQNN and pave the way for future developments in this field.}, archive = {J_NEUCOM}, author = {Erdi Acar and İhsan Yılmaz}, doi = {10.1016/j.neucom.2025.129404}, journal = {Neurocomputing}, month = {3}, pages = {129404}, shortjournal = {Neurocomputing}, title = {Unlocking the high dimensional’ potential: Comparative analysis of qubits and qutrits in variational quantum neural networks}, volume = {623}, year = {2025}, }
• (2025). Deep learning based image classification for embedded devices: A systematic review. NEUCOM, 623, 129402. (www)

  @article{ , abstract = {Deep learning models are widely employed to solve complex problems in different areas, particularly for image classification, because of their high performance in pattern recognition tasks. The demand for image classification extends beyond traditional computing environments and often requires deployment of embedded and low-cost devices in real-world scenarios to meet low-latency applications and user requirements. Embedding deep learning in low-cost devices is challenging due to their constrained resources, whereas deep learning models require many resources. In the literature, there are different approaches to make this embedding viable, such as reducing model complexity or improve system efficiency. Understanding these particularities is essential for proposing new low-cost model-embedding methods. Hence, this paper presents a systematic review of deep learning models for image classification using embedded devices. This review covers studies published between 2013 and 2023 and indexed in the ACM Digital Library, IEEE Xplore, PubMed, and Scopus. Our analysis included 111 studies, and we categorized eligible papers based on various attributes from the deep learning models. The key contributions of this study include identifying prevalent trends, challenges, and advancements in the field as well as summarizing techniques that enable the deployment of high-performance models on resource-constrained devices. Our findings are expected to significantly benefit this area by pointing to the perspectives and challenges inherent in the use of low-cost devices for deep learning image applications.}, archive = {J_NEUCOM}, author = {Larissa Ferreira Rodrigues Moreira and Rodrigo Moreira and Bruno Augusto Nassif Travençolo and André Ricardo Backes}, doi = {10.1016/j.neucom.2025.129402}, journal = {Neurocomputing}, month = {3}, pages = {129402}, shortjournal = {Neurocomputing}, title = {Deep learning based image classification for embedded devices: A systematic review}, volume = {623}, year = {2025}, }
• (2025). A noval dual-parameter structural model with enhanced traffic flow representations. NEUCOM, 623, 129401. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) are widely applied in traffic flow prediction due to their excellent capability for information aggregation. However, the dependence of GNNs on node connections often leads to high computational costs. Previous observations indicate that time series traffic data of different nodes share significant similarities, suggesting potential for capturing spatial relationships with limited common factors. Towards this end, our approach introduces an innovative Dual-Parameter Structural model(DPSM) to enhance the data representation of traffic flow time series, which extracts common factors with a reduced number of nodes for modeling similarities, and learns the inherent characteristics of nodes as specific factors simultaneously. Additionally, an attention-based mechanism is developed to exchange information between these two types of factors effectively. Comprehensive evaluations of four real-world datasets demonstrate the effectiveness of our framework for traffic prediction.}, archive = {J_NEUCOM}, author = {Yuxin Wu and Xudong Zhang and Haina Tang}, doi = {10.1016/j.neucom.2025.129401}, journal = {Neurocomputing}, month = {3}, pages = {129401}, shortjournal = {Neurocomputing}, title = {A noval dual-parameter structural model with enhanced traffic flow representations}, volume = {623}, year = {2025}, }
• (2025). Exponential stability of continuous-time and discrete-time neural networks with saturated impulses. NEUCOM, 623, 129400. (www)

  @article{ , abstract = {This paper investigates the exponential stability of neural networks (NNs) with saturated impulses that have different impulsive effects (i.e., stabilizing and destabilizing impulses) in both continuous-time and discrete-time situations. For the purpose of generating discrete-time analogues that maintain the dynamic characteristics of the considered systems, the semi-discretization approach is first applied to the continuous-time NNs with saturated impulses. Secondly, for the prescribed initial condition, the bounds of states of the system at the impulsive instants were estimated by using the quadratic Lyapunov function method and the average impulsive interval (AII) technique, which ensure that the saturation nonlinearities at the impulsive instants can be handled by the sector nonlinear model approach (SNMA) and the polyhedral representation approach (PRA), respectively. Furthermore, less conservative criteria for the exponential stability of the saturated impulsive continuous neural networks and their discrete-time analogues have been derived. Thirdly, an algorithm based on linear matrix inequalities (LMIs) has been presented to achieve an enlarged estimation of the region of attraction (ROA) under stabilizing impulses. This algorithm addresses the pertinent optimization problems and provides the design of an impulsive controller. Finally, the validity of the obtained results is verified through numerical examples and simulations, and the theoretical results are applied to the exponential synchronization of chaotic neural networks under saturated impulsive control.}, archive = {J_NEUCOM}, author = {Zhilong He and Chuandong Li and Hongjuan Wu and Linfei Nie and Zhiyong Yu}, doi = {10.1016/j.neucom.2025.129400}, journal = {Neurocomputing}, month = {3}, pages = {129400}, shortjournal = {Neurocomputing}, title = {Exponential stability of continuous-time and discrete-time neural networks with saturated impulses}, volume = {623}, year = {2025}, }
• (2025). Synchronization of fractional-order neural networks with inertia terms via cumulative reduced-order method. NEUCOM, 623, 129398. (www)

  @article{ , abstract = {This work investigates the synchronization issue of fractional-order inertial neural networks (FOINNs) by employing a cumulative reduced-order method. Firstly, a novel cumulative reduced-order method is presented under the framework of FOINNs, whose idea is that the highest order of the original model is gradually reduced between 0 and β . In particular, the corresponding network models are given for different values of k ( k = 1 , 2 ; k > 3 ) according to the inequality k β ≤ α ≤ ( k + 1 ) β , respectively. Secondly, for the studied FOINNs, a feedback control mechanism is proposed to get three sufficient conditions for achieving synchronization. Besides, considering the flexibility of the control gain, an adaptive control scheme is also designed to achieve synchronization. Lastly, three numerical simulations are performed to confirm the validity of the theoretical results.}, archive = {J_NEUCOM}, author = {Lianyang Hu and Haijun Jiang and Cheng Hu and Yue Ren and Shanshan Chen}, doi = {10.1016/j.neucom.2025.129398}, journal = {Neurocomputing}, month = {3}, pages = {129398}, shortjournal = {Neurocomputing}, title = {Synchronization of fractional-order neural networks with inertia terms via cumulative reduced-order method}, volume = {623}, year = {2025}, }
• (2025). Deep convolutional neural networks meet variational shape compactness priors for image segmentation. NEUCOM, 623, 129395. (www)

  @article{ , abstract = {Integrating spatial priors into image segmentation methods is essential for enhancing prediction accuracy and robustness. In this work, we focus on the shape compactness prior, which measures the circularity of objects, for variational image segmentation tasks. Segmentation of circular objects has significant applications across various fields. We propose utilizing shape compactness as a tool to address these applications. Existing algorithms face challenges such as computational inefficiency, model nonconvexity, and the need for extensive hyperparameter fine-tuning. To overcome these issues, we introduce two novel optimization algorithms coupled with the primal–dual formulation of the original problem: the primal–dual threshold dynamics (PD-TD) and the primal–dual soft threshold-dynamics algorithm (PD-STD). The PD-STD is a relaxed version of the PD-TD and offers superior computational efficiency. Building on the variational explanation of the sigmoid function, the proposed PD-STD algorithm can be integrated into deep neural networks (DNNs) to enhance segmentation results for circular objects. Conventional DNNs often act as black boxes, making them susceptible to noise and adversarial attacks, which leads to inconsistencies in segmentation when input images are corrupted. Our approach improves the robustness and stability of DNNs by enforcing the shape compactness prior information. Extensive experiments demonstrate that the proposed algorithms outperform state-of-the-art methods in both numerical efficiency and effectiveness. This is particularly evident when the algorithms are tested on Iris datasets with high levels of noise. Our findings highlight the potential of incorporating shape compactness priors to enhance image segmentation in challenging scenarios.}, archive = {J_NEUCOM}, author = {Kehui Zhang and Lingfeng Li and Hao Liu and Jing Yuan and Xue-Cheng Tai}, doi = {10.1016/j.neucom.2025.129395}, journal = {Neurocomputing}, month = {3}, pages = {129395}, shortjournal = {Neurocomputing}, title = {Deep convolutional neural networks meet variational shape compactness priors for image segmentation}, volume = {623}, year = {2025}, }
• (2025). Unsupervised domain adaptation segmentation algorithm with cross-domain data augmentation and category contrast. NEUCOM, 623, 129393. (www)

  @article{ , abstract = {Unsupervised domain adaptation (UDA) approaches have shown promise in improving the generalization ability of segmentation models by adapting from a source domain to a target domain using various adaptation strategies. However, most existing UDA approaches primarily focus on adapting from normal-to-normal conditions, while the adaptation from normal-to-adverse conditions is often overlooked. Compared with the former, the latter faces two critical obstacles, 1) The domain gap becomes more pronounced in the input space; 2) The issues of category imbalance and category merging are further exacerbated. In this study, we specifically address the UDA from normal conditions to multiple adverse conditions. We propose a UDA segmentation algorithm with cross-domain data augmentation and category contrast, named UDA c a c c . Our data augmentation strategy involves cross-domain fused style transfer and cross-domain incremental image mixing, which aim to mitigate the larger domain gap in the input space. Additionally, our cross-domain category contrast strategy aims to alleviate the challenges of category imbalance and category merging by encouraging the model to learn domain-invariant yet category-discriminative information in both the feature and output spaces. Experimental results on the Cityscapes → ACDC benchmarks demonstrate the superiority of our proposed method over previous state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Wenyong Dong and Zhixue Liang and Liping Wang and Gang Tian and Qianhui Long}, doi = {10.1016/j.neucom.2025.129393}, journal = {Neurocomputing}, month = {3}, pages = {129393}, shortjournal = {Neurocomputing}, title = {Unsupervised domain adaptation segmentation algorithm with cross-domain data augmentation and category contrast}, volume = {623}, year = {2025}, }
• (2025). Neural networks based finite-time cluster quasi-consensus for unknown nonlinear multi-agent systems with input saturations. NEUCOM, 623, 129391. (www)

  @article{ , abstract = {This paper studies the finite-time cluster quasi-consensus problems for a class of second-order unknown nonlinear multi-agent systems (MASs) with input saturations. Firstly, neural networks are utilized for approximations of unknown functions to deal with the control difficulty caused by unknown systems. In addition, the hyperbolic tangent function is introduced to handle the impact of input saturation. Then Lyapunov stability method and graph theory are used to demonstrate the quasi-consensus of the systems. Finally, simulation results verify the effectiveness of the proposed protocols.}, archive = {J_NEUCOM}, author = {Jia Wu and Wentao Wang and Wenyan Tang and Yongfang Xie and Xiaohong Nian}, doi = {10.1016/j.neucom.2025.129391}, journal = {Neurocomputing}, month = {3}, pages = {129391}, shortjournal = {Neurocomputing}, title = {Neural networks based finite-time cluster quasi-consensus for unknown nonlinear multi-agent systems with input saturations}, volume = {623}, year = {2025}, }
• (2025). Lifted proximal operator machine-based deep nonlinear dictionary learning with multilayer regularization. NEUCOM, 623, 129390. (www)

  @article{ , abstract = {Nonlinear dictionary learning (NLDL) can mine nonlinear information in data better than linear models. Mainstream NLDL methods are based on kernel methods; however, their large storage requirements pose computational challenges and are difficult to scalability. Moreover, the single-layer NLDL framework cannot capture deep features and requires the reversibility of nonlinear functions, leaving room for improving nonlinear signal representation. Inspired by deep dictionary learning, we design a generalized deep nonlinear dictionary learning model called DNLDL_Sp . This model features deep and nonlinear structures. In the proposed model, data are decomposed into nonlinear mappings of several dictionary layers, and the coefficients of the deeper layers are learned via dictionary learning in the preceding layers. We impose sparsity constraints on the coefficients of each layer of DNLDL_Sp to achieve a concise and accurate representation of the data, and this representation proves to be more effective than single-layer structures in capturing deep nonlinear representations. To optimize DNLDL_Sp and prevent issues such as loss of accuracy due to the inverse operation of the nonlinear function, we employ the lifted proximal operator machine (LPOM) technology to transform the nonlinear function into an equivalent proximal operator, which is then incorporated into the model as a penalty. This method converts the nonconvex optimization problem into a series of convex subproblems, allowing us to obtain the optimal dictionaries and coefficients through alternating updates. We demonstrate the effectiveness of our algorithm via image classification and denoising experiments involving different nonlinear functions and sparse constraints.}, archive = {J_NEUCOM}, author = {Jie Lin and Benying Tan and Yujie Li and Yang Qin and Shuxue Ding}, doi = {10.1016/j.neucom.2025.129390}, journal = {Neurocomputing}, month = {3}, pages = {129390}, shortjournal = {Neurocomputing}, title = {Lifted proximal operator machine-based deep nonlinear dictionary learning with multilayer regularization}, volume = {623}, year = {2025}, }
• (2025). Cooperative multi-agent actor-critic approach using adaptive value decomposition and parallel training for traffic network flow control. NEUCOM, 623, 129384. (www)

  @article{ , abstract = {Multi-agent deep reinforcement learning (MADRL) is a promising learning-based data-driven approach, particularly when it is applied to traffic signal control (TSC) in a large-scale urban road network. However, the present MADRL-based TSC methods still deserve further studies at the aspects of mutual coordination among multiple agents and computing efficiency of training algorithms. In this study, we propose an adaptive value decomposition-based multi-agent actor-critic (AVDMAC) approach using a parallel training algorithm for cooperative traffic network flow control. Firstly, the original centralized actor-critic reinforcement learning tasks are decomposed into several local learning tasks, where each actor-critic agent optimizes its control decisions in response to local environment feedback. Additionally, multiple agents striving towards a common objective are coordinated through an adaptive factorization model. In this model, weight coefficients are adaptively assigned to the action-value function of distributed agents by aggregating local rewards, so as to form a global joint action-value function that considers the individual contribution differences among local decisions. Further, we develop the parallel training formulations and implementation algorithm for AVDMAC approach based on Spark cloud in order to relieve of the time consumption load during MADRL training. Numerical experiments demonstrate the advantages at convergence, computational efficiency, and control performance of the AVDMAC approach.}, archive = {J_NEUCOM}, author = {Yongnan Zhang and Yonghua Zhou}, doi = {10.1016/j.neucom.2025.129384}, journal = {Neurocomputing}, month = {3}, pages = {129384}, shortjournal = {Neurocomputing}, title = {Cooperative multi-agent actor-critic approach using adaptive value decomposition and parallel training for traffic network flow control}, volume = {623}, year = {2025}, }
• (2025). Adaptive strategy templates using deep reinforcement learning for multi-issue bilateral negotiation. NEUCOM, 623, 129381. (www)

  @article{ , abstract = {Negotiating in uncertain environments, where user preferences are only partially known, poses a challenge for traditional negotiation models that rely on rigid, pre-defined strategies. These models struggle to adapt to changing conditions or transfer knowledge across different negotiation contexts, making them ineffective in dynamic environments. To address this research gap, we propose a novel negotiation model that uses deep reinforcement learning (DRL) to enable agents learn adaptable, generalizable strategies through the notion of “strategy templates”. These templates include (a) choice parameters to select tactics, (b) time parameters to control when tactics are activated, and (c) attribute-value parameters to guide acceptance and inform bidding decisions. As a result, we enable negotiation agents dynamically adapt their strategies, through pre-training on teacher strategies and refining them via online learning in diverse environments. Our agents also derive a user model to approximate partially specified user preferences, thus handling preference uncertainty more effectively. We developed a proof-of-concept prototype using an actor-critic architecture based on DRL, supplemented by stochastic search techniques for the estimation of user model and multi-objective optimization for making mutually beneficial offers. Experimental evaluations show that our model outperforms state-of-the-art approaches in terms of both individual and social-welfare utilities, demonstrating its ability to transfer experience across domains and excel in previously unseen scenarios. This work provides a robust framework for dynamic, adaptable strategy formation, bridging the gap in current negotiation models by addressing uncertainty in user preferences and strategy flexibility.}, archive = {J_NEUCOM}, author = {Pallavi Bagga and Nicola Paoletti and Kostas Stathis}, doi = {10.1016/j.neucom.2025.129381}, journal = {Neurocomputing}, month = {3}, pages = {129381}, shortjournal = {Neurocomputing}, title = {Adaptive strategy templates using deep reinforcement learning for multi-issue bilateral negotiation}, volume = {623}, year = {2025}, }
• (2025). Multi-source attention autoencoder network for hyperspectral unmixing with LiDAR data. NEUCOM, 623, 129380. (www)

  @article{ , abstract = {Autoencoder (AE) based on Convolutional Neural Networks (CNN) has demonstrated excellent capabilities in capturing high-dimensional features, finding widespread application in hyperspectral unmixing (HU). However, the challenge of ”same-spectrum with different material” phenomenon imposes a performance bottleneck when relying solely on hyperspectral data. In order to address this issue by incorporating the elevation information inherent in Light Detection and Ranging (LiDAR) data, we propose a Spectral-spatial Fusion Attention Network (SFANet). To mitigate semantic disparities among diverse information sources, we devised the Exponential Local Attention Module (ELAM) to achieve complementary information fusion for heterogeneous features. Additionally, dynamic group convolutions are introduced, enabling the network to concurrently learn convolutional grouping strategies and weights. Experimental evaluations conducted on three multi-source remote sensing datasets, comparing our model with state-of-the-art fusion models, demonstrate the superior performance and effectiveness of our proposed approach. The source code is available at https://github.com/qiwenjjin/SFANet .}, archive = {J_NEUCOM}, author = {Jiwei Hu and Yangrui Bai and Zijun Li and Qiwen Jin and Chengli Peng}, doi = {10.1016/j.neucom.2025.129380}, journal = {Neurocomputing}, month = {3}, pages = {129380}, shortjournal = {Neurocomputing}, title = {Multi-source attention autoencoder network for hyperspectral unmixing with LiDAR data}, volume = {623}, year = {2025}, }
• (2025). Robustness evaluation of image classification models based on edge features: Tight and non-tight boundaries. NEUCOM, 623, 129378. (www)

  @article{ , abstract = {State-of-the-art deep neural networks (DNNs) are susceptible to adversarial images, posing a significant threat to safety-critical tasks. The white-box attack is an effective method for testing the robustness of models. Compared to modifying the entire image, perturbing specific image features that impact model decisions while remaining imperceptible to human eyes is more threatening. Besides, more informative and accurate evaluation metrics are needed for robustness assessment. Therefore, this paper initially proposes a set of metrics to demonstrate the significance of edge features in image classification models and validate it across 3 different datasets. Drawing inspiration from the minimum perturbation distance and boundary thickness, we propose Non-Tight boundary width and Tight boundary width as robustness evaluation metrics based on statistical analysis. Finally, a method called E-IFGSM is proposed to generate robustness testing examples on both sides of the decision boundary against the edge feature. To validate the proposed robustness evaluating approach, extensive experiments are conducted on 14 CIFAR10 models with varying structures and training stages, obtaining interesting conclusions regarding the robustness of models. Additionally, the robustness evaluation method is versatile and also extended to the Deepfool method to evaluate the robustness of models from the perspective of entire images.}, archive = {J_NEUCOM}, author = {Hui Lu and Ting Mei and Shiqi Wang and Ruoliu Zhang and Kefei Mao}, doi = {10.1016/j.neucom.2025.129378}, journal = {Neurocomputing}, month = {3}, pages = {129378}, shortjournal = {Neurocomputing}, title = {Robustness evaluation of image classification models based on edge features: Tight and non-tight boundaries}, volume = {623}, year = {2025}, }
• (2025). Multi-time knowledge distillation. NEUCOM, 623, 129377. (www)

  @article{ , abstract = {The purpose of knowledge distillation is to transfer knowledge from a powerful and complex teacher model to a compact student model, enabling the student model to achieve performance similar to that of the teacher model. In recent years, many algorithms meticulously designed for benchmark tasks have been proposed, driving advancements in knowledge distillation. However, these advances have also increased the complexity of distillation frameworks, making hyperparameter tuning and task migration more challenging. Our work combines vanilla knowledge distillation with reverse knowledge distillation and proposes Multi-Time Knowledge Distillation (MTKD). MTKD extends the knowledge distillation process from one single time to multiple times, allowing teacher and student models to alternate as the ones being trained. In this manner, the teacher model enhances its generalization capability while maintaining a student-friendly characteristic, thereby enhancing the performance of the student model. Extensive experiments demonstrate that MTKD achieves performance comparable to representative knowledge distillation methods, without introducing any additional knowledge transfer modules or requiring adjustments to hyperparameters.}, archive = {J_NEUCOM}, author = {Guozhao Chen and Zheyi Fan and Yixuan Zhu and Tingting Zhang}, doi = {10.1016/j.neucom.2025.129377}, journal = {Neurocomputing}, month = {3}, pages = {129377}, shortjournal = {Neurocomputing}, title = {Multi-time knowledge distillation}, volume = {623}, year = {2025}, }
• (2025). MMTF-DES: A fusion of multimodal transformer models for desire, emotion, and sentiment analysis of social media data. NEUCOM, 623, 129376. (www)

  @article{ , abstract = {Desires, emotions, and sentiments are pivotal in understanding and predicting human behavior, influencing various aspects of decision-making, communication, and social interactions. Their analysis, particularly in the context of multimodal data (such as images and texts) from social media, provides profound insights into cultural diversity, psychological well-being, and consumer behavior. Prior studies overlooked the use of image–text pairwise feature representation, which is crucial for the task of human desire understanding. In this research, we have proposed a unified multimodal-based framework with image–text pair settings to identify human desire, sentiment, and emotion. The core of our proposed method lies in the encoder module, which is built using two state-of-the-art multimodal vision-language models (VLMs). To effectively extract visual and contextualized embedding features from social media image and text pairs, we jointly fine-tune two pre-trained multimodal VLMs: Vision-and-Language Transformer (ViLT) and Vision-and-Augmented-Language Transformer (VAuLT). Subsequently, we use an early fusion strategy on these embedding features to obtain combined diverse feature representations. Moreover, we leverage a multi-sample dropout mechanism to enhance the generalization ability and expedite the training process of our proposed method. To evaluate our proposed approach, we used the multimodal dataset MSED for the human desire understanding task. Through our experimental evaluation, we demonstrate that our method excels in capturing both visual and contextual information, resulting in superior performance compared to other state-of-the-art techniques. Specifically, our method outperforms existing approaches by 3% for sentiment analysis, 2.2% for emotion analysis, and approximately 1% for desire analysis.}, archive = {J_NEUCOM}, author = {Abdul Aziz and Nihad Karim Chowdhury and Muhammad Ashad Kabir and Abu Nowshed Chy and Md. Jawad Siddique}, doi = {10.1016/j.neucom.2025.129376}, journal = {Neurocomputing}, month = {3}, pages = {129376}, shortjournal = {Neurocomputing}, title = {MMTF-DES: A fusion of multimodal transformer models for desire, emotion, and sentiment analysis of social media data}, volume = {623}, year = {2025}, }
• (2025). Enhancing few-shot KB-VQA with panoramic image captions guided by large language models. NEUCOM, 623, 129373. (www)

  @article{ , abstract = {Current state-of-the-art (SOTA) KB-VQA techniques involve transforming images into image captions as prompts to harness the potent reasoning capabilities of large language models (LLMs) for generating answers. However, generic image captions often fall short in capturing crucial visual details, essential for LLMs to deliver precise responses . To address this challenge, we propose an image captioning model that effectively utilizes a set of visual language models, such as BLIP2, GRiT, OCR, etc., to extract rich visual information from images. Subsequently, we employ the inferential and summarization capabilities of LLM to generate panoramic image descriptions enriched with intricate details . Simultaneously, we employ Contextual Constraint Examples and Constraint Instruction to mitigate the potential hallucination issues arising from LLM-generated image captions. Extensive experiments validate the superiority and scalability of our proposed method, achieving significant improvements over SOTA methods in challenging few-shot settings. For instance, on the challenging OK-VQA, our method outperforms PICa by 6.5%. On the VQAv2 dataset, our method surpasses the SOTA approach by 5.4%.}, archive = {J_NEUCOM}, author = {Pengpeng Qiang and Hongye Tan and Xiaoli Li and Dian Wang and Ru Li and Xinyi Sun and Hu Zhang and Jiye Liang}, doi = {10.1016/j.neucom.2025.129373}, journal = {Neurocomputing}, month = {3}, pages = {129373}, shortjournal = {Neurocomputing}, title = {Enhancing few-shot KB-VQA with panoramic image captions guided by large language models}, volume = {623}, year = {2025}, }
• (2025). A comprehensive survey on recent feature selection methods for mixed data: Challenges, solutions and future directions. NEUCOM, 623, 129372. (www)

  @article{ , abstract = {Feature selection plays a crucial role in data analysis, pattern recognition, and machine learning by reducing feature dimensions, improving execution time, and enhancing model accuracy. It helps prevent overfitting by removing irrelevant and redundant features, ultimately leading to better model performance. The demand for robust feature selection methods tailored to mixed data has grown due to the increasing focus on multi-modal decision-making and classification in areas such as healthcare, finance, and social media analysis. Selecting features for mixed data presents unique challenges, including correlation assessment, model compatibility, and interpretability, requiring different approaches compared to homogeneous data. This comprehensive review delves into feature selection methodologies specifically designed for mixed data, offering insights into the suitability and effectiveness of various techniques. The surveyed methods cover both supervised and unsupervised approaches, addressing the considerations, challenges, and potential pitfalls inherent in feature selection for mixed data. Additionally, the review introduces benchmark datasets and evaluation metrics crucial for assessing these methods, concluding with a detailed comparison of recent approaches, highlighting their contributions, strengths, and weaknesses. Finally, the review sheds light on future trends and research directions in the field, providing a roadmap for continued advancements in feature selection for mixed data.}, archive = {J_NEUCOM}, author = {Mahshid Ebrahimi Warkiani and Mohammad Hossein Moattar}, doi = {10.1016/j.neucom.2025.129372}, journal = {Neurocomputing}, month = {3}, pages = {129372}, shortjournal = {Neurocomputing}, title = {A comprehensive survey on recent feature selection methods for mixed data: Challenges, solutions and future directions}, volume = {623}, year = {2025}, }
• (2025). Dynamic label correlations and dual-semantic enhancement learning for cross-modal retrieval. NEUCOM, 623, 129371. (www)

  @article{ , abstract = {With the rapid growth of multi-modal data, Cross-Modal Hashing (CMH) is widely applied due to its outstanding performance in both search and storage. Nevertheless, there are two issues to be further addressed: (1) most existing methods neglect dynamic learning of the importance of different labels; and (2) many methods fail to purify the consistency of data extracted from different feature spaces. For this purpose, we propose a method called Dynamic Label Correlations and Dual-Semantic Enhancement Learning for Cross-Modal Retrieval (DLCDE) in this study. This method is formed of two parts: Label Semantic Enhancement with Dynamic Label Reconstruction (LSEDLR) and Sample Semantic Enhancement with Consistency Purification and Structure Maintenance (SECPSM). The former first utilizes label-wise self-expression to dynamically explore the latent correlations between different labels and then employs a graph-based manifold regularizer to explore the structural relationships in the transformed label space to enhance label semantics, the latter leverages Hadamard-Product-based Matrix Factorization to enhance the common relationships between samples, thereby enhancing the sample semantics of the latent shared space. Moreover, dual-semantic enhancement learning is achieved by integrating enhanced label semantics and sample semantics in Distance-Distance Difference Minimization (DDDM). Numerous experiments on four benchmark datasets reveal that DLCDE surpasses a number of state-of-the-art CMH methods. The source code for DLCDE is publicly available at https://github.com/Fizzyf/DLCDE .}, archive = {J_NEUCOM}, author = {Shaohua Teng and Ziye Fang and Zefeng Zheng and Naiqi Wu and Wei Zhang and Luyao Teng}, doi = {10.1016/j.neucom.2025.129371}, journal = {Neurocomputing}, month = {3}, pages = {129371}, shortjournal = {Neurocomputing}, title = {Dynamic label correlations and dual-semantic enhancement learning for cross-modal retrieval}, volume = {623}, year = {2025}, }
• (2025). An adaptive gradient method with transformation bound. NEUCOM, 623, 129369. (www)

  @article{ , abstract = {Deep learning typically uses optimizers in two categories. The first category contains the stochastic gradient descent (SGD) family, which demonstrates superior generalization performance during the later stages of training. The second category features adaptive methods, such as Adam, which achieve faster convergence in the early stages of training. To ensure effective training, transitioning from adaptive methods to the SGD family after a certain number of iterations is crucial. We propose the AdaTB algorithm, which utilizes a specialized transformation function to enable a smooth transition from Adam to SGDM. Furthermore, we introduce a novel framework and provide a proof of the convergence of AdaTB within this framework in the non-convex setting. Finally, we verify the effectiveness of AdaTB using the Rastrigin function, CIFAR-10/100, and PTB datasets. The code can be accessed at https://github.com/Wenhan-Jiang/AdaTB.git .}, archive = {J_NEUCOM}, author = {Wenhan Jiang and Kun Zeng and Zhixia Jiang and Dongpo Xu}, doi = {10.1016/j.neucom.2025.129369}, journal = {Neurocomputing}, month = {3}, pages = {129369}, shortjournal = {Neurocomputing}, title = {An adaptive gradient method with transformation bound}, volume = {623}, year = {2025}, }
• (2025). CSDSE: An efficient design space exploration framework for deep neural network accelerator based on cooperative search. NEUCOM, 623, 129366. (www)

  @article{ , abstract = {The design and optimization of deep neural network accelerators necessitates thoughtful consideration of numerous design parameters and various resource/physical constraints that render their design spaces massive in scale and complex in distribution. When faced with these massive and complex design spaces, previous works on design space exploration confront the exploration–exploitation dilemma, struggling to concurrently ensure optimization efficiency and stability. To address the exploration–exploitation dilemma, we present a novel design space exploration method entitled CSDSE. CSDSE implements heterogeneous agents separately accountable for exploration or exploitation to cooperatively search the design space. In order to enable CSDSE to adapt design spaces with various space distributions and expanding scales, we extend CSDSE with mechanism of adaptive agent organization and multi-scale search. Furthermore, we introduce a weighted compact buffer that encourages agents to search in diverse directions and bolsters their global exploration ability. CSDSE is implemented to optimize accelerator design. Compared to former DSE methods, it achieves latency speedups of up to 15.68x and energy-delay-product reductions of up to 16.22x under different constraint scenarios.}, archive = {J_NEUCOM}, author = {Kaijie Feng and Xiaoya Fan and Jianfeng An and Haoyang Wang and Chuxi Li}, doi = {10.1016/j.neucom.2025.129366}, journal = {Neurocomputing}, month = {3}, pages = {129366}, shortjournal = {Neurocomputing}, title = {CSDSE: An efficient design space exploration framework for deep neural network accelerator based on cooperative search}, volume = {623}, year = {2025}, }
• (2025). A nearly optimal adaptive saturation function tuning method for quasi-sliding mode control based on integral reinforcement learning. NEUCOM, 623, 129363. (www)

  @article{ , abstract = {The problem we focus on in this paper is to find a nearly optimal quasi-sliding mode controller of continuous-time nonlinear systems that can both eliminate chattering and minimize the cost function, which is a measure of the performance index of dynamic systems. First, the deficiency of chattering in traditional sliding mode control (SMC) is analyzed. In quasi-SMC, the signum function of the traditional SMC is replaced with a continuous function to eliminate chattering. Then, an optimal adaptive saturation function tuning algorithm based on integral reinforcement learning (IRL) is proposed. Under an initial SMC policy, the proposed method can learn the nearly optimal saturation function using policy iteration. To satisfy the requirements of the learned saturation function, we treat the problem of training the saturation function as an optimization problem with constraints. The online neural networks implementation of the proposed algorithm is presented based on symmetric radius basis functions and a regularized batch least-squares (BLS) algorithm to train the control law in this paper. An optimal adaptive quasi-SMC control algorithm is obtained. The system property of input-to-state stability (ISS) is guaranteed. Finally, three examples are simulated and compared with the quasi-sliding mode controller to verify the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Lei Guo and Wenbo Xiong and Han Zhao and Yuan Song and Dongming Gan}, doi = {10.1016/j.neucom.2025.129363}, journal = {Neurocomputing}, month = {3}, pages = {129363}, shortjournal = {Neurocomputing}, title = {A nearly optimal adaptive saturation function tuning method for quasi-sliding mode control based on integral reinforcement learning}, volume = {623}, year = {2025}, }
• (2025). How to form brain-like memory in spiking neural networks with the help of frequency-induced mechanism. NEUCOM, 623, 129361. (www)

  @article{ , abstract = {Understanding how memory is formed in the large-scale biological brain with sparse structure has haunted human beings for many decades. Spiking Neural Networks (SNNs), which emulate brain-like spiking communication and neural dynamics, offer a promising model for modeling brain-like memory. However, achieving practical brain-like memory in SNNs remains a challenging frontier. Inspired by the brain’s focus on frequent spike patterns during memorization, this work introduces frequency-induced potentiation (FIP) and frequency-induced depression (FID) to guide the formation and retrieval of brain-like memory in SNNs. Neurons exhibiting high-frequency firing enter the FIP state, enhancing their excitability and promoting neuroplasticity, while neurons in the FID state become inactive, inhibiting plasticity. By combining FIP/FID with structural and weight plasticity, SNNs generate cell assemblies whose structure and weight distributions represent memory. This approach enables memory formation with properties such as convergence, recollection, separation, and conceptualization, within sparse network topologies. The FIP/FID mechanism also promotes network stability and long-term memory retention. Furthermore, this memory proves immune to interference and can be accurately recalled in subsequent tasks. This work could be further combined with cognitive science to help researchers better understand the nature of the memory forming process in human brain.}, archive = {J_NEUCOM}, author = {Yunlin Lei and Huiqi Li and Mingrui Li and Yaoyu Chen and Yu Zhang and Zihui Jin and Xu Yang}, doi = {10.1016/j.neucom.2025.129361}, journal = {Neurocomputing}, month = {3}, pages = {129361}, shortjournal = {Neurocomputing}, title = {How to form brain-like memory in spiking neural networks with the help of frequency-induced mechanism}, volume = {623}, year = {2025}, }
• (2025). VI-PINNs: Variance-involved physics-informed neural networks for fast and accurate prediction of partial differential equations. NEUCOM, 623, 129360. (www)

  @article{ , abstract = {Although physics-informed neural networks (PINNs) have progressed a lot in many real-world applications recently, there remain problems to be further studied. Quantifying predictive uncertainty in PINNs is one of the most challenging problems and has not been thoroughly solved yet. In this work, we consider the problems of predictive uncertainty in PINNs and propose our variance-involved physics-informed neural networks (VI-PINNs) to give better predictions. Different from a single output in vanilla PINNs, the final layer of our model outputs two values to represent the predicted mean and variance respectively, and the latter is used to represent the predictive uncertainty of the output. A modified negative log-likelihood (MNLL) loss is introduced and treated as an auxiliary task for fast and accurate training. We perform several experiments including high-dimensional partial differential equations to highlight the advantages of our approach. The results convey that our method not only gives more accurate predictions but also converges faster.}, archive = {J_NEUCOM}, author = {Bin Shan and Ye Li and Sheng-Jun Huang}, doi = {10.1016/j.neucom.2025.129360}, journal = {Neurocomputing}, month = {3}, pages = {129360}, shortjournal = {Neurocomputing}, title = {VI-PINNs: Variance-involved physics-informed neural networks for fast and accurate prediction of partial differential equations}, volume = {623}, year = {2025}, }
• (2025). Deep clustering of tabular data by weighted gaussian distribution learning. NEUCOM, 623, 129359. (www)

  @article{ , abstract = {Deep learning methods are primarily proposed for supervised learning of images or text with limited applications to clustering problems. In contrast, tabular data with heterogeneous feature space pose unique challenges in representation learning, where deep learning has yet to replace traditional machine learning. This paper addresses these challenges by developing one of the first deep clustering methods for tabular data, Gaussian Cluster Embedding in Autoencoder Latent Space (G-CEALS). G-CEALS is an unsupervised deep clustering framework for learning the parameters of multivariate Gaussian cluster distributions by iteratively updating individual cluster weights. The G-CEALS method presents average rank orderings of 2.9 (1.7) and 2.8 (1.7) based on clustering accuracy and adjusted Rand index (ARI) scores on sixteen tabular data sets, respectively, and outperforms nine state-of-the-art clustering methods. G-CEALS substantially improves clustering performance compared to traditional K-means and GMM, which are still de facto methods for clustering tabular data. Similar computationally efficient and high-performing deep clustering frameworks are imperative to reap the myriad benefits of deep learning on tabular data over traditional machine learning.}, archive = {J_NEUCOM}, author = {Shourav B. Rabbani and Ivan V. Medri and Manar D. Samad}, doi = {10.1016/j.neucom.2025.129359}, journal = {Neurocomputing}, month = {3}, pages = {129359}, shortjournal = {Neurocomputing}, title = {Deep clustering of tabular data by weighted gaussian distribution learning}, volume = {623}, year = {2025}, }
• (2025). Mi-CGA: Cross-modal graph attention network for robust emotion recognition in the presence of incomplete modalities. NEUCOM, 623, 129342. (www)

  @article{ , abstract = {Multimodal Emotion Recognition in Conversation (Multimodal ERC) is crucial for understanding human communication across various applications. However, the challenge of missing modalities impedes the development of robust models. Existing approaches often overlook scenarios where multiple modalities are absent simultaneously and fail to explore deep semantic interactions between modalities. Additionally, learning high-dimensional interactive features from limited samples is challenging due to missing data. This paper proposes Mi-CGA, a framework tailored for incomplete multimodal learning in conversational contexts. Mi-CGA comprises two main components: Incomplete Multimodal Representation (IMR) and Cross-modal Graph Attention Network (CGA-Net). IMR simulates incomplete modalities, while CGA-Net extracts rich information from conversational graphs. CGA-Net consists of three key modules: Modality Feature Estimation reconstructs missing data, Multi-head Graph Attention Network enhances utterance-level representation, and Cross-modal Attention Network improves conversation-level representation. Experimental results on three benchmark datasets (IEMOCAP, CMU-MOSI, and CMU-MOSEI) consistently demonstrate that Mi-CGA outperforms several representative baseline models, marking a significant advancement for the Multimodal ERC task. Source code for Mi-CGA is available at https://github.com/dangkh/Mi-CGA .}, archive = {J_NEUCOM}, author = {Cam-Van Thi Nguyen and Hai-Dang Kieu and Quang-Thuy Ha and Xuan-Hieu Phan and Duc-Trong Le}, doi = {10.1016/j.neucom.2025.129342}, journal = {Neurocomputing}, month = {3}, pages = {129342}, shortjournal = {Neurocomputing}, title = {Mi-CGA: Cross-modal graph attention network for robust emotion recognition in the presence of incomplete modalities}, volume = {623}, year = {2025}, }
• (2025). An analysis of graph neural networks for fake review detection: A systematic literature review. NEUCOM, 623, 129341. (www)

  @article{ , abstract = {Over the past decade, detecting fake reviews has emerged as a critical challenge in ensuring the credibility of online businesses. The capability of Graph Neural Networks (GNNs) to handle data structured in graphs is a key advantage, enabling a nuanced representation of the complex relationships among reviews, users, and products. With the growing demand for robust solutions, there has been a rising focus in the research community on integrating GNNs into fake review detection mechanisms. This research conducts a thorough systematic literature review covering GNN-based articles, conference papers, and scholarly works from 2019 to January 2024, examining the current state and challenges. Initially, we provide an introduction to the problem’s background and a broad overview of spam review methods based on GNNs. Additionally, we highlight existing techniques for detecting fake reviews, specifically emphasizing GNNs-based models categorized by types. Subsequently, we compare key concepts, merits, and drawbacks of these methods within their respective categories. Finally, we outline several limitations and unresolved issues that merit future attention in this domain and present potential avenues for further research. We hope this study can serve as a valuable resource for researchers, system practitioners, and newcomers to overcome current challenges and navigate forthcoming scenarios when deploying a spam review detection system with GNNs.}, archive = {J_NEUCOM}, author = {Ramadhani A. Duma and Zhendong Niu and Ally S. Nyamawe and Ali Asghar Manjotho}, doi = {10.1016/j.neucom.2025.129341}, journal = {Neurocomputing}, month = {3}, pages = {129341}, shortjournal = {Neurocomputing}, title = {An analysis of graph neural networks for fake review detection: A systematic literature review}, volume = {623}, year = {2025}, }
• (2025). A comprehensive survey of transfer dictionary learning. NEUCOM, 623, 129322. (www)

  @article{ , abstract = {Despite the noteworthy advancements in the field of transfer dictionary learning, a comprehensive survey remains conspicuously absent. Moreover, since cross-domain learning and domain-adaptive learning are two main directions of transfer learning. Consequently, this paper aims to extensively review the advancements achieved in the cross-domain dictionary learning algorithms as well as the domain-adaptive dictionary learning algorithms. Firstly, based on the learning strategies, the cross-domain dictionary learning algorithms can be segmented into supervised cross-domain dictionary learning algorithms, semi-supervised cross-domain dictionary learning algorithms, and unsupervised cross-domain dictionary learning algorithms. Correspondingly, the domain-adaptive dictionary learning algorithms can also be categorized into three distinct groups: supervised domain-adaptive dictionary learning algorithms, semi-supervised domain-adaptive dictionary learning algorithms, and unsupervised domain-adaptive dictionary learning algorithms. Then, we provide detailed analyses of the research progress and faced challenges, and elaborate on the working principles of a typical algorithm for each category. Secondly, to validate the competence of the domain-adaptive dictionary learning and cross-domain dictionary learning algorithms, we conducted a comparative evaluation against several forefront dictionary learning algorithms, shallow transfer learning algorithms, and deep transfer learning algorithms. Experimental results show that the transfer dictionary learning algorithms exhibit superior performance than the forefront dictionary learning algorithms and transfer learning algorithms on the five well-known datasets. Compared to the single-domain limitation of traditional dictionary learning, cross-domain dictionary learning uses multi-domain data to enrich feature representation. This not only enhances the dictionary discriminative but also improves model generalization. Moreover, domain-adaptive dictionary learning integrates domain-adaptive characteristics while retaining the advantages of dictionary learning. This achieves dual optimization for sparse representation of features and cross-domain transfer. Lastly, we analyze the challenges faced by the transfer dictionary learning algorithms and present five future research directions. This survey would aid readers in achieving a systematic understanding of dictionary learning and transfer learning, selecting appropriate cross-domain and domain-adaptive dictionary learning algorithms for different scenarios.}, archive = {J_NEUCOM}, author = {Mengyao Li and Yang Li and Zhengming Li}, doi = {10.1016/j.neucom.2024.129322}, journal = {Neurocomputing}, month = {3}, pages = {129322}, shortjournal = {Neurocomputing}, title = {A comprehensive survey of transfer dictionary learning}, volume = {623}, year = {2025}, }
• (2025). Active inference tree search in large POMDPs. NEUCOM, 623, 129319. (www)

  @article{ , abstract = {The ability to plan ahead efficiently is key for both living organisms and artificial systems. Model-based planning and prospection are widely studied in cognitive neuroscience and artificial intelligence (AI), but from different perspectives—and with different desiderata in mind (biological realism versus scalability) that are difficult to reconcile. Here, we introduce a novel method to plan in POMDPs— Active Inference Tree Search (AcT)—that combines the normative character and biological realism of a leading planning theory in neuroscience (Active Inference) and the scalability of tree search methods in AI. This unification enhances both approaches. On the one hand, tree searches enable the biologically grounded, first principle method of active inference to be applied to large-scale problems. On the other hand, active inference provides a principled solution to the exploration–exploitation dilemma, which is often addressed heuristically in tree search methods. Our simulations show that AcT successfully navigates binary trees that are challenging for sampling-based methods, problems that require adaptive exploration, and the large POMDP problem ‘ RockSample’ —in which AcT reproduces state-of-the-art POMDP solutions. Furthermore, we illustrate how AcT can simulate neurophysiological responses (e.g., in the hippocampus and prefrontal cortex) of humans and other animals that solve large planning problems. These numerical analyses show that Active Tree Search is a principled realisation of neuroscientific and AI planning theories, offering biological realism and scalability.}, archive = {J_NEUCOM}, author = {Domenico Maisto and Francesco Gregoretti and Karl J. Friston and Giovanni Pezzulo}, doi = {10.1016/j.neucom.2024.129319}, journal = {Neurocomputing}, month = {3}, pages = {129319}, shortjournal = {Neurocomputing}, title = {Active inference tree search in large POMDPs}, volume = {623}, year = {2025}, }
• (2025). L2 cluster synchronization for discrete-time neural networks with intermittent control. NEUCOM, 623, 129314. (www)

  @article{ , abstract = {This work addresses the cluster synchronization for discrete-time neural networks over energy-constrained communication channels subjecting to packet loss. A new intermittent control strategy including packet loss is employed, which not only reduces communication load but also save control energy. Additionally, the sufficient condition to achieve cluster synchronization is obtained by constructing interval-dependent function, and the l 2 -gain performance of the synchronization error is ensured to reduce the impact of unexpected exogenous perturbations via the intermittent control. Finally, the effectiveness and demonstrated of the results are illustrated by a numerical example.}, archive = {J_NEUCOM}, author = {Zijing Xiao and Zitian Li and Yumei Zhou and Yao Zhao and Tianyang Zhang}, doi = {10.1016/j.neucom.2024.129314}, journal = {Neurocomputing}, month = {3}, pages = {129314}, shortjournal = {Neurocomputing}, title = {L2 cluster synchronization for discrete-time neural networks with intermittent control}, volume = {623}, year = {2025}, }
• (2025). HiREN: Towards higher supervision quality for better scene text image super-resolution. NEUCOM, 623, 129309. (www)

  @article{ , abstract = {Scene text image super-resolution (STISR) is an important pre-processing technique for text recognition from low-resolution scene images. Nowadays, various methods have been proposed to extract text-specific information from high-resolution (HR) images to supervise STISR model training. However, due to uncontrollable factors ( e.g. shooting equipment, focus, and environment) in manually photographing HR images, the quality of HR images cannot be guaranteed, which unavoidably impacts STISR performance. Observing the quality issue of HR images, in this paper we propose a novel idea to boost STISR by first enhancing the quality of HR images and then using the enhanced HR images as supervision to do STISR. Concretely, we develop a new STISR framework, called H igh- R esolution EN hancement (HiREN) that consists of two branches and a quality estimation module. The first branch is developed to recover the low-resolution (LR) images, and the other is an HR quality enhancement branch aiming at generating high-quality (HQ) text images based on the HR images to provide more accurate supervision to the LR images. As the degradation from HQ to HR may be diverse, and there is no pixel-level supervision for HQ image generation, we design a kernel-guided enhancement network to handle various degradation, and exploit the feedback from a recognizer and text-level annotations as weak supervision signal to train the HR enhancement branch. Then, a quality estimation module is employed to evaluate the qualities of HQ images, which are used to suppress the erroneous supervision information by weighting the loss of each image. Extensive experiments on TextZoom show that HiREN can work well with most existing STISR methods and significantly boost their performances.}, archive = {J_NEUCOM}, author = {Minyi Zhao and Yi Xu and Bingjia Li and Jie Wang and Jihong Guan and Shuigeng Zhou}, doi = {10.1016/j.neucom.2024.129309}, journal = {Neurocomputing}, month = {3}, pages = {129309}, shortjournal = {Neurocomputing}, title = {HiREN: Towards higher supervision quality for better scene text image super-resolution}, volume = {623}, year = {2025}, }
• (2025). An efficient action proposal processing approach for temporal action detection. NEUCOM, 623, 129294. (www)

  @article{ , abstract = {Temporal action detection is a fundamental yet challenging task in video understanding. It is important to process the action proposals for action classification and temporal boundary localization. Some methods process action proposals by exploiting the relations between them. However, learning the relations between numerous action proposals is time-consuming and requires huge computation and memory storage. Each proposal contains contextual information extracted from video segments, and redundant information aggregation has a negative impact on the final detection performance. In this paper, we exploit an efficient model which processes each proposal individually and learn intra-proposal features adequately, avoiding the interference of redundant information to achieve more effective detection. We also design relational learning models based on mean pooling, self-attention, and temporal convolution to compare with the intra-proposal learning model. Extensive experiments show that our method outperforms the relation learning models and achieves competitive performance on the two standard benchmarks. Moreover, efficiency experiments also verify that our model is more efficient than the relation learning methods.}, archive = {J_NEUCOM}, author = {Xuejiao Hu and Jingzhao Dai and Ming Li and Yang Li and Sidan Du}, doi = {10.1016/j.neucom.2024.129294}, journal = {Neurocomputing}, month = {3}, pages = {129294}, shortjournal = {Neurocomputing}, title = {An efficient action proposal processing approach for temporal action detection}, volume = {623}, year = {2025}, }
• (2025). Improved young fruiting apples target recognition method based on YOLOv7 model. NEUCOM, 623, 129186. (www)

  @article{ , abstract = {Accurate detection of apples is a key factor in improving the efficiency of bagging robots. During the young fruiting stage, factors such as similarities between young apples and background leaves, occlusion, and small targets pose a challenge to apple detection. In this study, an apple dataset of 3590 images is constructed at the apple base in Yuncheng City, Shanxi Province. The Squeeze Excitation Block (SE Block), the Atrous Spatial Pyramid Pooling and Convolutional Block Attention Module (ASPP-CBAM), along with an additional low-level prediction head (P2) are fused in the You Only Look Once v7 (YOLOv7) for detecting young fruiting apples, named YOLOv7-SAP. Firstly, to diminish the rate of false detections caused by the color similarity between fruit and background, the SE Block is added to the backbone network of YOLOv7 to selectively emphasize the information characteristics. Secondly, to minimize the loss of information about young fruits due to occlusion, the ASPP is added in front of the SE Block to complement the extracted young apple features. Finally, to address the issue of model oversampling affecting small target detection, the P2 layer is introduced within our enhanced model structures. To mitigate discrepancies between predicted and actual frames, CIoU loss in the original network is substituted with EIoU loss. The performance of the YOLOv7-SAP model is documented with a precision (P), the recall (R), the mean average precision (mAP@.5), and the F1 score of the YOLOv7-SAP model reach 89.30%, 88.20%, 91.20%, and 88.75%, respectively. Compared with YOLOv3, YOLOv5, YOLOv7, SSD, and Faster R-CNN, the YOLOv7-SAP model significantly improved the detection of young fruiting apples.}, archive = {J_NEUCOM}, author = {Bingxiu Shi and Chengkai Hou and Xiaoli Xia and Yunhong Hu and Hua Yang}, doi = {10.1016/j.neucom.2024.129186}, journal = {Neurocomputing}, month = {3}, pages = {129186}, shortjournal = {Neurocomputing}, title = {Improved young fruiting apples target recognition method based on YOLOv7 model}, volume = {623}, year = {2025}, }
• (2025). Focusing on feature-level domain alignment with text semantic for weakly-supervised domain adaptive object detection. NEUCOM, 622, 129435. (www)

  @article{ , abstract = {Domain adaptive object detection applies the knowledge learned in the source domain to the target domain, thereby improving object detection performance. The existing work focuses on introducing image-level annotations for the target domain and enhancing visual semantics through Weakly-Supervised Domain Adaptive (WSDA) learning paradigms. However, we observe that solely leveraging vision annotations fails to effectively guide the network in learning the intrinsic knowledge of indistinct boundary features, because the similarity between the source and target domain features can lead to entanglement issues. Fortunately, recent research has shown that the Contrastive Language Image Pre-training (CLIP) effectively enhances visual feature representation by introducing text semantics. Inspired by this, we propose the DACLIP, which utilizes text semantics to encourage feature-level domain alignment and solve the problem of feature entanglement. Firstly, we design the Domain-wise Text Assistant (DTA) module, which learns the inherent attributes of the global feature by providing specific text prompts for the domain distribution. Secondly, we design the Class-wise Text Assistant (CTA) module, which utilizes text semantics to guide visual learning and eliminate interference from similar instance attributes at a finer granularity. Unlike the previous method of directly using CLIP instead of detection heads, DACLIP focuses on the feature-level domain alignment, matching the inherent attributes of different domains in the shared feature space from coarse to fine. The proposed DACLIP achieves state-of-the-art performance on multiple datasets, particularly in the challenging PASCAL VOC → Watercolor scenario where it achieves 62.2% mAP.}, archive = {J_NEUCOM}, author = {Zichong Chen and Jian Cheng and Ziying Xia and Yongxiang Hu and Xiaochen Li and Zhicheng Dong and Nyima Tashi}, doi = {10.1016/j.neucom.2025.129435}, journal = {Neurocomputing}, month = {3}, pages = {129435}, shortjournal = {Neurocomputing}, title = {Focusing on feature-level domain alignment with text semantic for weakly-supervised domain adaptive object detection}, volume = {622}, year = {2025}, }
• (2025). Data-driven optimal shared control of unmanned aerial vehicles. NEUCOM, 622, 129428. (www)

  @article{ , abstract = {Cooperation between humans and autonomy is a critical topic of unmanned aerial vehicle (UAV) control. How to co-pilot the UAV with human operator to achieve optimal performance presents a significant challenge. In this paper, we propose a novel data-driven optimal shared control method for UAV using the Koopman operators to predict the nonlinear dynamics of the UAVs. An original shared control mechanism is established to allocate the relationship between optimal and human control inputs. The model of the system is learned from human maneuver data via the Koopman operator approach, and the optimal controller is approximated online using reinforcement learning techniques. The Lyapunov theory analyzes the stability of the proposed method. Compared with offline RL methods, the proposed method can learn the optimal controller online without a precise UAV dynamics model from human maneuver data. The effectiveness of the proposed method is demonstrated by numerical and Human-in-the-loop (HiTL) simulation.}, archive = {J_NEUCOM}, author = {Junkai Tan and Shuangsi Xue and Zihang Guo and Huan Li and Hui Cao and Badong Chen}, doi = {10.1016/j.neucom.2025.129428}, journal = {Neurocomputing}, month = {3}, pages = {129428}, shortjournal = {Neurocomputing}, title = {Data-driven optimal shared control of unmanned aerial vehicles}, volume = {622}, year = {2025}, }
• (2025). A data-driven predictive control strategy with application to tandem cold rolling process. NEUCOM, 622, 129418. (www)

  @article{ , abstract = {The thickness and tension control is crucial for the improvement of strip product quality in the tandem cold rolling process, which is essentially a coupling control problem. In addition, since the great scale and complexity of the rolling process, it is hard to establish an accurate mechanism model. As a result, the traditional model-based decoupling control methods are infeasible. To cope with such problem, the gated recurrent unit network is first used to predict the thickness and tension. And the Bayesian optimization algorithm is utilized to select the network hyperparameters automatically. Then, based on the predicted values, the coupling control problem is transformed into a multi-objective optimization problem and further solved by using an adaptive gradient descent method. Specifically, a watermarking based encryption scheme is given to enhance the security of data transmission. Furthermore, the convergence of the network predictive model and the stability of the proposed control method are also proved theoretically. Finally, a tandem cold rolling process is applied to show the effectiveness of the developed method.}, archive = {J_NEUCOM}, author = {Zhen-Lei Ma and Xiao-Jian Li}, doi = {10.1016/j.neucom.2025.129418}, journal = {Neurocomputing}, month = {3}, pages = {129418}, shortjournal = {Neurocomputing}, title = {A data-driven predictive control strategy with application to tandem cold rolling process}, volume = {622}, year = {2025}, }
• (2025). A lightweight human pose estimation model based on nonlinear spiking neural convolution model. NEUCOM, 622, 129413. (www)

  @article{ , abstract = {Human pose estimation is an important task in the field of computer vision. Existing high-precision models need more parameters and computational complexity, making them unsuitable for resource-constrained devices. On the other hand, current lightweight models often have limited accuracy for reduced parameters and computations. To address the challenge, this paper proposes a novel lightweight human pose estimation model based on nonlinear spiking neural P system (NSNP). It adopts a lightweight backbone network similar to the HRNet framework. Moreover, the nonlinear spiking neural convolutional model introduces a novel Dual-channel Atrous Spatial Pooling (DCASP) module. This module facilitates the interaction of information from different scales of receptive fields to capture details in the image. Experimental results demonstrate that the proposed model achieves 89.0% accuracy on the MPII dataset with only 4.5 million parameters and 71.5% average accuracy on the COCO2017 validation set with a mere 4.2 Gflops computational complexity. The experimental results of this paper’s model surpass those of the latest lightweight models, indicating that this approach effectively balances accuracy and complexity. The code is available at https://github.com/YuxuanZhang-web/Lite-NSNP-Pose .}, archive = {J_NEUCOM}, author = {Yuxuan Zhang and Meichen Xia and Hong Peng and Zhicai Liu and Jun Guo}, doi = {10.1016/j.neucom.2025.129413}, journal = {Neurocomputing}, month = {3}, pages = {129413}, shortjournal = {Neurocomputing}, title = {A lightweight human pose estimation model based on nonlinear spiking neural convolution model}, volume = {622}, year = {2025}, }
• (2025). Semantic-guided complementary fusion network for salient object detection. NEUCOM, 622, 129383. (www)

  @article{ , abstract = {Most of salient object detection methods are based on U-shape architecture. Although they have achieved good performance, this structure tends to gradually dilute the high-level semantic information during the top-down fusion process, leading to imprecise localization of salient objects. Additionally, due to the limitation of receptive fields, CNN-based methods often underperform in detecting irregular-scale objects and multiple salient objects. To address the aforementioned issues, we propose a semantic-guided complementary fusion network for salient object detection, which includes global semantic extraction (GSE) module, bilateral complementary fusion (BCF) module and context feature refinement (CFR) module. Specifically, GSE is designed to obtain global semantic features for better localization of salient objects. BCF is proposed to fuse the global semantic features from GSE with the local detailed features from the backbone network, reducing the loss of semantic features during the top-down fusion process. CFR is then employed to refine the features through suppressing background noise and non-salient regions. Experiments on five public datasets verify that our proposed model achieves satisfactory performance compared with state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Kunqian Yang and Caitou He}, doi = {10.1016/j.neucom.2025.129383}, journal = {Neurocomputing}, month = {3}, pages = {129383}, shortjournal = {Neurocomputing}, title = {Semantic-guided complementary fusion network for salient object detection}, volume = {622}, year = {2025}, }
• (2025). WAE: An evaluation metric for attribution-based XAI on time series forecasting. NEUCOM, 622, 129379. (www)

  @article{ , abstract = {Explainable Artificial Intelligence (XAI) sheds light on the decision-making ground of black-box models by offering explanations. However, the difficulty for humans to qualitatively evaluate XAI’s explanations on time series has led to a very limited development of XAI in the field of time series. Thus, helping humans to evaluate explanations on time series becomes a critical and challenging task. In this paper, we propose a metric to evaluate the explanations generated from attribution-based XAI method on time series forecasting problem, called Window-based Attribution Root Mean Squared Error (WAE). We suppose that more advanced XAI methods as well as more explainable datasets can achieve better performance on this metric. The experimental results show that SHAP and LEMNA, as improved methods of LIME, both of them have smaller WAE than LIME. In addition, we artificially designed explainable dataset, semi-explainable dataset, and weakly explainable dataset. They have sequentially increasing WAE under the same XAI method. The experimental outcomes indicate that the two hypotheses essentially stand, implying the feasibility of our proposed measure. Furthermore, we employed this metric on two real-world datasets to demonstrate its applicability, and the combination of TCN and SHAP yields an optimal explanation. Finally, we analyze the effect of different parameter values on WAE, focusing on two crucial parameters, the prediction step lag and the time window length.}, archive = {J_NEUCOM}, author = {Yueshan Chen and Sihai Zhang}, doi = {10.1016/j.neucom.2025.129379}, journal = {Neurocomputing}, month = {3}, pages = {129379}, shortjournal = {Neurocomputing}, title = {WAE: An evaluation metric for attribution-based XAI on time series forecasting}, volume = {622}, year = {2025}, }
• (2025). Adversarial detection by approximation of ensemble boundary. NEUCOM, 622, 129364. (www)

  @article{ , abstract = {Despite being effective in many application areas, Deep Neural Networks (DNNs) are vulnerable to being attacked. In object recognition, the attack takes the form of a small perturbation added to an image, that causes the DNN to misclassify, but to a human appears no different. Adversarial attacks lead to defences that are themselves subject to attack, and the attack/defence strategies provide important information about the properties of DNNs . In this paper, a novel method of detecting adversarial attacks is proposed for an ensemble of Deep Neural Networks (DNNs) solving two-class pattern recognition problems. The ensemble is combined using Walsh coefficients which are capable of approximating Boolean functions and thereby controlling the decision boundary complexity. The hypothesis in this paper is that decision boundaries with high curvature allow adversarial perturbations to be found, but change the curvature of the decision boundary, which is then approximated in a different way by Walsh coefficients compared to the clean images. Besides controlling boundary complexity, the coefficients also measure the correlation with class labels, which may aid in understanding the learning and transferability properties of DNNs . While the experiments here use images, the proposed approach of modelling two-class ensemble decision boundaries could in principle be applied to any application area.}, archive = {J_NEUCOM}, author = {Terry Windeatt}, doi = {10.1016/j.neucom.2025.129364}, journal = {Neurocomputing}, month = {3}, pages = {129364}, shortjournal = {Neurocomputing}, title = {Adversarial detection by approximation of ensemble boundary}, volume = {622}, year = {2025}, }
• (2025). VQA and visual reasoning: An overview of approaches, datasets, and future direction. NEUCOM, 622, 129345. (www)

  @article{ , abstract = {Visual question answering (VQA) is a problem that researchers in both computer vision and natural language processing are interested in studying. In VQA, a system is given an image and a question in natural language about that image. The VQA system is then expected to answer in natural language. To find the right answer, a VQA algorithm may need to use common sense to make sense of the information in the image and external knowledge. In this paper, we discuss some of the main ideas behind VQA systems and provide a comprehensive literature survey of the current state of the art in VQA and visual reasoning from four perspectives: problem definition and challenges, approaches, existing datasets, and evaluation matrices. We conclude our survey with a discussion and some potential future research directions in this area to generate new ideas and creative approaches to solving current problems and developing new applications.}, archive = {J_NEUCOM}, author = {Rufai Yusuf Zakari and Jim Wilson Owusu and Ke Qin and Hailin Wang and Zaharaddeen Karami Lawal and Tao He}, doi = {10.1016/j.neucom.2025.129345}, journal = {Neurocomputing}, month = {3}, pages = {129345}, shortjournal = {Neurocomputing}, title = {VQA and visual reasoning: An overview of approaches, datasets, and future direction}, volume = {622}, year = {2025}, }
• (2025). Optimal scheduling of shared autonomous electric vehicles with multi-agent reinforcement learning: A MAPPO-based approach. NEUCOM, 622, 129343. (www)

  @article{ , abstract = {The advent of shared autonomous electric vehicles (SAEVs) is expected to decarbonize the transport sector and enable large-scale mobility on demand. Considering the highly dynamic nature of operation environments, it is appealing to investigate the optimal schedule decisions, comprising order-matching, relocation, and charge/ discharge with the V2G technology. Recent literature has investigated these schedule decisions with deep reinforcement learning (DRL) independently, neglecting the collaborative interactions among multiple autonomous vehicles. To this end, the fully dynamic schedule task is integrally described as a partially observable Markov decision process (POMDP) under the constraints of time-dependent demand, spatial-temporal varying electricity price, available vehicles and chargers. Then, we design a robust and extensible framework that combines multi-agent proximal policy optimization (MAPPO) and a binary linear programming (BLP) model to solve the optimization problem, in which SAEVs are regarded as agents to make decisions according to their local observations. Additionally, observation normalization and action masking are utilized to improve training efficiency. Numerical experiments are conducted with New York City yellow taxi data to simulate the real-time SAEV operating environment. These results demonstrate that the proposed MAPPO approach is expected to enhance the operators' revenue while improving the customer service rate compared to benchmark algorithms.}, archive = {J_NEUCOM}, author = {Jingjing Tian and Hongfei Jia and Guanfeng Wang and Qiuyang Huang and Ruiyi Wu and Heyao Gao and Chao Liu}, doi = {10.1016/j.neucom.2025.129343}, journal = {Neurocomputing}, month = {3}, pages = {129343}, shortjournal = {Neurocomputing}, title = {Optimal scheduling of shared autonomous electric vehicles with multi-agent reinforcement learning: A MAPPO-based approach}, volume = {622}, year = {2025}, }
• (2025). Enhancing hexapod robot mobility on challenging terrains: Optimizing CPG-generated gait with reinforcement learning. NEUCOM, 622, 129328. (www)

  @article{ , abstract = {Hexapod robots, with more legs and freedoms, excel in stability and terrain adaptability over wheeled or quadruped robots. They are suited for tasks like geological exploration, outdoor monitoring, and disaster relief. Central Pattern Generators (CPG) enable hexapod robots to walk naturally without external control, but gait adaptability on complex terrains is limited. Deep Reinforcement Learning (DRL), though commonly used in robot motion control, faces more challenges for hexapod robots due to the high-dimensional continuous control problem and their complex dynamics, especially in complex terrains. This paper presents a novel method that employs DRL to enhance the gait output of CPG, thereby enhancing the terrain adaptability of hexapod robots. Experiments are carried out on three challenging terrains: rough terrains, slopes, and stairs. The results indicate that the presented method provides better learning efficiency and effectiveness than cases when DRL is used alone. It also shows that the presented method allows hexapod robots to traverse complex terrains with enhanced flexibility and efficacy in comparison with cases when only CPG is used.}, archive = {J_NEUCOM}, author = {Shichang Huang and Minhua Zheng and Zhongyu Hu and Peter Xiaoping Liu}, doi = {10.1016/j.neucom.2024.129328}, journal = {Neurocomputing}, month = {3}, pages = {129328}, shortjournal = {Neurocomputing}, title = {Enhancing hexapod robot mobility on challenging terrains: Optimizing CPG-generated gait with reinforcement learning}, volume = {622}, year = {2025}, }
• (2025). Few-shot aero-engine bearing fault diagnosis with denoising diffusion based data augmentation. NEUCOM, 622, 129327. (www)

  @article{ , abstract = {As a core component of aero-engines, the health condition of rolling bearings is crucial to the safety and stability of the aviation power system. In the fault diagnosis of aero-engine rolling bearings, the issue of insufficient fault data has long persisted. The training process of the widely introduced Generative Adversarial Networks (GANs) is difficult to control, and their generalization ability is relatively poor. Therefore, this paper proposes a data augmentation method based on denoising diffusion probabilistic model (DDPM) for aero-engine bearing fault diagnosis under few-shot. The proposed method is grounded in a well-defined probabilistic model and mathematical principles, allowing it to avoid instability and mode collapse during training. Specifically, we first propose using the synchro-squeezed wavelet transform (SST) to convert one-dimensional time-series signals into time-domain images as model input, addressing the insufficient feature extraction of traditional time–frequency analysis methods and clearly illustrating the fault characteristics of frequency variation with time. Next, we design a CRR-UNet based on residual connections to mitigate overfitting caused by insufficient data during the reverse denoising process of DDPM, thereby improving the quality of generated samples. Finally, a deep residual shrinkage network (DRSN) is employed to conduct fault diagnosis on the augmented fault dataset. The research results show that the samples generated by this method are highly similar to the original samples, outperforming existing data augmentation methods, and the fault diagnosis accuracy on the augmented dataset reaches 98%, significantly improving fault diagnosis performance.}, archive = {J_NEUCOM}, author = {Zuowei Ping and Dewen Wang and Yong Zhang and Bo Ding and Yaqiong Duan and Wei Zhou}, doi = {10.1016/j.neucom.2024.129327}, journal = {Neurocomputing}, month = {3}, pages = {129327}, shortjournal = {Neurocomputing}, title = {Few-shot aero-engine bearing fault diagnosis with denoising diffusion based data augmentation}, volume = {622}, year = {2025}, }
• (2025). Cardiac cavity segmentation review in the past decade: Methods and future perspectives. NEUCOM, 622, 129326. (www)

  @article{ , abstract = {Medical imaging technology has played a vital role in modern medicine and medical care. Cardiovascular imaging and computing technology are essential for diagnosing and treating diseases. Cardiac image segmentation is the foundation of cardiovascular imaging and computing technology. This article aims to review the mainstream technologies, datasets, evaluation indicators, and prospects of cardiac chamber segmentation in the past decade. The mainstream technologies for cardiac chamber segmentation are mainly divided into two categories: those based on traditional theories and those based on deep learning techniques. The cardiac chamber segmentation dataset is divided into three categories: ventricular, atrial, and whole heart segmentation dataset. Evaluation metrics are mainly divided into three categories: regional similarity metrics, contour similarity metrics, and clinical metrics. This review found that existing cardiac chamber segmentation methods have achieved good segmentation performance on specific data domains and have the potential for broad applicability in clinical settings, such as calculating ejection fraction and implementing catheter ablation procedures. However, fully applying heart chamber segmentation technology in clinical practice still poses challenges, such as domain adaptivity, interpretability, and spatiotemporal consistency issues that have not been fully resolved. These challenges are worthy of future research in cardiac chamber segmentation.}, archive = {J_NEUCOM}, author = {Feiyan Li and Weisheng Li and Yucheng Shu and Yidong Peng and Bin Xiao}, doi = {10.1016/j.neucom.2024.129326}, journal = {Neurocomputing}, month = {3}, pages = {129326}, shortjournal = {Neurocomputing}, title = {Cardiac cavity segmentation review in the past decade: Methods and future perspectives}, volume = {622}, year = {2025}, }
• (2025). BERT-DXLMA: Enhanced representation learning and generalization model for english text classification. NEUCOM, 622, 129325. (www)

  @article{ , abstract = {In recent years, the rapid development of artificial intelligence technology has significantly improved the performance of BERT-based deep learning models in text classification tasks. However, these models still have shortcomings in capturing the deep semantics of text, emerging text classification, and handling minority class samples in imbalanced data. To address these challenges, we proposed an innovative BERT-DXLMA hybrid model. This model enhances the semantic feature extraction and representation learning capabilities by integrating xLSTM, and uses semantic fusion technology to reduce the information loss caused by the increase in the number of network layers. In addition, we re-improved the loss function based on focal loss to improve the model’s attention to minority class samples and generalization ability. We conducted extensive experiments and in-depth analysis on six public datasets, and the results showed that BERT-DXLMA outperformed multiple baseline methods in terms of precision and overall accuracy, demonstrating its superiority in text classification tasks.}, archive = {J_NEUCOM}, author = {Xingliang Mao and Zhuhao Li and Qingxi Li and Shichao Zhang}, doi = {10.1016/j.neucom.2024.129325}, journal = {Neurocomputing}, month = {3}, pages = {129325}, shortjournal = {Neurocomputing}, title = {BERT-DXLMA: Enhanced representation learning and generalization model for english text classification}, volume = {622}, year = {2025}, }
• (2025). Distributed task allocation of fleet-level maintenance: Dealing with stochastic durations. NEUCOM, 622, 129324. (www)

  @article{ , abstract = {In scenarios where systems must execute a series of missions with limited breaks, the difficulty in assigning maintenance activities is of paramount importance. Generally, the efficiency of formation maintenance is dependent of maintenance planning as well as groups scheduling. However, few existing research takes into account the influence of maintenance providers scheduling, which results in inappropriate maintenance decisions. This paper proposes a selective maintenance optimization model for a single equipment system and a fleet-level maintenance task allocation model with random maintenance time. A distributed algorithm is introduced for the addressed fleet-level maintenance task allocation with stochastic maintenance durations. In comparison to those existing models where task durations are assumed to be deterministic, this approach utilizes a selective maintenance model in combination with a decentralized task allocation framework, guaranteeing successful completion of subsequent missions. On basis of a customized heuristic algorithm with an enhanced consensus-based bundle method, the proposed approach seeks the optimal solution to the allocation of maintenance tasks across a fleet by minimizing downtime while maximizing success probability. The effectiveness and validity of the established model is finally verified via a case study involving agricultural harvesters, showing its potential application in dynamic and resource-constrained environments.}, archive = {J_NEUCOM}, author = {Pengxiang Wang and Xiaojian Yi}, doi = {10.1016/j.neucom.2024.129324}, journal = {Neurocomputing}, month = {3}, pages = {129324}, shortjournal = {Neurocomputing}, title = {Distributed task allocation of fleet-level maintenance: Dealing with stochastic durations}, volume = {622}, year = {2025}, }
• (2025). Global multi-scale extraction and local mixed multi-head attention for facial expression recognition in the wild. NEUCOM, 622, 129323. (www)

  @article{ , abstract = {Facial expression recognition (FER) in the wild is a challenging computer vision task due to various uncontrollable factors such as facial occlusion and head pose variation. To address the above issues, we propose a global multi-scale extraction and local mixed multi-head attention network (GLMEA), which consists of a global multi-scale extraction module (GMEM), a local mixed multi-head attention module (LMAM), and an attention separation loss (ASL). Specifically, GMEM extracts global multi-scale features across the channel dimension using parallel multi-scale convolution, and then adaptively enhances discriminative features at different scales through a simple yet effective gating mechanism. This design achieves a richer global representation, reducing sensitivity to occlusion and pose variation. Subsequently, LMAM focuses on local salient features in a coarse-to-fine manner through multiple parallel mixed attention heads, enabling the model to learn a more comprehensive facial representation while preserving the integrity of the feature space. Finally, ASL reduces the consistency of attention maps learned by different mixed attention heads, mitigating attention overlap and encouraging a more diverse facial focus. Extensive experimental results demonstrate that our proposed GLMEA outperforms several current state-of-the-art methods on three challenging in-the-wild datasets, indicating superiority and robustness. In addition, considering the limited research on FER for elderly facial expressions and the potential applications of recognizing elderly emotional changes in intelligent diagnosis and treatment, we collected and labeled a new Chinese Elderly Facial Expression (CEFE) dataset to validate the potential of our method for recognizing elderly facial expressions.}, archive = {J_NEUCOM}, author = {Zixiang Fei and Bo Zhang and Wenju Zhou and Xia Li and Yukun Zhang and Minrui Fei}, doi = {10.1016/j.neucom.2024.129323}, journal = {Neurocomputing}, month = {3}, pages = {129323}, shortjournal = {Neurocomputing}, title = {Global multi-scale extraction and local mixed multi-head attention for facial expression recognition in the wild}, volume = {622}, year = {2025}, }
• (2025). Detection of fake bitrate videos based on high-frequency and deblocking filtering difference map features. NEUCOM, 622, 129320. (www)

  @article{ , abstract = {Video bitrate up-conversion is a video processing technique that improves the quality by increasing the bitrate, typically for attracting more viewers. Detection of fake bitrate videos is beneficial for copyright protection and aids in verifying the authenticity of the video content. In this work, a novel algorithm is proposed to detect fake bitrate videos. Theoretical and statistical analysis of the artifacts in compression and decompression domains of fake bitrate video is established, which provides a basis for feature modeling. According to the analysis, the Difference Map Feature ( F D M ) which consists of The High-frequency Component Difference map Feature ( F H C D ) in I-frames and Deblocking Filtering features Difference map Feature ( F D F D ) in P-frames is proposed. The videos undergo an additional compression operation to calculate the difference maps which are fed into a dual-branch hybrid deep-learning network to obtain GOP-wise detection results. Extensive experiments have demonstrated that our algorithm has state-of-the-art detection capabilities under different encoding parameter settings.}, archive = {J_NEUCOM}, author = {Yikun Ao and Tanfeng Sun and Qiang Xu}, doi = {10.1016/j.neucom.2024.129320}, journal = {Neurocomputing}, month = {3}, pages = {129320}, shortjournal = {Neurocomputing}, title = {Detection of fake bitrate videos based on high-frequency and deblocking filtering difference map features}, volume = {622}, year = {2025}, }
• (2025). M-GENE: Multiview genes expression network ensemble for bone metabolism-related gene classification. NEUCOM, 622, 129318. (www)

  @article{ , abstract = {In this paper, a novel multi-view gene expression network ensemble (M-GENE) pipeline has been proposed to predict correlations between genes and phenotypes using bulk RNA-seq gene expression data. The framework incorporates a preprocessing strategy that reshapes raw RNA-seq data into a structured format compatible with neural network training. To mitigate the effects of data imbalance, a refined application of SMOTE is employed, enhancing the uniformity of data distribution and addressing training bias. A multi-view domain alignment network further strengthens the model by identifying shared gene expression features under varying experimental conditions, enabling a comprehensive and accurate analysis. Extensive experiments validate M-GENE’s superior accuracy and robustness, surpassing contemporary models and offering significant potential for practical applications.}, archive = {J_NEUCOM}, author = {Keyi Yu and Weilong Tan and Jirong Ge and Xinyu Li and Yingbei Wang and Jingwen Huang and Xuan Chen and Shengqiang Li and Nianyin Zeng}, doi = {10.1016/j.neucom.2024.129318}, journal = {Neurocomputing}, month = {3}, pages = {129318}, shortjournal = {Neurocomputing}, title = {M-GENE: Multiview genes expression network ensemble for bone metabolism-related gene classification}, volume = {622}, year = {2025}, }
• (2025). Distributed set-membership filtering for complex networks with amplify-and-forward relays: A two-step ellipsoid method. NEUCOM, 622, 129317. (www)

  @article{ , abstract = {This work investigates distributed set-membership filtering for complex networks with amplify-and-forward relays, considering unknown but bounded noises and Lipschitz-type nonlinear dynamics. To address issues of communication delay and signal attenuation due to long-distance transmission between network nodes and remote filter, amplify-and-forward relays are introduced for the purpose of extending the communication range and enhancing signal strength. A distributed filter featuring a one-step prediction and update is developed to efficiently utilize neighbor information in complex networks. The gain of this distributed filter is obtained by invoking a two-step ellipsoidal method. Ellipsoidal constraints on the state and prediction error are calculated, and these values are subsequently used to derive ellipsoidal constraints on the filtering error. Then, the P i ( k ) -dependent constraints that guarantee the filtering error are achieved. An illustrative example is designed to demonstrate the validity of the proposed filtering scheme.}, archive = {J_NEUCOM}, author = {Yi Zhong and Tianyang Zhang and Zitian Li and Shenggui Bao and Yumei Zhou}, doi = {10.1016/j.neucom.2024.129317}, journal = {Neurocomputing}, month = {3}, pages = {129317}, shortjournal = {Neurocomputing}, title = {Distributed set-membership filtering for complex networks with amplify-and-forward relays: A two-step ellipsoid method}, volume = {622}, year = {2025}, }
• (2025). Let storytelling tell vivid stories: A multi-modal-agent-based unified storytelling framework. NEUCOM, 622, 129316. (www)

  @article{ , abstract = {Storytelling via image sequences demands the creation of compelling and vivid narratives that adhere to the visual content while maintaining engaging plot divergence. Previous works have incrementally refined the alignment of multiple modalities, yet often resulted in the generation of simplistic storylines for image sequences. In this study, we introduce the LLaMS framework, designed to generate multi-modal human-level stories characterized by expressiveness and consistency. Our approach involves the utilization of a multi-modal agent framework to elevate the expression of factual content and applies a textual reasoning architecture for the purpose of expressive story generation and prediction. Furthermore, we introduce the Story-Adapter module, tailored for long image sequence illustration, with a focus on maintaining prolonged story consistency as opposed to short-term object consistency. Extensive experiments are conducted to validate the superior performance of the proposed LLaMS with human evaluation. Evaluations demonstrate that LLaMS achieves state-of-the-art storytelling performance, with an 86% correlation and 100% consistency win rate compared to prior state-of-the-art methods. Additionally, we conduct ablation experiments to confirm the efficacy of the proposed multi-modal agent framework and Story-Adapter module. Our code is accessible at https://anonymous.4open.science/status/LLams-FF83 .}, archive = {J_NEUCOM}, author = {Rongsheng Zhang and Jiji Tang and Chuanqi Zang and Mingtao Pei and Wei Liang and Zeng Zhao and Zhou Zhao}, doi = {10.1016/j.neucom.2024.129316}, journal = {Neurocomputing}, month = {3}, pages = {129316}, shortjournal = {Neurocomputing}, title = {Let storytelling tell vivid stories: A multi-modal-agent-based unified storytelling framework}, volume = {622}, year = {2025}, }
• (2025). SEDA-EEG: A semi-supervised emotion recognition network with domain adaptation for cross-subject EEG analysis. NEUCOM, 622, 129315. (www)

  @article{ , abstract = {In this paper, a cross-subject emotion recognition network based on semi-supervised and domain adversarial learning with electroencephalogram (EEG) signals (SEDA-EEG) is proposed, which addresses the challenge of high EEG variability from different subjects in brain-computer interface (BCI) research. By employing the differential entropy features of EEG signal processed by the linear dynamic systems as input, influences of noises can be greatly reduced and the essential changes in EEG can be well reflected while capturing the implicit signal features over time. Simultaneously, a deep neural network is designed to match similarity relationships between samples and labels by capturing key patterns in the data, which improves the performance of recognizing EEG signals. Moreover, a domain transfer module based on domain adversarial loss and offline feature decomposition semi-supervised learning is proposed to enhance inter-domain knowledge generalization, which achieves cross-domain feature alignment and enables the model more adapt to the target data. Experimental results show that the proposed SEDA-EEG yields the state-of-the-art performance, which outperforms other advanced models by 12.90% on the SEED-IV dataset with stronger robustness, indicating the potential of applying into the EEG-oriented cross-subject emotion recognition.}, archive = {J_NEUCOM}, author = {Weilong Tan and Hongyi Zhang and Yingbei Wang and Weimin Wen and Liang Chen and Han Li and Xingen Gao and Nianyin Zeng}, doi = {10.1016/j.neucom.2024.129315}, journal = {Neurocomputing}, month = {3}, pages = {129315}, shortjournal = {Neurocomputing}, title = {SEDA-EEG: A semi-supervised emotion recognition network with domain adaptation for cross-subject EEG analysis}, volume = {622}, year = {2025}, }
• (2025). Geometric edge convolution for rigid transformation invariant features in 3D point clouds. NEUCOM, 622, 129313. (www)

  @article{ , abstract = {Extracting rigid transformation invariant features is still a challenge on 3D point clouds because rigid transformation changes the point coordinates, and relying on the point coordinates, most existing deep learning models fail. This paper addresses this challenge by proposing geometric edge convolution (GEConv). To facilitate rigid transformation invariant feature extraction, GEConv uses geometric features that are invariant to rigid transformations to construct a local graph and use shared multi-layer perceptron (MLP) on the graph edges to extract deep features. By leveraging these invariant geometric features to build the graph, GEConv effectively ensures the extraction of deep features that are also invariant to rigid transformations. By stacking GEConv, we construct a deep model called GEConvNet that hierarchically extracts different level rigid transformation invariant features for various point cloud tasks. To prove the effectiveness of GEConv, we provide experiments on four point cloud processing tasks; classification, parts segmentation, semantic segmentation, and point set registration. In the classification and segmentation tasks, GEConvNet performs better on robustness to rotation compared to existing models. Furthermore, our feature-matching-based approach for point set registration demonstrates strong performance, providing further evidence of GEConv’s ability to handle rotation and translation effectively. Our code is available at http://www.github.com/saifabel/GEConv .}, archive = {J_NEUCOM}, author = {Saifullahi Aminu Bello and Saghir Alfasly and Jiawei Mao and Jian Lu and Lin Li and Chen Xu and Yuru Zou}, doi = {10.1016/j.neucom.2024.129313}, journal = {Neurocomputing}, month = {3}, pages = {129313}, shortjournal = {Neurocomputing}, title = {Geometric edge convolution for rigid transformation invariant features in 3D point clouds}, volume = {622}, year = {2025}, }
• (2025). Temporal difference learning with multi-step returns for intelligent optimal control of dynamic systems. NEUCOM, 622, 129311. (www)

  @article{ , abstract = {In this article, the adaptive neural control algorithm based on the λ -return mechanism [ANC( λ )] is developed under the temporal difference learning framework to address the optimal control problems of complex dynamic systems. First, in the light of the classical value iteration (VI) framework, a novel iteration mechanism is established, including the general value function, control policy, and the multi-step return value function, where the multi-step value function is constructed through multi-step returns and temporal difference with long-term prediction factor λ [TD( λ )]. Meanwhile, the generalized VI idea is incorporated into the specific implementation process of the ANC( λ ) algorithm. The general iterative value function and the λ -return iterative value function are respectively initialized from zero and a larger positive definite function. Subsequently, the convergence of the ANC( λ ) algorithm is rigorously analyzed, that is, the general iterative value function and the λ -return iterative value function converge to the optimal value in a monotonically increasing and monotonically decreasing manner, respectively. Finally, the effectiveness and progressiveness of the ANC( λ ) algorithm are demonstrated by conducting the power system application and the nonlinear system simulation.}, archive = {J_NEUCOM}, author = {Peng Xin and Ding Wang and Ao Liu and Junfei Qiao}, doi = {10.1016/j.neucom.2024.129311}, journal = {Neurocomputing}, month = {3}, pages = {129311}, shortjournal = {Neurocomputing}, title = {Temporal difference learning with multi-step returns for intelligent optimal control of dynamic systems}, volume = {622}, year = {2025}, }
• (2025). Analysis of coupled distributed stochastic approximation for misspecified optimization. NEUCOM, 622, 129310. (www)

  @article{ , abstract = {We consider an n agents distributed optimization problem with misspecified information characterized in a parametric sense, where the unknown parameter can be solved by a distinct distributed parameter learning problem. Though each agent only has access to its local parameter learning and computational problem, they mean to collaboratively minimize the average of their local cost functions through local communications over a connected network. Such problem is obviated in some distributed computing scenarios but really significant. To address the special optimization problem, we propose a coupled distributed stochastic approximation algorithm, in which every agent updates the current estimate of its unknown parameter and decision variable by stochastic approximation method; then averages the estimates and decision variables of its neighbors over network in consensus protocol. Our interest lies in the fundamental theoretical convergence analysis of this algorithm. We rigorously capture the factors that affect the algorithm performance, and prove that the mean-squared error of the decision variable is bounded by O ( 1 n k ) + O 1 n ( 1 − ρ w ) 1 k 1 . 5 + O ( 1 ( 1 − ρ w ) 2 ) 1 k 2 , where k is the iteration count and ( 1 − ρ w ) is the spectral gap of the network weighted adjacency matrix. It reveals that the network connectivity characterized by ( 1 − ρ w ) only influences the high order of convergence rate, whereas the domain rate still acts the same as the centralized algorithm. In addition, we analyze that the transient iteration needed for reaching its dominant rate O ( 1 n k ) is O ( n ( 1 − ρ w ) 2 ) . This can provide valuable insight for distributed algorithm design. Numerical experiments are carried out to demonstrate the theoretical results by taking different CPUs as agents, which is more applicable to real-world distributed scenarios.}, archive = {J_NEUCOM}, author = {Yaqun Yang and Jinlong Lei}, doi = {10.1016/j.neucom.2024.129310}, journal = {Neurocomputing}, month = {3}, pages = {129310}, shortjournal = {Neurocomputing}, title = {Analysis of coupled distributed stochastic approximation for misspecified optimization}, volume = {622}, year = {2025}, }
• (2025). Mamba-driven hierarchical temporal multimodal alignment for referring video object segmentation. NEUCOM, 622, 129308. (www)

  @article{ , abstract = {The objective of Referring Video Object Segmentation (RVOS) is to segment target objects within video sequences described by natural language expressions. Previous approaches often overlook the latent temporal information in the language modality, and they fail to align this information with the temporal dimension of the visual modality, leading to a lack of comprehensive cross-modal understanding between visual and linguistic spaces. Additionally, the absence of pixel-level inter-frame information aggregation hinders fine-grained space–time consistency modeling of the target object, thereby reducing segmentation accuracy. To address these issues in RVOS, we propose a novel mamba-driven hierarchical temporal multimodal alignment framework, abbreviated as MHTMA. Specifically, we introduce a hierarchical multimodal interaction and temporal integration module designed to enhance the temporal correlation of multimodal features, achieving cross-modal temporal alignment of visual and linguistic features to enable more accurate decoding of the target object. Moreover, we develop a mamba-based temporal representation learning module that learns the spatio-temporal features of the referred object at the pixel level, resulting in more precise generation of spatio-temporally consistent target object masks. Extensive experimental results on four benchmarks demonstrate that our MHTMA outperforms state-of-the-art methods, validating the effectiveness of the proposed modules.}, archive = {J_NEUCOM}, author = {Le Liang and Lefei Zhang}, doi = {10.1016/j.neucom.2024.129308}, journal = {Neurocomputing}, month = {3}, pages = {129308}, shortjournal = {Neurocomputing}, title = {Mamba-driven hierarchical temporal multimodal alignment for referring video object segmentation}, volume = {622}, year = {2025}, }
• (2025). Global distilling framework with cognitive gravitation for multimodal emotion recognition. NEUCOM, 622, 129306. (www)

  @article{ , abstract = {Multimodal emotion recognition (MER) which aims to recognize human emotions from multimodal data, is significant for many research tasks and applications. The inherent heterogeneity across different modalities creates information redundancy and distribution gap that pose significant challenges to MER. Previous feature-decoupling methods have alleviated these challenges and achieved promising achievements by decomposing each modality into modality-shared/ -specific subspaces for better representation learning. However, existing feature-decoupling methods have not reduced the redundant information within modalities effectively and the discriminability of modality-specific features obtained is insufficient. Moreover, existing feature-decoupling methods neglect the distribution gap between the decoupled modality-shared and modality-specific features, resulting in suboptimal performance. This paper proposes a novel global distilling framework (GDF) with cognitive gravitation to mitigate these limitations. Specifically, the feature of each modality is decoupled into modality-shared and modality-specific features using a feature decoupling module. Then a novel modal cognitive gravitation loss (MCGL) is used to reduce redundant information and obtain the discriminative features of each modality. Finally, a global graph distillation unit (GGD-Unit) is proposed to mitigate the distribution gap between the decoupled two types of features for effective feature fusion. Comprehensive experimental results on the CMU-MOSI and CMU-MOSEI datasets show that the proposed model achieves promising performance.}, archive = {J_NEUCOM}, author = {Huihui Li and Haoyang Zhong and Chunlin Xu and Xiaoyong Liu and Guihua Wen and Lianqi Liu}, doi = {10.1016/j.neucom.2024.129306}, journal = {Neurocomputing}, month = {3}, pages = {129306}, shortjournal = {Neurocomputing}, title = {Global distilling framework with cognitive gravitation for multimodal emotion recognition}, volume = {622}, year = {2025}, }
• (2025). Dlordinal: A python package for deep ordinal classification. NEUCOM, 622, 129305. (www)

  @article{ , abstract = {dlordinal is a new Python library that unifies many recent deep ordinal classification methodologies available in the literature. Developed using PyTorch as underlying framework, it implements the top performing state-of-the-art deep learning techniques for ordinal classification problems. Ordinal approaches are designed to leverage the ordering information present in the target variable. Specifically, it includes loss functions, various output layers, dropout techniques, soft labelling methodologies, and other classification strategies, all of which are appropriately designed to incorporate the ordinal information. Furthermore, as the performance metrics to assess novel proposals in ordinal classification depend on the distance between target and predicted classes in the ordinal scale, suitable ordinal evaluation metrics are also included. dlordinal is distributed under the BSD-3-Clause license and is available at https://github.com/ayrna/dlordinal .}, archive = {J_NEUCOM}, author = {Francisco Bérchez-Moreno and Rafael Ayllón-Gavilán and Víctor M. Vargas and David Guijo-Rubio and César Hervás-Martínez and Juan C. Fernández and Pedro A. Gutiérrez}, doi = {10.1016/j.neucom.2024.129305}, journal = {Neurocomputing}, month = {3}, pages = {129305}, shortjournal = {Neurocomputing}, title = {Dlordinal: A python package for deep ordinal classification}, volume = {622}, year = {2025}, }
• (2025). H∞ state estimation for multiplex networks with randomly packet loss and sensor saturations. NEUCOM, 622, 129304. (www)

  @article{ , abstract = {The state information of multiplex networks (MNs) is a prerequisite for research in various fields, such as network dynamics, synchronization control, topology identification, fault diagnosis, and more. In the context of MNs, it is common to encounter sensor saturations and random packet loss while exchanging information. The state estimation of the networks is greatly influenced by both phenomena. Therefore, engaging in the examination of state estimation for MNs within scenarios characterized by sensor saturations and randomly occurring packet losses represents a considerable challenge. To tackle this challenge, we propose the problem of H ∞ state estimation problem in MNs with sensor saturations and randomly occurring packet loss. A suite of H ∞ state estimators is devised based on the available measurement outputs to guarantee that the estimation error exhibits exponential stability while simultaneously meeting the H ∞ performance requirements. To comprehensively describe the random phenomena of sensor saturations, data loss, and normal operation encountered in MNs, this paper employs the Kronecker delta function to set up a measurement model that effectively represents various combinations of random events by adjusting their respective probabilities. The sought-after estimator gains are subsequently acquired by resolving matrix inequalities through the application of semidefinite programming techniques. Subsequently, a numerical illustration complemented by simulations is presented to corroborate the efficacy and viability of the proposed methodology.}, archive = {J_NEUCOM}, author = {Hanqi Shu and Zufan Zhang}, doi = {10.1016/j.neucom.2024.129304}, journal = {Neurocomputing}, month = {3}, pages = {129304}, shortjournal = {Neurocomputing}, title = {H∞ state estimation for multiplex networks with randomly packet loss and sensor saturations}, volume = {622}, year = {2025}, }
• (2025). BoostSF-SHAP: Gradient boosting-based software for protein–ligand binding affinity prediction with explanations. NEUCOM, 622, 129303. (www)

  @article{ , abstract = {Machine learning-based (ML-based) scoring functions (SFs) for protein–ligand binding affinity prediction have exhibited remarkable performance in the field of structure-based drug discovery. However, little attention has been given to the interpretability of these SFs. In this study, we propose a software called BoostSF-SHAP for protein–ligand binding affinity prediction. Specifically, we employed gradient boosting decision trees (GBDTs) to construct the ML-based SF. Forty-one intermolecular interaction features were used as the input of this SF. Notably, the proposed software can provide local and global explanations for the SF by using the SHapley Additive exPlanations (SHAP) approach. This paper presents a description of the architecture, functionalities, and implementation details of the proposed software. An assessment and illustrative examples of how to use this software are also provided. BoostSF-SHAP is written in Python and available on GitHub under the Apache License.}, archive = {J_NEUCOM}, author = {Xingqian Chen and Shuangbao Song and Zhenyu Song and Shuangyu Song and Junkai Ji}, doi = {10.1016/j.neucom.2024.129303}, journal = {Neurocomputing}, month = {3}, pages = {129303}, shortjournal = {Neurocomputing}, title = {BoostSF-SHAP: Gradient boosting-based software for protein–ligand binding affinity prediction with explanations}, volume = {622}, year = {2025}, }
• (2025). Observer-based edge-consensus for fractional-order uncertain nodal networks with positive constraint. NEUCOM, 622, 129302. (www)

  @article{ , abstract = {The paper focuses on investigating the consensus problem of edges in nodal networks whose edges are depicted by fractional-order uncertain dynamics with positive constraint. An observer-based distributed edge-consensus protocol is built. Based on fractional Laplace transform, positive system theory, and graph theory, some sufficient criteria are derived for the positive edge-consensus of fractional-order uncertain nodal networks with the order belonging to ( 0 , 1 ] under a directed graph. Furthermore, a sufficient edge-consensus criterion for the system in an undirected graph, only relying on the count of nodes and edges, is derived. What is more, a necessary and sufficient criterion is given for the positive edge-consensus of nodal networks with certain fractional-order dynamics. Finally, a simulation example is provided to demonstrate the effectiveness of results.}, archive = {J_NEUCOM}, author = {Yanyan Ye and Yanfang Rong and Yaping Sun and Housheng Su and Viacheslav P. Shkodyrev}, doi = {10.1016/j.neucom.2024.129302}, journal = {Neurocomputing}, month = {3}, pages = {129302}, shortjournal = {Neurocomputing}, title = {Observer-based edge-consensus for fractional-order uncertain nodal networks with positive constraint}, volume = {622}, year = {2025}, }
• (2025). New sufficient conditions on the global exponential stability of delayed inertial neural networks. NEUCOM, 622, 129301. (www)

  @article{ , abstract = {In this paper, we utilize the characteristics method to deduce several novel sufficient conditions for the global exponential stability of delayed inertial neural networks (DINNs). The proposed criteria are presented as a set of linear scalar inequalities, which notably do not entail the use of either reduced order method or Lyapunov-Krasovskii functionals (LKFs), distinguishing them from existing results and offering straightforward solvability. Lastly, we substantiate the analytical outcomes through three numerical examples accompanied by respective simulations.}, archive = {J_NEUCOM}, author = {Wentao Wang and Wei Zeng and Wei Chen}, doi = {10.1016/j.neucom.2024.129301}, journal = {Neurocomputing}, month = {3}, pages = {129301}, shortjournal = {Neurocomputing}, title = {New sufficient conditions on the global exponential stability of delayed inertial neural networks}, volume = {622}, year = {2025}, }
• (2025). Neural observer-based formation for multi-UAVs against deception and desired trajectory attacks. NEUCOM, 622, 129297. (www)

  @article{ , abstract = {This paper investigates secure formation control for multi-UAV systems in the presence of deception and desired trajectory attacks. It is crucial to note that previous papers studying the secure formation control problem for multi-UAV systems usually apply to cases where the multi-UAV systems are subject to sensor attacks and actuator attacks. In this paper, we propose a formation control scheme to address both desired trajectory and deception attacks, incorporating attack modeling, estimation, and compensation at its core. First, dynamic models are constructed for multi-UAV systems and cyber attacks, respectively. In particular, desired trajectory and deception attacks are mapped to control input channels during modeling. Subsequently, an adaptive neural network observer is introduced to reconstruct desired trajectory attack signals. The online updating of neural network weights avoids the need for manual parameter selection. Next, we propose an event-triggered secure formation controller with an attack compensation approach aimed at reducing transmission resources under an event-triggered mechanism. The uniform ultimate boundedness of the system is derived. Finally, the formation tracking protocol is substantiated and validated by providing simulation results.}, archive = {J_NEUCOM}, author = {Kunpeng Pan and Feisheng Yang and Yang Lyu and Mingyue Ji and Quan Pan}, doi = {10.1016/j.neucom.2024.129297}, journal = {Neurocomputing}, month = {3}, pages = {129297}, shortjournal = {Neurocomputing}, title = {Neural observer-based formation for multi-UAVs against deception and desired trajectory attacks}, volume = {622}, year = {2025}, }
• (2025). A new generative adversarial networks-based fault diagnosis framework: Learning a mapping to estimate fault. NEUCOM, 622, 129288. (www)

  @article{ , abstract = {Thanks to the strong generalization and generation capabilities, Generative Adversarial Networks (GANs) have been developed for fault diagnosis schemes in recent years. Most of them are constructed as data enhancers for generating few-shot faulty samples and serve for the fault classification task. However, classifier-based fault diagnosis frameworks have difficulty in recognizing new fault classes and estimating fault magnitudes. In order to exploit the potential of GAN for fault detection and estimation, this paper proposes a Fault-estimable AutoEncoder-GAN (FAE-GAN), which uses a new framework to learn the fault elimination mapping. In FAE-GAN, faulty data are first expanded by sampling from non-normal data domains. Then, pairs of normal and faulty samples are fed into the FAE-GAN for mapping training, aiming to learn the conditional distribution from faulty to normal. In this way, the fault contained in the process data can be eliminated adaptively, thereby achieving the fault estimation purpose. Finally, a numerical example and a chemical simulation verify the proposed approach.}, archive = {J_NEUCOM}, author = {Ling Li and Zhuofu Pan and Yucheng Ma and Zhiwen Chen and Jie Chen and Yalin Wang}, doi = {10.1016/j.neucom.2024.129288}, journal = {Neurocomputing}, month = {3}, pages = {129288}, shortjournal = {Neurocomputing}, title = {A new generative adversarial networks-based fault diagnosis framework: Learning a mapping to estimate fault}, volume = {622}, year = {2025}, }
• (2025). Knowledge fusion distillation and gradient-based data distillation for class-incremental learning. NEUCOM, 622, 129286. (www)

  @article{ , abstract = {Deep neural networks, despite their exceptional performance on individual tasks, face the challenge of catastrophic forgetting when incrementally learning new scenarios. Existing class-incremental learning methods have shown promising results by rehearsing past samples from auxiliary memory. However, preserving only a limited subset of previous samples fails to capture the complete distribution, leading to a decline in memory quality as new data is added. Additionally, issues like exemplar distribution collapse and task-recency bias impede the effective transfer of knowledge from teacher to student during distillation. To address these problems, we propose a novel replay-based learning framework called K nowledge fusion D istillation and gradient-based D ata D istillation (K3D). K3D simultaneously optimizes both the classifier and exemplars during the learning phase. By parameterizing exemplars, we enable their optimization through gradient-based data distillation, enhancing their representativeness instead of discarding them, as done in other replay strategies. Furthermore, we enhance the classifier using knowledge fusion distillation, ensuring that decision boundaries are maintained across all classes. Extensive experiments on the CIFAR100 and ImageNet100 benchmarks show that synthetic exemplars optimized by K3D are more representative than selected ones. By combining optimizable exemplars and knowledge fusion distillation, K3D outperforms several state-of-the-art methods and effectively mitigates catastrophic forgetting in class-incremental learning scenarios.}, archive = {J_NEUCOM}, author = {Lin Xiong and Xin Guan and Hailing Xiong and Kangwen Zhu and Fuqing Zhang}, doi = {10.1016/j.neucom.2024.129286}, journal = {Neurocomputing}, month = {3}, pages = {129286}, shortjournal = {Neurocomputing}, title = {Knowledge fusion distillation and gradient-based data distillation for class-incremental learning}, volume = {622}, year = {2025}, }
• (2025). MCFN: Multi-scale crossover feed-forward network for high performance watermarking. NEUCOM, 622, 129282. (www)

  @article{ , abstract = {In recent years, deep learning-based watermarking for digital images has seen substantial progress, particularly in embedding secret messages in an invisible manner and extracting them after attacks. However, existing methods face challenges in effectively integrating spatial and frequency domain features, which limits overall performance. To address this, we propose the Multi-scale Crossover Feed-forward Network ( MCFN ), an innovative solution explicitly designed to process features of secret messages and cover images in a robust watermarking framework. The MCFN features a crossover structure that facilitates global communication between single-dimensional and double-dimensional features, enhancing the capacity to simultaneously capture crucial characteristics from both secret messages and cover images. A key element of our design is the Discrete Wavelet Transform (DWT) convolution block, which enables multi-scale feature extraction in the frequency domain, overcoming the limitations of conventional watermarking networks. We also incorporate perceptual loss to ensure high visual quality of the watermarked image and introduce a Bi-Directional Attack Simulation Layer (BD-ASL) to improve resilience against various attacks during training. Extensive experimental results validate that our framework achieves superior capacity, robustness, and imperceptibility, significantly outperforming current state-of-the-art methods in the deep learning-based watermarking field.}, archive = {J_NEUCOM}, author = {Yufeng Wu and Baowei Wang and Guiling Wang and Xin Liao}, doi = {10.1016/j.neucom.2024.129282}, journal = {Neurocomputing}, month = {3}, pages = {129282}, shortjournal = {Neurocomputing}, title = {MCFN: Multi-scale crossover feed-forward network for high performance watermarking}, volume = {622}, year = {2025}, }
• (2025). A region based salient stacking optimized detector (ReSOD) for an effective anomaly detection and video tracking in surveillance systems. NEUCOM, 622, 129281. (www)

  @article{ , abstract = {Identifying unusual occurrences in videos is among one of the most frequently addressed topics in the field of computer vision. Because of its personal or context-specific definitions, it is regarded as a challenging endeavor in the field of video analysis. To deal with the identification of anomalies issues, multiple approaches have been put forth. The use of deep learning and customized methods are among such methods. A large number of investigators have attempted to identify the most effective approach for efficiently identifying anomalies in video streams with a relatively low risk of false alarms. For use in video surveillance applications, this study presents a unique and lightweight video tracking technology called Region based Salient Stacking Optimized Detector (ReSOD). This framework uses a Correlation based Moving Position with Region Tracking Model (CoMP-RT) to perform an efficient and precise anomaly identification. In addition, motion prediction is carried out using a Salient Point based Stacking Pixel Hypothesis Model (S 2 P 2 Hypo) technique to provide safety and control. Furthermore, a unique optimization method called the hybrid Forensic Choatic Kernel Search Optimizer (FCKSO) is presented in order to estimate the optimal value and enhance the motion prediction process. Background subtraction is then carried out using the optimal value that is acquired. Furthermore, system analysis and assessment are being conducted using some of the most recent and well-known video surveillance datasets, including the mall dataset, UCSD, UMN, ShanghaiTech dataset, and violent flow dataset. By integrating CoMP-RT, S 2 P 2 Hypo, and FCKO techniques, the overall anomaly detection efficiency is enhanced to 99 % with high accuracy of 99.4 % and error rate of 1.2 %.}, archive = {J_NEUCOM}, author = {Areej Alasiry and Mohammed Qayyum}, doi = {10.1016/j.neucom.2024.129281}, journal = {Neurocomputing}, month = {3}, pages = {129281}, shortjournal = {Neurocomputing}, title = {A region based salient stacking optimized detector (ReSOD) for an effective anomaly detection and video tracking in surveillance systems}, volume = {622}, year = {2025}, }
• (2025). Cross-domain data-driven reinforcement learning for IGSO satellite coverage optimization. NEUCOM, 622, 129278. (www)

  @article{ , abstract = {The beam coverage area of Inclined GeoSynchronous Orbit (IGSO) satellites suffers from a diurnal north-south shift caused by the orbital inclination. Thus, it is crucial to adjust the satellite’s attitude for stabilizing the beam coverage area and enhancing the communication signal quality. However, traditional optimization methods fail to optimize attitude due to the intricate orbital environment and the time-consuming simulation solving process. In this paper, we propose a Cross-Domain Data-Driven Reinforcement Learning (C3DRL) algorithm for IGSO satellite coverage optimization, which utilizes offline expert data to assist agent learning and exploration during interactions within the online environment. To overcome the dynamic differences between online and offline environments, a practical cross-domain adaptation approach is proposed. This approach distinguishes dynamics changes and penalizes agents for learning transitions with significant dynamics differences. By addressing the issue of action value function overestimation, we effectively utilize offline data to drive agent learning process. The experimental results show that C3DRL effectively stabilizes the coverage area compared with other algorithms. In addition, several ablation experiments are conducted to demonstrate the effectiveness of the cross-domain adaptation method.}, archive = {J_NEUCOM}, author = {Dongyu Xu and Yifan Bo and Biao Leng}, doi = {10.1016/j.neucom.2024.129278}, journal = {Neurocomputing}, month = {3}, pages = {129278}, shortjournal = {Neurocomputing}, title = {Cross-domain data-driven reinforcement learning for IGSO satellite coverage optimization}, volume = {622}, year = {2025}, }
• (2025). A support vector machine formulation for linear and kernel discriminant analysis. NEUCOM, 622, 129273. (www)

  @article{ , abstract = {Originally formulated as a trace quotient criterion, linear discriminant analysis (LDA) and kernel discriminant analysis (KDA) have been shown to be equivalent to a multivariate regression problem. In this paper, considering a K-class problem ( K ≥ 2 ), we prove that LDA as well KDA are equivalent to solve independently K binary SVM classifiers with specific constraints. The key point of this reformulation is the definition of a new class indicator matrix which preserves the equivalence with the LDA’ s solution. Our proposition has several interests: first, we introduce naturally sparsity in the model using the concept of support vectors. Secondly, we improve robustness against outliers by selecting an adequate loss function. Thirdly, empirical results reveal that the proposed approach is less sensitive to the class imbalance problem than standard SVM. Several experiments on synthetic and real data sets show the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {F. Dufrenois and K. Jbilou}, doi = {10.1016/j.neucom.2024.129273}, journal = {Neurocomputing}, month = {3}, pages = {129273}, shortjournal = {Neurocomputing}, title = {A support vector machine formulation for linear and kernel discriminant analysis}, volume = {622}, year = {2025}, }
• (2025). Large language model-based evolutionary optimizer: Reasoning with elitism. NEUCOM, 622, 129272. (www)

  @article{ , abstract = {Large Language Models (LLMs) have demonstrated remarkable reasoning abilities, prompting interest in their application as black-box optimizers. This paper asserts that LLMs possess the capability for zero-shot optimization across diverse scenarios, including multi-objective and high-dimensional problems. We introduce a novel population-based method for numerical optimization using LLMs called Large Language-Model-Based Evolutionary Optimizer (LEO). Our hypothesis is supported through numerical examples, spanning benchmark and industrial engineering problems such as supersonic nozzle shape optimization, heat transfer, and windfarm layout optimization. We compare our method to several gradient-based and gradient-free optimization approaches. While LLMs yield comparable results to state-of-the-art methods, their imaginative nature and propensity to hallucinate demand careful handling. We provide practical guidelines for obtaining reliable answers from LLMs and discuss method limitations and potential research directions.}, archive = {J_NEUCOM}, author = {Shuvayan Brahmachary and Subodh M. Joshi and Aniruddha Panda and Kaushik Koneripalli and Arun Kumar Sagotra and Harshil Patel and Ankush Sharma and Ameya D. Jagtap and Kaushic Kalyanaraman}, doi = {10.1016/j.neucom.2024.129272}, journal = {Neurocomputing}, month = {3}, pages = {129272}, shortjournal = {Neurocomputing}, title = {Large language model-based evolutionary optimizer: Reasoning with elitism}, volume = {622}, year = {2025}, }
• (2025). Generalization bounds for pairwise learning with the huber loss. NEUCOM, 622, 129265. (www)

  @article{ , abstract = {Pairwise learning is a machine learning and pattern recognition technique that focuses on learning and making predictions by comparing pairs of data points or instances. It has drawn particular attention over the last decades in machine learning. The performance of pairwise learning is investigated in this paper. The main contribution is to conduct refined learning theory analysis based on the Huber loss function. It not only obtains improved rates of learning in the present of heavy-tailed noise, but also verifies the theoretical results by numerical experiments, which means our work have potential usages in reality.}, archive = {J_NEUCOM}, author = {Shouyou Huang and Zhiyi Zeng and Siying Jiang}, doi = {10.1016/j.neucom.2024.129265}, journal = {Neurocomputing}, month = {3}, pages = {129265}, shortjournal = {Neurocomputing}, title = {Generalization bounds for pairwise learning with the huber loss}, volume = {622}, year = {2025}, }
• (2025). Classifying infinite-dimensional data with unified basis functions: An effective machine learning approach. NEUCOM, 622, 129245. (www)

  @article{ , abstract = {Many real-world applications involve functional data, where each sample is a continuous function rather than a finite-dimensional vector or matrix. This infinite-dimensional nature poses significant challenges for classification tasks. Despite the progress in functional data analysis, existing methods and machine learning algorithms often struggle with the complexities of infinite-dimensional spaces. This paper introduces a novel approach that aims to simplify and improve the accuracy of infinite-dimensional classification tasks within functional data analysis. By addressing the algorithmic and computational complexities of existing high-dimensional classification methods, our approach provides an effective solution for the non-linear classification of infinite-dimensional functional data. The primary innovation lies in identifying a set of unified basis functions to accurately represent all observations originally defined in infinite-dimensional space. Subsequently, these observed functions in an infinite-dimensional space can be represented by their parameter vectors in a finite-dimensional space, and the original functional classification problem is converted to the conventional classification problem in a vector space (parameters) to solve, which leads to the simplicity of the proposed approach. We present the detailed design and implementation of this approach, including an effective framework for identifying functional data representations and B-spline knot placements. This approach not only facilitates the transformation between infinite-dimensional function spaces and finite-dimensional representative parameter spaces but also demonstrates significant accuracy improvements across various datasets in this field, as evidenced by empirical findings and comparative experimental results.}, archive = {J_NEUCOM}, author = {Jiayu Shang and Xiao-Jun Zeng}, doi = {10.1016/j.neucom.2024.129245}, journal = {Neurocomputing}, month = {3}, pages = {129245}, shortjournal = {Neurocomputing}, title = {Classifying infinite-dimensional data with unified basis functions: An effective machine learning approach}, volume = {622}, year = {2025}, }
• (2025). Personalized dual transformer network for sequential recommendation. NEUCOM, 622, 129244. (www)

  @article{ , abstract = {Sequential Recommendation (SR) seeks to anticipate the item that users will interact with at the next moment, utilizing their historical sequences of interactions. Its core task is to mine users’ interests. Most existing transformer-based models focus on extracting user’s interests by leveraging local context information from individual interaction sequences, ignoring the user’s personalized characteristics. Furthermore, when the sequence length is significantly shorter than the specified threshold, these transformer-based models encounter challenges associated with the cold-start issue. To tackle the mentioned issues, we propose a novel transformer-based sequential model, named Personalized Dual Transformer Network (PDTN), that extends the length of the user’s sequences by leveraging the user’s historical behaviors to search for relevant items. Specifically, a personalized feature extraction module is proposed to extract characteristics of both the user and the user’s historical behaviors. Then, a Dual Transformer is designed to retrieve a user’s relevant historical behaviors, which can increase the length of the user’s historical interaction sequences and fully extract the user’s interests. Furthermore, Stochastic Shared Embeddings (SSE) regularization is developed into a Transformer-based sequential model to reduce overfitting and errors in the processing of training. Extensive experiments are conducted to compare PDTN with current methods on four publicly available real-world datasets for prediction tasks. The results consistently demonstrate that PDTN outperforms advanced sequential recommendation methods. The source code is publicly available at https://github.com/QinLab-WFU/PDTN .}, archive = {J_NEUCOM}, author = {Meiling Ge and Chengduan Wang and Xueyang Qin and Jiangyan Dai and Lei Huang and Qibing Qin and Wenfeng Zhang}, doi = {10.1016/j.neucom.2024.129244}, journal = {Neurocomputing}, month = {3}, pages = {129244}, shortjournal = {Neurocomputing}, title = {Personalized dual transformer network for sequential recommendation}, volume = {622}, year = {2025}, }
• (2025). Application domains of aspect and sentiment classification techniques: A survey. NEUCOM, 622, 129237. (www)

  @article{ , abstract = {Recently, social media has changed the way, information is produced, transferred and consumed. User-generated content in the form of posts, blogs, comments, feedback, and reviews, has established a new connection between the producers and users of information. Tracking such content has enabled businesses to collect customer’s feedback to provide better services. The abundance of information from diverse sources helps users tap into the wisdom of crowds and aid in making more informed decisions. This raised the question of overcoming information overload and providing a rich and coherent user experience. This question has opened a rich venue for researchers on how to analyze such a huge amount of customer feedback to get actionable insights. A lot of research has been conducted in this area, which largely depends on opinion mining, sentiment analysis and text mining algorithms to interpret and make sense of large amounts of textual data. The proposed survey aims to identify application domains. Exploring each application domain to determine which one is more complex than others. Evaluating state-of-the-art NLP and machine learning techniques against application domains. Setting a future direction by deliberating the complexity of the dataset, lack of benchmark dataset, performance comparison, and future work. Moreover, this study critically analyzed surveys on document-level sentiment analysis and aspect-based sentiment analysis surveys. In this way, the last ten years' surveys (document and aspect level) have been investigated to find out what areas these surveys are not addressing. The main focus is on aspect-based sentiment analysis techniques (ABSTs). On the other hand, this study provides such a novelty by identifying limitations in machine learning techniques, the complexity of the dataset, and identifying application domains and NLP challenges. In addition, six application domains and four limitations have been identified by exploring the recent ten years of aspect-based sentiment analysis techniques. Finally, it thoroughly explores the application domains and limitations of aspect-based sentiment analysis techniques.}, archive = {J_NEUCOM}, author = {Jibran Mir and Azhar Mahmood and Shaheen Khatoon and Saddam Hussain and Syed Sajid Ullah and Jawaid Iqbal}, doi = {10.1016/j.neucom.2024.129237}, journal = {Neurocomputing}, month = {3}, pages = {129237}, shortjournal = {Neurocomputing}, title = {Application domains of aspect and sentiment classification techniques: A survey}, volume = {622}, year = {2025}, }
• (2025). Dataset assembly for training spiking neural networks. NEUCOM, 622, 129207. (www)

  @article{ , abstract = {Spiking Neural Networks (SNNs) are the next generation of biologically plausible Artificial Neural Networks (ANNs) which utilize neurons that communicate via binary spikes distributed over time rather than scalar values. These sparse neuron-to-neuron communications offer significant power savings when realized on purpose-built neuromorphic hardware. However, training SNNs is challenging because backpropagation, the conventional technique used for ANNs, cannot handle the discrete pulses spread across differing neurons during activation. Various SNN training platforms have been put forth that provide workarounds for this problem by loosening the constraints on neuron activation. Such approximations have the negative side-effect of forming dependencies to the specific viewpoint in which the training spike data was collected. In this paper, we show that the mechanistic models used in state-of-the-art SNN platforms perform poorly when presented with samples of differing spike viewpoints caused by changes in light intensity, simulation parameters, or Dynamic Vision Sensor (DVS) camera settings. We present a platform-agnostic approach for dataset assembly, the maximization of model potential through methodical dataset selection at the sample-level granularity. Our approach mitigates spike viewpoint dependencies by combining established SNN training approximations with a Generative Adversarial Network (GAN) to augment training datasets with broader spike distribution data. Dataset assembly improves model robustness on targeted viewpoints, without introducing significant computational overhead, by extracting untapped performance from extant starting datasets. Models trained using our approach boast a 187.28% increase in accuracy robustness when evaluated on differing spike viewpoints using the IBM DVSGesture dataset.}, archive = {J_NEUCOM}, author = {Anthony Baietto and Christopher Stewart and Trevor J. Bihl}, doi = {10.1016/j.neucom.2024.129207}, journal = {Neurocomputing}, month = {3}, pages = {129207}, shortjournal = {Neurocomputing}, title = {Dataset assembly for training spiking neural networks}, volume = {622}, year = {2025}, }
• (2025). LCL_FDA: Local context learning and full-level decoder aggregation network for large-scale point cloud semantic segmentation. NEUCOM, 621, 129321. (www)

  @article{ , abstract = {Point cloud semantic segmentation is a key techno-logy for autonomous driving, intelligent power inspection, and smart city planning, etc. However, many existing methods still face limitations in capturing local contextual information and learning long-range dependencies in point cloud representations. To adder-ss these problems, we propose a novel local context learning and full-level aggregation network for point cloud semantic segmen-tation in large-scale scenes. Firstly, we construct a local geometry feature extraction module using both coordinate and angular information. This module aims to acquire global information within local regions, thereby enhancing locally discriminative features. Secondly, we combine single-point channel attention pooling and rotation axis cross attention pooling to effectively enhance the capability of network to exploit local details. Finally, we devised a full-level decoder aggregation module, which expands the receptive field of each decoder to enhance the learning capability of the network for long-range dependencies. We evaluate our method on three large-scale point cloud datasets, including our self-constructed DW_3D dataset and the publicly available datasets Toronto3D and S3DIS. The experiment results validate the effectiveness of our method. The mIoU of the proposed method on DW_3D, Toronto-3D and S3DIS dataset can reach 72.9 %, 82.2 %, 72.3 %(6-fold). Code is avaiable https://github.com/yzq-stu/LCL_FDA .}, archive = {J_NEUCOM}, author = {Yong Li and Zhenqin Ye and Xingwen Huang and Yubin HeLi and Feng Shuang}, doi = {10.1016/j.neucom.2024.129321}, journal = {Neurocomputing}, month = {3}, pages = {129321}, shortjournal = {Neurocomputing}, title = {LCL_FDA: Local context learning and full-level decoder aggregation network for large-scale point cloud semantic segmentation}, volume = {621}, year = {2025}, }
• (2025). ChatGPT in threefold: As attacker, target, and evaluator in adversarial attacks. NEUCOM, 621, 129307. (www)

  @article{ , abstract = {ChatGPT is one of the most popular large language models (LLMs). It has achieved state-of-the-art (SOTA) performance in various natural language processing (NLP) tasks. Recent studies have shown that deep learning (DL) models reveal vulnerability to adversarial texts. However, current research on the security of ChatGPT is still in its nascent stages. In this paper, ChatGPT plays three roles in adversarial attacks: an attack method, a target model, and an evaluation tool. We propose a novel black-box word-level adversarial text generation method, ChatGPT Tricker (CGT), for the text classification task. In the framework of adversarial text generation based on word importance, CGT introduces imperceptible perturbations to the original text by utilizing ChatGPT prompts, aiming to craft adversarial texts. Subsequently, we leverage the transferability of adversarial texts to attack ChatGPT and assess its robustness against such attacks. Additionally, we introduce an innovative fluency evaluation method for adversarial texts based on ChatGPT. We define a novel concept termed offset average difference (OAD) to provide a practical formula for the computation of fluency scores. The proposed method enables automated evaluation without human intervention. We utilize CGT to attack both BERT and ChatGPT on real-world English and Chinese text classification datasets. Experimental results demonstrate that CGT balances various aspects of attack performance well. CGT produces more fluent adversarial texts while achieving an attack success rate close to the SOTA baselines. Concurrently, the results also indicate that ChatGPT resists English adversarial texts but is vulnerable to Chinese ones. On the Chinese dataset, CGT can attack ChatGPT with a success rate exceeding 30%.}, archive = {J_NEUCOM}, author = {Yunting Zhang and Lin Ye and Kai Tan and Zeshu Tian and Baisong Li and Hongli Zhang}, doi = {10.1016/j.neucom.2024.129307}, journal = {Neurocomputing}, month = {3}, pages = {129307}, shortjournal = {Neurocomputing}, title = {ChatGPT in threefold: As attacker, target, and evaluator in adversarial attacks}, volume = {621}, year = {2025}, }
• (2025). New h∞ controller for neural networks subject to time-varying delay by state estimation based approach. NEUCOM, 621, 129300. (www)

  @article{ , abstract = {This study focuses on the H ∞ control issue for neural networks (NNs) subject to time-varying delay and external disturbances. The main objective is to construct a new H ∞ controller to ensure the asymptotical stability of the closed-loop system in the case when the neuron states are not measurable. By establishing an appropriate Lyapunov–Krasovskii functional (LKF) with single and double integral terms and utilizing some effective integral inequalities, the state estimator is presented. Then, according to the estimated state, a more effective H ∞ controller is further developed, and by utilizing a decoupling technique, some less conservative results are derived which guarantees that the concerned system is asymptotically stable and achieves given H ∞ performance. At the end, the advantages and less conservativeness of the presented method is shown by one well-known simulation example.}, archive = {J_NEUCOM}, author = {Haibo Liu and Wei Qian and Yunji Zhao}, doi = {10.1016/j.neucom.2024.129300}, journal = {Neurocomputing}, month = {3}, pages = {129300}, shortjournal = {Neurocomputing}, title = {New h∞ controller for neural networks subject to time-varying delay by state estimation based approach}, volume = {621}, year = {2025}, }
• (2025). Adaptive neural fault-tolerant consensus tracking for input-delayed multi-agent systems with quantized states/input and application on mechanical systems. NEUCOM, 621, 129299. (www)

  @article{ , abstract = {This article studies the adaptive neural fault-tolerant consensus control for the multi-agent systems (MASs) with quantized states/input and time-varying input delay. Different from the results of adaptive control based on continuous states, this article presents two adaptive control algorithms for MASs with unquantized and quantized states, respectively. The entire design procedure of the control algorithms can be divided into two steps. First of all, the conventional adaptive consensus algorithm is designed for MASs with unquantized information. Secondly, a new consensus control algorithm is developed for MASs with states and input quantization. Since the quantized signal is discontinuous, which causes great difficulties in the design of consensus algorithm. This article addresses the discontinuity of quantization by establishing the relationship between the signal before and after quantization. An auxiliary system is designed to compensate for the time-varying input delay in the system. The radial basis function neural networks (RBFNNs) are applied to handle unknown terms in the system and reduce the computational complexity. Based on the designed fault-tolerant controller, the sensor faults are compensated well, and all signals of MASs can maintain semi-globally uniformly ultimately bounded (SGUUB) and the consensus errors can converge to the neighborhood near the origin. Lastly, the effectiveness and feasibility of developed control algorithm is validated by simulation results.}, archive = {J_NEUCOM}, author = {Yadong Yang and Jiyu Zhu and Qikun Shen}, doi = {10.1016/j.neucom.2024.129299}, journal = {Neurocomputing}, month = {3}, pages = {129299}, shortjournal = {Neurocomputing}, title = {Adaptive neural fault-tolerant consensus tracking for input-delayed multi-agent systems with quantized states/input and application on mechanical systems}, volume = {621}, year = {2025}, }
• (2025). A critical analysis of the theoretical framework of the extreme learning machine. NEUCOM, 621, 129298. (www)

  @article{ , abstract = {Despite several successful applications of the Extreme Learning Machine (ELM) as a new neural network training method that combines random selection with deterministic computation, we show that some fundamental principles of ELM lack a rigorous mathematical basis. In particular, we refute the proofs of two fundamental claims and construct datasets that serve as counterexamples to the ELM algorithm. Finally, we provide alternative claims to the basic principles that justify the effectiveness of ELM in some theoretical cases.}, archive = {J_NEUCOM}, author = {Irina Perfilieva and Nicolás Madrid and Manuel Ojeda-Aciego and Piotr Artiemjew and Agnieszka Niemczynowicz}, doi = {10.1016/j.neucom.2024.129298}, journal = {Neurocomputing}, month = {3}, pages = {129298}, shortjournal = {Neurocomputing}, title = {A critical analysis of the theoretical framework of the extreme learning machine}, volume = {621}, year = {2025}, }
• (2025). Adaptive population sizing for multi-population based constrained multi-objective optimization. NEUCOM, 621, 129296. (www)

  @article{ , abstract = {In recent years, the prevalence of constrained multi-objective optimization problems across numerous scenarios has incited a surge of interest in the advancement of constrained multi-objective evolutionary algorithms (CMOEAs). Multi-population CMOEAs have demonstrated effectiveness in balancing between objective optimization and constraint satisfaction, where auxiliary populations can mine infeasible regions to help main populations get grid of local optimums. However, the evolution of the auxiliary population often necessitates an equal or even greater number of function evaluations compared to the main population, leading to substantial expenditure of computational resources. To save function evaluations, this paper suggests an adaptive population sizing method to dynamically shrink the auxiliary population according to the current evolutionary state. Subsequently, a multi-stage evolutionary algorithm is developed, which integrates a variety of strategies to more effectively evolve the auxiliary population and ultimately eliminate it, thereby saving function evaluations for the main population. The proposed CMOEA is empirically evaluated against nine state-of-the-art algorithms on challenging test suites, which exhibits superior performance and versatility.}, archive = {J_NEUCOM}, author = {Ye Tian and Ruiqin Wang and Yajie Zhang and Xingyi Zhang}, doi = {10.1016/j.neucom.2024.129296}, journal = {Neurocomputing}, month = {3}, pages = {129296}, shortjournal = {Neurocomputing}, title = {Adaptive population sizing for multi-population based constrained multi-objective optimization}, volume = {621}, year = {2025}, }
• (2025). Deep keypoints adversarial attack on face recognition systems. NEUCOM, 621, 129295. (www)

  @article{ , abstract = {Face recognition systems based on deep learning have recently demonstrated an outstanding success in solving complex issues. Yet they turn out to be very vulnerable to attack. Therefore, the vulnerability of such systems has to be studied. An efficient attack strategy that deceives the face recognition system creates adversarial examples. The system may mistakenly reject a real subject as a result of such attack. Current methods for creating adversarial face images have poor perceptual quality and take too long to produce. In this paper, we introduce a novel Adversarial Attack named D K A 2 that combines both geometry and intensity based attack categories. The attack consists of three main parts: keypoints detection, Geometrically keypoints Perturbation and adversarial mask Generation. Unlike other attacks which perturb every pixel in the image, our method perturbs only the salient regions of the face represented by the keypoints (2% of the image) and the automatic generated adversarial mask. Limiting the perturbed points minimizes the distortion caused by the attack and resulting images that seem natural. Besides, the suggested approach produces stronger adversarial examples that can avoid black-box face matchers with attack success rates as high as 97,03%.}, archive = {J_NEUCOM}, author = {Emna BenSaid and Mohamed Neji and Marwa Jabberi and Adel M. Alimi}, doi = {10.1016/j.neucom.2024.129295}, journal = {Neurocomputing}, month = {3}, pages = {129295}, shortjournal = {Neurocomputing}, title = {Deep keypoints adversarial attack on face recognition systems}, volume = {621}, year = {2025}, }
• (2025). Decoupling SQL query hardness parsing for text-to-SQL. NEUCOM, 621, 129293. (www)

  @article{ , abstract = {The fundamental goal of the Text-to-SQL task is to translate natural language question into SQL query. Current research, which has made significant progress, primarily emphasizes the information coupling between natural language questions and schemas. However, the correlation between natural language questions and SQL queries is often overlooked. Decoupling this correlation may simplify the task. In this paper, we introduce an innovative framework based on Decoupling SQL Query Hardness Parsing for Text-to-SQL. The framework decouples the Text-to-SQL task based on query hardness by analyzing natural language questions and highly relevant schemas. It introduces a distributed structure to simplify multi-hardness task into single-hardness task, thereby significantly reducing the parsing burden on language models. We have evaluated our proposed framework and achieved a new state-of-the-art performance in the methods with model scales less than 3B on the Spider 1.0 dev set. Our code is shown at https://github.com/JarvenYi/DQHP .}, archive = {J_NEUCOM}, author = {Jiawen Yi and Guo Chen and Xiaojun Zhou}, doi = {10.1016/j.neucom.2024.129293}, journal = {Neurocomputing}, month = {3}, pages = {129293}, shortjournal = {Neurocomputing}, title = {Decoupling SQL query hardness parsing for text-to-SQL}, volume = {621}, year = {2025}, }
• (2025). A new prediction strategy for dynamic multi-objective optimization using hybrid fuzzy C-means and support vector machine. NEUCOM, 621, 129291. (www)

  @article{ , abstract = {A dynamic multi-objective optimization problem (DMOP) involves optimizing multiple conflicting objectives that change over time. It presents a significant challenge in rapidly adapting to evolving environments and tracking the Pareto Front. Recently, there has been increased interest in prediction-based methods due to their effectiveness and broad applicability. However, many predictive methods focus mainly on the correlation of solutions between successive generations, neglecting the overall connectedness, which leads to predicted solutions deviating from the evolutionary direction. This paper proposes a novel dynamic multi-objective optimization prediction method, called FCM-SVM-DMOEA. This method utilizes the Fuzzy C-Means (FCM) algorithm to cluster historical Pareto sets. The outcomes of Fuzzy C-Means are divided into high-quality and low-quality sets employing the Fast Non-Dominated Sorting method based on the Center Point of the Cluster. Subsequently, a Support Vector Machine (SVM) classifier is adopted to train these two sets. When environmental changes occur, the initial population consists of both variations of high-quality set and random solutions identified by the SVM classifier. Experimental results from DF suite benchmark test problems, in comparison with selected state-of-the-art algorithms, indicate that the Pareto set traced by the proposed method exhibits superior diversity and distribution, markedly improving the performance of dynamic multi-objective optimization algorithms.}, archive = {J_NEUCOM}, author = {Tao Zhang and Qing Tao and Linjun Yu and Haohao Yi and Jiawei Chen}, doi = {10.1016/j.neucom.2024.129291}, journal = {Neurocomputing}, month = {3}, pages = {129291}, shortjournal = {Neurocomputing}, title = {A new prediction strategy for dynamic multi-objective optimization using hybrid fuzzy C-means and support vector machine}, volume = {621}, year = {2025}, }
• (2025). A rule-guided interpretable lightweight framework for fetal standard ultrasound plane capture and biometric measurement. NEUCOM, 621, 129290. (www)

  @article{ , abstract = {Automating prenatal ultrasound analysis is challenging due to the need for accurate detection of standard planes and biometric measurements. Existing methods typically rely on a multi-stage process, which suffers from errors due to sequential dependencies, limited anatomical perception, and over-confidence issues. To address these challenges, we propose a novel, rule-guided, and lightweight framework that directly captures standard planes and performs biometric measurements in a single, unified process. Our model incorporates a redundant feature selection module to handle ultrasound noise and capture rich anatomical representations. By designing separable dense blocks and controlling upsampling ratios, we optimize the framework for real-time processing on constrained devices. The expert knowledge-based scoring mechanism evaluates anatomical morphology, edge clarity, and model confidence, offering interpretability and ensuring high-quality plane selection. Additionally, a gradient refinement module fine-tunes measurement points to further enhance biometric accuracy. Comparative experiments demonstrate that our method significantly outperforms other state-of-the-art approaches in fetal anatomical structure segmentation, optimal plane capture, and biometric measurement. Its computational complexity and inference speed are also impressive, showcasing substantial clinical application potential.}, archive = {J_NEUCOM}, author = {Jintang Li and Zhan Gao and Chunlian Wang and Bin Pu and Kenli Li}, doi = {10.1016/j.neucom.2024.129290}, journal = {Neurocomputing}, month = {3}, pages = {129290}, shortjournal = {Neurocomputing}, title = {A rule-guided interpretable lightweight framework for fetal standard ultrasound plane capture and biometric measurement}, volume = {621}, year = {2025}, }
• (2025). ISTD-DETR: A deep learning algorithm based on DETR and super-resolution for infrared small target detection. NEUCOM, 621, 129289. (www)

  @article{ , abstract = {Detection of small infrared targets holds significant importance in fields such as emergency rescue operations and autonomous driving. However, the low contrast and high noise sensitivity nature of infrared images often lead to poor detection performance. To address these limitations, this paper proposes ISTD-DETR, a novel approach that integrates super-resolution preprocessing with an improved Real-Time Detection Transformer (RT-DETR) model. First, this approach significantly enhances the detail information of infrared images by introducing EDSR super-resolution techniques and extensive image augmentation methods. Second, the backbone of the ISTD-DETR model is designed using an Efficient Multi-Scale Attention (EMA) structure and state space model, improving the model’s ability for feature extraction and long-range dependency modeling. Third, a new S2 feature layer and micro-target detection EncoderHead are added to the model to achieve better feature fusion and target localization, making ISTD-DETR well-suited for infrared small target detection. Additionally, the method incorporates an SPD-EMA module into the feature fusion segment, further enhancing the model’s ability to detect small-scale and low-resolution targets. Experimental results on the public datasets demonstrate that ISTD-DETR significantly outperforms existing methods, reducing false positives and false negatives while maintaining strong real-time performance.}, archive = {J_NEUCOM}, author = {Huanyu Yang and Jun Wang and Yuming Bo and Jiacun Wang}, doi = {10.1016/j.neucom.2024.129289}, journal = {Neurocomputing}, month = {3}, pages = {129289}, shortjournal = {Neurocomputing}, title = {ISTD-DETR: A deep learning algorithm based on DETR and super-resolution for infrared small target detection}, volume = {621}, year = {2025}, }
• (2025). CCL-MPC: Semi-supervised medical image segmentation via collaborative intra-inter contrastive learning and multi-perspective consistency. NEUCOM, 621, 129287. (www)

  @article{ , abstract = {Semi-supervised image segmentation extracts specific regions and tissues by utilizing extensive unlabeled images and limited labeled images, which can alleviate the dependence on plenty of accurately labeled data. However, it is difficult to learn robust feature representations due to the potential noise in pseudo-labels caused by inefficient consistency learning, and poor class diversity in feature spaces. To address this issue, we propose a semi-supervised medical image segmentation method using Collaborative intra-inter Contrastive Learning and Multi-Perspective Consistency (CCL-MPC). First, we propose a collaborative intra-inter contrastive learning strategy that includes symmetric bidirectional contrastive learning and certainty-guided contrastive learning, to exploit the intrinsic differences between inter-image and intra-image feature representations. Second, we design a multi-perspective consistency learning strategy to improve the class diversity by utilizing a dual-branch network and two augmented views. Additionally, we dynamically partition the pseudo-label certainty area for auxiliary consistency learning to reduce the potential noise during the training process. Experimental results on the publicly available datasets demonstrate that CCL-MPC can achieve better segmentation performance than the state-of-the-art methods for semi-supervised medical image segmentation tasks. The source code is available at https://github.com/EmarkZOU/CCL-MPC .}, archive = {J_NEUCOM}, author = {Xiaogang Du and Yibin Zou and Tao Lei and Weichuan Zhang and Yingbo Wang and Asoke K. Nandi}, doi = {10.1016/j.neucom.2024.129287}, journal = {Neurocomputing}, month = {3}, pages = {129287}, shortjournal = {Neurocomputing}, title = {CCL-MPC: Semi-supervised medical image segmentation via collaborative intra-inter contrastive learning and multi-perspective consistency}, volume = {621}, year = {2025}, }
• (2025). EPCTS: Enhanced prompt-aware cross-prompt essay trait scoring. NEUCOM, 621, 129283. (www)

  @article{ , abstract = {For many years, automatic essay scoring (AES) has become one of the primary applications of natural language processing in the field of education, aiming to assess the quality of the given essays automatically and accurately. However, due to conflicts between prompt-specific scoring and real educational contexts, the research scope has expanded from prompt-specific scoring to cross-prompt scoring, and further extended to scoring based on different traits. However, cross-prompt scoring imposes higher demands on the model’s ability to extract feature representations of essays and its generalization capabilities, posing one of the main challenges in current AES research. Currently, there has been considerable research on representing essays themselves, but scoring tasks that integrate the relationship between essays and prompts remain scarce. In this paper, we propose a robust model, named EPCTS (Enhanced Prompt-aware Cross-prompt Essay Trait Scoring), which effectively utilizes known prompt information to assist in the scoring task. We also design a prompt embedding method that combines part-of-speech (POS) features and syntactic features, and further develop an algorithm for prompt-essay relationship features. Experimental results demonstrate that our EPCTS model achieves state-of-the-art performance and higher consistency. The significant improvement in essay scoring under low-resource prompts also demonstrates the superiority of our model, alleviating the challenge of data scarcity for essays under low-resource prompts.}, archive = {J_NEUCOM}, author = {Jiangsong Xu and Jian Liu and Mingwei Lin and Jiayin Lin and Shenbao Yu and Liang Zhao and Jun Shen}, doi = {10.1016/j.neucom.2024.129283}, journal = {Neurocomputing}, month = {3}, pages = {129283}, shortjournal = {Neurocomputing}, title = {EPCTS: Enhanced prompt-aware cross-prompt essay trait scoring}, volume = {621}, year = {2025}, }
• (2025). DSTF: A diversified spatio-temporal feature extraction model for traffic flow prediction. NEUCOM, 621, 129280. (www)

  @article{ , abstract = {Traffic flow prediction forms a critical foundation for the management and planning of urban transportation systems. However, the complex spatial interactions among road segments and the dynamic patterns of traffic flow variations across multiple time scales pose significant challenges to improving forecasting accuracy. To address these complexities, this paper introduces a Diversified Spatio-Temporal Feature Extraction (DSTF) Model, designed to effectively mine temporal, spatial, and spatio-temporal cross-correlations. Specifically, in the temporal dimension, a gated convolution-enhanced Res2Net architecture is employed to capture diverse traffic flow patterns across varying time scales. In the spatial dimension, the model integrates global and local perspectives by employing spatial attention mechanisms and a dual-view Geom-GCN, enabling it to capture global spatial dependencies, local geographic neighborhood relationships, and semantic similarity-driven spatial correlations among nodes within the urban road network. For spatio-temporal cross-correlations, a dynamic synchronization aggregation module is developed using spatio-temporal attention, effectively capturing the evolving interactions between nodes or regions across different time slices. Experimental evaluations conducted on four real-world highway traffic datasets demonstrate that DSTF outperforms PDFormer, achieving average improvements of 1.35%, 1.61% and 0.81% in MAE, MAPE and RMSE metrics, respectively. These results highlight the model’s superior capability in extracting and leveraging temporal and spatial features for traffic flow prediction.}, archive = {J_NEUCOM}, author = {Xing Wang and Xiaojun Wang and Faliang Huang and Fumin Zou and Lyuchao Liao and Ruihao Zeng}, doi = {10.1016/j.neucom.2024.129280}, journal = {Neurocomputing}, month = {3}, pages = {129280}, shortjournal = {Neurocomputing}, title = {DSTF: A diversified spatio-temporal feature extraction model for traffic flow prediction}, volume = {621}, year = {2025}, }
• (2025). A survey of super-resolution image quality assessment. NEUCOM, 621, 129279. (www)

  @article{ , abstract = {Image super-resolution (SR) technology plays a crucial role in enhancing image resolution and has found extensive applications in areas such as medical imaging, surveillance, and security. The growing need for effective assessment methodologies in SR image quality (SR-IQA) has spurred significant scholarly research. This paper provides a comprehensive review of the recent advancements in SR-IQA, including both the evolution of algorithms and the creation of subjective evaluation databases. It critically examines and analyzes various SR-IQA techniques across different datasets, starting with an overview of the development of SR algorithms and highlighting the importance of SR-IQA research. The paper then introduces and compares mainstream SR-IQA datasets, categorizing and evaluating SR-IQA methodologies. It concludes with an assessment of key image quality assessment algorithms, supported by detailed dataset analyses. The conclusion summarizes the current state of SR-IQA, identifies ongoing challenges, and suggests future research directions to advance the field.}, archive = {J_NEUCOM}, author = {Lei Shu and Qinru Zhu and Yujie He and Wei Chen and Jiebin Yan}, doi = {10.1016/j.neucom.2024.129279}, journal = {Neurocomputing}, month = {3}, pages = {129279}, shortjournal = {Neurocomputing}, title = {A survey of super-resolution image quality assessment}, volume = {621}, year = {2025}, }
• (2025). TNPNet: An approach to few-shot open-set recognition via contextual transductive learning. NEUCOM, 621, 129276. (www)

  @article{ , abstract = {In the field of Few-Shot Open-Set Recognition (FSOSR), identifying images from known and unknown categories with limited data is a pressing challenge. Existing approaches often falter due to complex learning strategies or reliance on pseudo-negative sample features. To address this, we introduce the Transductive Negative Prototype learning Network (TNPNet), a straightforward yet effective architecture for FSOSR. Leveraging both local and global context, our novel Contextual Transductive Learning (CTL) method enhances prototypes for both known and unknown categories. TNPNet also features a parameter-free classifier designed to optimize CTL-enhanced prototypes and introduces a multi-line loss for robust supervision. Through extensive evaluations, TNPNet demonstrates superior AUROC metrics across multiple scenarios and datasets, affirming its robustness and practicality in real-world applications. Code is available at: https://github.com/wushaoling540/TNPNet/}, archive = {J_NEUCOM}, author = {Shaoling Wu and Huilan Luo and Xiaoming Lin}, doi = {10.1016/j.neucom.2024.129276}, journal = {Neurocomputing}, month = {3}, pages = {129276}, shortjournal = {Neurocomputing}, title = {TNPNet: An approach to few-shot open-set recognition via contextual transductive learning}, volume = {621}, year = {2025}, }
• (2025). Adaptive neural networks-based event-triggered formation control for multi-robot source localization. NEUCOM, 621, 129275. (www)

  @article{ , abstract = {In this paper, hybrid event-triggered formation control schemes in the presence/absence of velocity measurements are proposed for multiple robots in source localization of an unknown scalar field. A neural network model is used to approximate the unknown scalar field, then we put forward adaptive strategies to calculate the gradient of this field. Considering the obstacles in reality as well as the communication burden and calculation cost between robots, obstacle avoidance policies and event-triggered mechanisms are designed to deal with them. Moreover, we develop approximation functions to avoid the chattering phenomenon caused by the sign function in these schemes. In addition, we build passive filters and velocity estimators and further apply them to control schemes under no velocity measurements. Eventually, a simulation test is performed to verify the effectiveness of the proposed scheme.}, archive = {J_NEUCOM}, author = {Rui-Guo Li and Ze-Hao Shi and Huai-Ning Wu}, doi = {10.1016/j.neucom.2024.129275}, journal = {Neurocomputing}, month = {3}, pages = {129275}, shortjournal = {Neurocomputing}, title = {Adaptive neural networks-based event-triggered formation control for multi-robot source localization}, volume = {621}, year = {2025}, }
• (2025). A crisis event classification method based on a multimodal multilayer graph model. NEUCOM, 621, 129271. (www)

  @article{ , abstract = {Quickly obtaining and classifying relevant information about crisis events via social media platforms, such as Twitter and Weibo, plays a critical role in the subsequent rescue operations and post-disaster reconstruction. Current crisis event classification technologies are unable to comprehensively acquire data and are not applicable to a wide range of scenarios. To address this issue, this paper proposes a crisis event prediction method based on a multi-information multimodal deep graph model. Exploring multimodal graph information innovatively and deeply enhances the effectiveness and broadens the applicability of predicting crisis events. First, Bi-LSTM and a GCN structure based on word graphs are used to obtain the sentence and grammatical information of the text, constructing a text feature vector. The image feature vector is constructed by combining a CNN and a Transformer. Second, a graph of the text group and a graph of the image group are constructed on the basis of cosine similarity, and a multilayer structure combining autoencoders and a GCN is used to mine the integrated information of the data. Finally, multimodal feature vectors are converted into a multimodal semantic network, and a multilayer GCN structure with an attention mechanism is employed for crisis event classification prediction. The experimental results on the CrisisMMD dataset and its generated datasets demonstrate that the proposed model outperforms the current state-of-the-art models.}, archive = {J_NEUCOM}, author = {Jing Wang and Shuo Yang and Hui Zhao and Yanyan Chen}, doi = {10.1016/j.neucom.2024.129271}, journal = {Neurocomputing}, month = {3}, pages = {129271}, shortjournal = {Neurocomputing}, title = {A crisis event classification method based on a multimodal multilayer graph model}, volume = {621}, year = {2025}, }
• (2025). Transmission map and background light guided enhancement of unpaired underwater image. NEUCOM, 621, 129270. (www)

  @article{ , abstract = {Due to the difficulty of obtaining paired underwater images, unpaired underwater image enhancement using CycleGAN has received attention. However, existing unpaired methods suffer from over-enhancement or under-enhancement, due to (1) neglect the constraint of domain knowledge in underwater imaging, (2) lack of consideration about the characteristics adversarial of CycleGAN. To this end, we propose an unpaired underwater image enhancement algorithm based on the physical maps guidance. Specifically, to introduce the underwater imaging constraints into CycleGAN, we propose the UIIE-Subnet to estimate physical maps for modeling enhancement process and UDE-Subnet for degradation process. Then, based on cycle consistency, we derive the physical map consistency assumption to explore the relationship of physical maps in different processes, thereby proposing the transmission map loss and background light loss to promote cycle generative learning. Experiments on public datasets show that the proposed unpaired method produces more natural images, which outperforms most paired and unpaired enhancement methods.}, archive = {J_NEUCOM}, author = {Simeng Cheng and Zhigang Jin and Xiaodong Wu and Jiawei Liang}, doi = {10.1016/j.neucom.2024.129270}, journal = {Neurocomputing}, month = {3}, pages = {129270}, shortjournal = {Neurocomputing}, title = {Transmission map and background light guided enhancement of unpaired underwater image}, volume = {621}, year = {2025}, }
• (2025). UA-FER: Uncertainty-aware representation learning for facial expression recognition. NEUCOM, 621, 129261. (www)

  @article{ , abstract = {Facial Expression Recognition (FER) remains a challenging task due to unconstrained conditions like variations in illumination, pose, and occlusion. Current FER approaches mainly focus on learning discriminative features through local attention and global perception of visual encoders, while neglecting the rich semantic information in the text modality. Additionally, these methods rely solely on the softmax-based activation layer for predictions, resulting in overconfident decision-making that hampers the effective handling of uncertain samples and relationships. Such insufficient representations and overconfident predictions degrade recognition performance, particularly in unconstrained scenarios. To tackle these issues, we propose an end-to-end FER framework called UA-FER, which integrates vision–language pre-training (VLP) models with evidential deep learning (EDL) theory to enhance recognition accuracy and robustness. Specifically, to identify multi-grained discriminative regions, we propose the Multi-granularity Feature Decoupling (MFD) module, which decouples global and local facial representations based on image–text affinity while distilling the universal knowledge from the pre-trained VLP models. Additionally, to mitigate misjudgments in uncertain visual–textual relationships, we introduce the Relation Uncertainty Calibration (RUC) module, which corrects these uncertainties using EDL theory. In this way, the model enhances its ability to capture emotion-related discriminative representations and tackle uncertain relationships, thereby improving overall recognition accuracy and robustness. Extensive experiments on in-the-wild and in-the-lab datasets demonstrate that our UA-FER outperforms the state-of-the-art models.}, archive = {J_NEUCOM}, author = {Haoliang Zhou and Shucheng Huang and Yuqiao Xu}, doi = {10.1016/j.neucom.2024.129261}, journal = {Neurocomputing}, month = {3}, pages = {129261}, shortjournal = {Neurocomputing}, title = {UA-FER: Uncertainty-aware representation learning for facial expression recognition}, volume = {621}, year = {2025}, }
• (2025). TEG-DI: Dynamic incentive model for federated learning based on tripartite evolutionary game. NEUCOM, 621, 129259. (www)

  @article{ , abstract = {Federated Learning (FL), as a distributed machine learning approach that protects data privacy, has gained significant attention in recent years. However, some participants benefit from the global model without contributing private data (i.e., free-riding), which challenges FL’s development. Although existing studies attempt to encourage data contribution through game-theoretic models, they often overlook dynamic strategy changes across multiple training iterations, particularly the spontaneous transitions in FR behavior. Moreover, these studies do not adequately address participants contributing minimal data or only partially engaging in training and we define this behavior as partial free-riding (PFR). To address these issues, we propose a dynamic incentive model based on tripartite evolutionary game theory (TEG-DI), dividing participants into the federated organizer (FO) and the federated collaborator (FC) while constructing a novel tripartite game framework with the central server (CS). We treat PFR as a strategy adopted by collaborators and analyze the evolutionarily stable strategy (ESS) to explore the complex interactions between participants and the server across six scenarios. Based on the analysis, we dynamically adjust the incentive mechanisms and propose optimization suggestions to address potential non-ideal evolutionary states. The experimental results demonstrate that the TEG-DI model effectively encourages free-riders to transition into honest participants, reducing the free-riders gains by 10.51%.}, archive = {J_NEUCOM}, author = {Jiuyun Xu and Yingzhi Zhao and Xiaowen Li and Liang Zhou and Kongshang Zhu and Xiangrui Xu and Qiang Duan and Ruru Zhang}, doi = {10.1016/j.neucom.2024.129259}, journal = {Neurocomputing}, month = {3}, pages = {129259}, shortjournal = {Neurocomputing}, title = {TEG-DI: Dynamic incentive model for federated learning based on tripartite evolutionary game}, volume = {621}, year = {2025}, }
• (2025). EOOD: End-to-end oriented object detection. NEUCOM, 621, 129251. (www)

  @article{ , abstract = {Recently, significant advancements have been made in oriented object detectors based on convolutional networks. However, these models often rely on hand-designed post-processing techniques such as non-maximum suppression (NMS) to suppress redundant predictions, which impedes establishing an end-to-end detection system. In this paper, we explore how to build an end-to-end oriented detector? Firstly, our research demonstrates that prediction-oriented one-to-one label assignment (P O LA) can significantly reduce the performance gap between using and not using NMS and is an essential component of end-to-end detection. Additionally, the proposed Negative Sample Reweighted Focal Loss ( NFL ) can widen the classification confidence gap between positive and negative samples, separate positive samples from noise, and guarantee high classification scores for the single positive sample. Finally, in order to address the lack of supervision caused by one-to-one label assignment, a joint training pipeline is designed that unites multiple auxiliary heads and takes advantage of one-to-many label assignment on supervision to improve feature representation and increase performance. During inference, Only the main head trained with one-to-one label assignment is involved in prediction. Extensive experiments on publicly available datasets DOTA and DIOR-R demonstrate that the proposed EOOD exhibits significant performance improvements over baseline models and has the potential to overcome the limitations of NMS. Code is available at https://github.com/zhangiguang/EOOD .}, archive = {J_NEUCOM}, author = {Caiguang Zhang and Zilong Chen and Boli Xiong and Kefeng Ji and Gangyao Kuang}, doi = {10.1016/j.neucom.2024.129251}, journal = {Neurocomputing}, month = {3}, pages = {129251}, shortjournal = {Neurocomputing}, title = {EOOD: End-to-end oriented object detection}, volume = {621}, year = {2025}, }
• (2025). PINR: A physics-integrated neural representation for dynamic fluid scenes. NEUCOM, 621, 129250. (www)

  @article{ , abstract = {Recent research based on neural representation has achieved impressive results in dynamic reconstruction. However, reconstructing high-fidelity dynamic fluid scenes from multi-view videos remains challenging due to the complex fluid phenomena such as translucency, scattering, occlusion, and visual ambiguity. In this paper, we present a novel neural representation that simultaneously exploits physics knowledge and differentiable volume rendering techniques to construct a physically plausible neural field. We first propose a simple yet effective state field to separate fluid and non-fluid contents apart, which essentially reduces the amount of computation and appearance ambiguity. To achieve robust optimization and fast convergence, we introduce intermediate velocity in neural representation based on Helmholtz Decomposition to decouple physical quantities (velocity, pressure and density) that are tightly coupled in physical equations. In contrast to previous physics-informed reconstruction methods, we explore a novel solution by introducing a fast and stable fluid solver to update the intermediate velocity to a physically conserved incompressible velocity. This solver is based on a parallelizable multi-scale architecture, which can effectively avoid degenerate solutions. By adopting the proposed physics-aware loss, we achieved better convergence and superior performance in dynamic fluid scenes. Extensive experiments show that the proposed method can achieve state-of-the-art results for reconstructing dynamic fluid scenes, and the performance and efficiency have been significantly improved in different scenarios.}, archive = {J_NEUCOM}, author = {Ziyu Zhao and Sifan Zhou and Xiaobo Lu and Weili Zeng and Jian Qian}, doi = {10.1016/j.neucom.2024.129250}, journal = {Neurocomputing}, month = {3}, pages = {129250}, shortjournal = {Neurocomputing}, title = {PINR: A physics-integrated neural representation for dynamic fluid scenes}, volume = {621}, year = {2025}, }
• (2025). Incomplete multi-view subspace clustering based on robust matrix completion. NEUCOM, 621, 129240. (www)

  @article{ , abstract = {Multi-view clustering (MVC) is a popular approach used in data mining and pattern recognition to enhance clustering performance by leveraging complementary information from different data sources or features. However, real-world multi-view data often suffer from incomplete views, where some samples lack representation on one or more views due to various faults during data collection or processing. Efficiently exploiting the connections between multiple views from such incomplete data presents a significant challenge compared to complete-view clustering. Although several methods have been proposed in recent years to address the problem of incomplete multi-view clustering (IMVC), existing approaches have not fully considered: (1) the differential negative effects associated with varying missing data rates across different views, and (2) the impact of intrinsic noise arising from the diverse quality of these views. To address this problem, we propose a novel subspace clustering framework based on robust matrix completion for incomplete and unlabeled multi-view data. The proposed model consists of two modules. The first is a regularized missing data completion module from an information theoretic perspective, and the second is a centralized subspace model, combining low rank properties. In addition, we develop an efficient iterative algorithm using half-quadratic (HQ) and linearized alternating direction method (ADM). Finally, we conduct experiments on five benchmark multi-view datasets to demonstrate that our proposed method outperforms state-of-the-art methods in most datasets and metrics.}, archive = {J_NEUCOM}, author = {Lei Xing and Xinhu Zheng and Yao Lu and Badong Chen}, doi = {10.1016/j.neucom.2024.129240}, journal = {Neurocomputing}, month = {3}, pages = {129240}, shortjournal = {Neurocomputing}, title = {Incomplete multi-view subspace clustering based on robust matrix completion}, volume = {621}, year = {2025}, }
• (2025). A light-weight deep learning framework for low light image enhancement. NEUCOM, 621, 129236. (www)

  @article{ , abstract = {In computer vision, deep learning-based methods for improving low-light images have gained popularity. The proposed lightweight end-to-end deep neural network architecture is designed by minimizing the number of trainable parameters while optimizing design choices for efficiency and ensuring fast inference time. The proposed architecture consists of denoising, enhancing, and fusion modules designed to enhance image visibility, and contrast and reduce noise while preserving content and color information. We used a modified convolutional neural network (CNN)-based framework for exposure fusion that is designed to identify and rectify hidden degradation within dimly light images and highly adaptive to diverse lighting conditions. However, after conducting quantitative experiments, we have found that the proposed method outperforms the state-of-the-art TTST by about 0.48 dB and EDiffSR by 1.48 dB. Our lightweight method accounts for 8.28% and 6.77% of the computational cost (FLOPs) of TTST and EDiffSR respectively, and requires just 1.91% and 1.35% of their trainable parameters additionally.}, archive = {J_NEUCOM}, author = {Laraib Zainab and Hammad Afzal and Khawir Mahmood and Omar Arif}, doi = {10.1016/j.neucom.2024.129236}, journal = {Neurocomputing}, month = {3}, pages = {129236}, shortjournal = {Neurocomputing}, title = {A light-weight deep learning framework for low light image enhancement}, volume = {621}, year = {2025}, }
• (2025). Event-level supervised contrastive learning with back-translation augmentation for event causality identification. NEUCOM, 621, 129232. (www)

  @article{ , abstract = {How to identify event causality from natural language texts is becoming a heated research topic due to the significant role played by causality in AI domain. However, the limited scale and complex causality expressions of existing labeled datasets still impede better event causality identification (ECI) performance, especially in sentence-level ECI. Here, this study proposes the event-level supervised contrastive learning with back-translation augmentation method (BT-ESupCL) to address the aforementioned two obstacles simultaneously. Specifically, an efficient compound back-translation augmentation method with three parallel strategies is designed to mitigate data scarcity accompanied by more expression diversity and lower computational cost. To validly model the intricate causality expressions (in terms of both event semantics and casual structures) and then train a robust model, BT-ESupCL develops novel sampling strategies to construct contrastive pairs with more causality interactions. And, a fine-tuned supervised contrastive loss is devised to cope with imbalanced class distributions and data noise. Comprehensive experiments and analyses are implemented to verify the efficacy of our method. The results demonstrate that our approach achieves more promising performance than prior methods with 1.3% and 2.2% F1-score improvements on EventStoryLine Corpus and CausalTimeBank respectively, and helps to gain more diverse samples, more robustness to noise.}, archive = {J_NEUCOM}, author = {Shunhang Li and Gang Zhou and Jing Chen and Yepeng Sun and Ningbo Huang and Sisi Peng}, doi = {10.1016/j.neucom.2024.129232}, journal = {Neurocomputing}, month = {3}, pages = {129232}, shortjournal = {Neurocomputing}, title = {Event-level supervised contrastive learning with back-translation augmentation for event causality identification}, volume = {621}, year = {2025}, }
• (2025). On using derivatives and multiple kernel methods for clustering and classifying functional data. NEUCOM, 621, 129231. (www)

  @article{ , abstract = {In this paper, we propose a framework for rich representation of smooth functional data, leveraging a multiview approach that considers functions and their derivatives as complementary sources of information. Additionally, motivated by the non-linear nature of functional data, we advocate for kernel methods as a suitable modeling approach. We extend existing multiple kernel learning techniques for multivariate data to handle functional data. In particular, we introduce a general procedure for linearly combining different kernel functions. We apply this framework to both clustering and classification tasks, extending multiple kernel k -means and multiple kernel SVM methods to Sobolev functions in H q . Our experiments involve both simulated and real-world data, demonstrating the effectiveness of our proposed methods.}, archive = {J_NEUCOM}, author = {Julien Ah-Pine and Anne-Françoise Yao}, doi = {10.1016/j.neucom.2024.129231}, journal = {Neurocomputing}, month = {3}, pages = {129231}, shortjournal = {Neurocomputing}, title = {On using derivatives and multiple kernel methods for clustering and classifying functional data}, volume = {621}, year = {2025}, }
• (2025). SYN-PBOX: A large-scale benchmark dataset of box-shaped objects for scene understanding of bin picking. NEUCOM, 621, 129215. (www)

  @article{ , abstract = {Robotic grasping and sorting tasks require a large amount of comprehensive and high-quality data, but collecting real-world data is too expensive and time-consuming in practice. To address this issue and bridge the domain gap when training with synthetic datasets for specific scenarios, this paper utilizes domain randomization (DR) and its enhanced techniques, and generates synthetic dataset of box-shaped objects with complex scenarios. Our dataset contains 15,000 realistic images (including: point clouds, RGB images, depth images) of 21 box-shaped objects and detailed annotation data in different poses and cluttered conditions (including: 6D pose, segmentation mask, instance segmentation, and instance segmentation of the reasonable grasping surface for each object). Experiments demonstrate the superior performance of our dataset in segmentation, pose estimation, grasping, and its ability to be effectively applied to real-world scenarios. Furthermore, we propose a deep neural network called YOLOv8-D, which can achieve a grasping success rate of over 95% on previously unseen real objects after training on our dataset.}, archive = {J_NEUCOM}, author = {Jiale Cui and Jinlong Shi and Caisheng Liu and Suqin Bai and Xin Shu and Xin Zuo and Qiang Qian and Dan Xu and Haoran Deng}, doi = {10.1016/j.neucom.2024.129215}, journal = {Neurocomputing}, month = {3}, pages = {129215}, shortjournal = {Neurocomputing}, title = {SYN-PBOX: A large-scale benchmark dataset of box-shaped objects for scene understanding of bin picking}, volume = {621}, year = {2025}, }
• (2025). Adversarial purification of information masking. NEUCOM, 621, 129214. (www)

  @article{ , abstract = {Adversarial attacks meticulously generate minuscule, imperceptible perturbations that add to images to deceive neural networks. Adversarial purification methods seek to remove perturbations using generative models to achieve defense. However, residual perturbations lead to less-than-ideal results. Under the premise that perturbations are difficult to remove completely, we are the first to quantify the hazards of residual perturbations and explore how to achieve more robust defenses by reducing perturbations and resisting the impact of residual perturbations. Motivated by this, we propose a novel adversarial purification approach named Information Mask Purification (IMPure). Our method utilizes informative masks and a regional intersection reconstruction to generate images to reduce the perturbation residues. During training, we use the combination module to guide the generative model in recovering feature representations. Finally, we establish a combined constraint of pixel loss and perceptual loss to augment the model’s reconstruction adaptability. Extensive experiments on the complex dataset ImageNet with classifier models demonstrate that our approach achieves state-of-the-art results in defending against adversarial attack methods. Implementation code and pre-trained weights can be accessed at https://github.com/NoWindButRain/IMPure .}, archive = {J_NEUCOM}, author = {Sitong Liu and Zhichao Lian and Shuangquan Zhang and Liang Xiao}, doi = {10.1016/j.neucom.2024.129214}, journal = {Neurocomputing}, month = {3}, pages = {129214}, shortjournal = {Neurocomputing}, title = {Adversarial purification of information masking}, volume = {621}, year = {2025}, }
• (2025). Global-aware interaction network for RGB-D salient object detection. NEUCOM, 621, 129204. (www)

  @article{ , abstract = {Most existing RGB-D salient object detection (SOD) methods rely on depth images to complement RGB images, improving detection accuracy. However, under complex or low-light conditions, the poor quality of depth images often introduces substantial interference, negatively affecting model performance. Consequently, effectively integrating complementary information from both RGB and depth images remains a significant challenge in this field. In this paper, we propose a Global-aware Interaction Network (GAINet) for RGB-D SOD to better capture depth image cues and address modality discrepancies. Specifically, we introduce a Cross-Modal Feature Fusion Module (CMFFM), which consists of a Feature Extraction Enhancement (FEE) module and a Modal Interaction (MI) module. The FEE module refines and enhances the features from both RGB and depth images, while the MI module uses depth features to guide the detection of salient objects in the RGB features. Additionally, a Multi-layer Complementary Module (MCM) is designed to further enhance and refine multi-level complementary features, enabling step-by-step decoding of salient feature maps. A hybrid loss function is employed to optimize GAINet’s training process. Extensive experiments show that GAINet outperforms 15 state-of-the-art RGB-D SOD methods across six publicly available datasets. The codes can be found in https://github.com/wzxxmj/GAINet/ .}, archive = {J_NEUCOM}, author = {Zijian Jiang and Ling Yu and Yu Han and Junru Li and Fanglin Niu}, doi = {10.1016/j.neucom.2024.129204}, journal = {Neurocomputing}, month = {3}, pages = {129204}, shortjournal = {Neurocomputing}, title = {Global-aware interaction network for RGB-D salient object detection}, volume = {621}, year = {2025}, }
• (2025). Detection of cervical lesion cell based on the difference of context cells. NEUCOM, 621, 129198. (www)

  @article{ , abstract = {Cervical cancer is a prevalent malignant tumor that poses a significant threat to the lives of women. Nowadays, significant advancements have been made in the field of deep learning-based techniques for the detection of abnormal cervical cells. However, there are still some issues hindering the performance improvement of cervical abnormal cells detection, such as the differences between normal cells and abnormal cells, as well as the differences among abnormal cells. To solve the aforementioned problems and improve the detection performance of cervical abnormal cell, we have developed an attention-based model for detecting abnormal cervical cells. We propose the Global–Local Features Fusion Detection Head and the Local Features Contrastive Auxiliary Head. The global–local feature fusion method aims to mimic clinical practice, introducing global features and fusing them with local features to contextualize and enhance foreground while suppressing background in pathological images. The local feature contrast method leverages attention and grouped supervised contrastive learning, extracting more effective, typifying abnormal cell features through comparing abnormal cells, distinguishing those with similar appearance or different categories, and those with same category but different appearance. We validate on 28974 patches of abnormal cervical cells. Compared to existing methods, our proposed approach has achieved outstanding performance, achieving an AP of 50.7%, and effectively mitigated the issues raised, demonstrating its effectiveness in automated cervical abnormal cell detection.}, archive = {J_NEUCOM}, author = {Yizhou Chen and Qiuxiang He and Huiyan Jiang and Wenbo Pang}, doi = {10.1016/j.neucom.2024.129198}, journal = {Neurocomputing}, month = {3}, pages = {129198}, shortjournal = {Neurocomputing}, title = {Detection of cervical lesion cell based on the difference of context cells}, volume = {621}, year = {2025}, }
• (2025). Adaptive optimal consensus of nonlinear multi-agent systems with unknown dynamics using off-policy integral reinforcement learning. NEUCOM, 621, 129185. (www)

  @article{ , abstract = {Reinforcement learning (RL) has been identified as a promising approach for developing adaptive optimal consensus schemes for high-order strict-feedback nonlinear multi-agent systems (MASs). However, existing methods have limitations, as they can only be applied to systems with partially unknown dynamics and require an identifier–actor–critic framework. This paper proposes a novel approach that combines classical backstepping techniques and off-policy integral reinforcement learning (IRL) to circumvent these limitations and develop an adaptive optimal consensus scheme for nonlinear MASs with completely unknown dynamics. Specifically, we introduce an off-policy IRL-based adaptive optimal consensus scheme that can obtain optimal control inputs without knowledge of the system dynamics. The algorithm utilizes the actor–critic structure and updates the weight vectors using only one learning rule in each step based on the collected system trajectory data. We have proven that the optimal consensus is achieved, and the estimation errors of the optimal weight vectors are uniformly ultimately bounded (UUB). Finally, we present a simulation example to validate the effectiveness of the proposed approach.}, archive = {J_NEUCOM}, author = {Lei Yan and Zhi Liu and C.L. Philip Chen and Yun Zhang and Zongze Wu}, doi = {10.1016/j.neucom.2024.129185}, journal = {Neurocomputing}, month = {3}, pages = {129185}, shortjournal = {Neurocomputing}, title = {Adaptive optimal consensus of nonlinear multi-agent systems with unknown dynamics using off-policy integral reinforcement learning}, volume = {621}, year = {2025}, }
• (2025). SDI: A sparse drift identification approach for force/torque sensor calibration in industrial robots. NEUCOM, 620, 129292. (www)

  @article{ , abstract = {Calibrating force/torque (F/T) sensor drift is an enduring objective for robotic precise force control. This article presents a novel drift identification method to discover the dynamics of F/T sensor drifts from noisy measurement data, which is conducive to accurate sensor drift compensation. In the drift identification method, a linear dynamical model with measurement noise is formulated to characterize the evolution of sensor drift, and an expectation–maximization optimization framework which integrates Kalman smoothing with sparse Bayesian learning is put forward to identify the parameters of the linear dynamical model using F/T sensor measurement data. The effectiveness of the proposed drift identification method is validated on extensive robotic experiments including scenarios with unloaded mass, loaded mass, and contact force. Experimental results demonstrate the superiority of the proposed drift identification method compared to the conventional least square method for sensor calibration.}, archive = {J_NEUCOM}, author = {Xuechun Qiao and Chenxuan Xu and Yasen Wang and Guijun Ma}, doi = {10.1016/j.neucom.2024.129292}, journal = {Neurocomputing}, month = {3}, pages = {129292}, shortjournal = {Neurocomputing}, title = {SDI: A sparse drift identification approach for force/torque sensor calibration in industrial robots}, volume = {620}, year = {2025}, }
• (2025). PaintDiffusion: Towards text-driven painting variation via collaborative diffusion guidance. NEUCOM, 620, 129284. (www)

  @article{ , abstract = {Artistic image creation has recently attracted considerable interest. Style and content are two key components of an artistic image. Existing methods typically learn style from paintings/texts while extracting content from photographs, or alternatively, derive both style and content from texts. In this paper, we revisit the distinctive characteristics of style and content to explore a more reasonable way for artistic image creation. Intuitively, style is usually highly abstract and intricate, which poses a challenge when describing it in words. Therefore, representing style with a painting is often a better choice. Conversely, content is more concrete and tangible, making it more easily describable through texts. Based on above analysis, we propose a novel text-driven painting variation method called PaintDiffusion, which creates new artistic images by rearranging the style elements within a painting to align with the content structures described by a text prompt. PaintDiffusion is built upon the pretrained Stable Diffusion model conditioning on images/texts. To achieve our goal, we first design a condition seeking strategy to find an optimal condition code that can seamlessly integrate the style information from the image condition and the content information from the text condition. Moreover, a style-content calibration strategy is introduced to further refine and enhance the details by adjusting the noisy latent code with the gradients of well-designed calibration functions. The obtained condition code, along with the adjusted latent code, can collaboratively guide our model to produce desired artistic images with impressive quality. In addition, our PaintDiffusion does not require heavy model training or fine-tuning, significantly reducing the time and resources required. Extensive experiments are conducted to verify the effectiveness of our method.}, archive = {J_NEUCOM}, author = {Haibo Chen and Zikun Chen and Lei Zhao and Jun Li and Jian Yang}, doi = {10.1016/j.neucom.2024.129284}, journal = {Neurocomputing}, month = {3}, pages = {129284}, shortjournal = {Neurocomputing}, title = {PaintDiffusion: Towards text-driven painting variation via collaborative diffusion guidance}, volume = {620}, year = {2025}, }
• (2025). A hybrid algorithm with inlier-guided hough voting for point cloud registration. NEUCOM, 620, 129277. (www)

  @article{ , abstract = {Point cloud registration is a fundamental yet challenging task due to the contamination by large numbers of outliers in correspondences. Following the classical model generation-and-selection solution, many methods are dedicated to exploring effective outlier removal strategies to obtain inlier correspondences and generate candidate hypotheses. However, these methods usually suffer from inaccurate model selection, resulting in the failure of point cloud registration. In this paper, we propose a hybrid point cloud registration framework that considers both the quality of hypothesis generation and the accuracy of model selection. For model generation, a global compatibility-based inlier sampling is performed to obtain outlier-free consensus and produce reliable hypotheses. Further, we present an inlier-guided Hough voting network for model selection, which incorporates the inlier count prior into the voting process and conducts maximum likelihood estimation to select the best transformation accurately. Extensive experiments on 3DMatch, 3DLoMatch, and KITTI datasets indicate that the proposed point cloud registration method achieves state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Qun Jiang and Zhi Chen and Fan Yang and Lin Guo and Luxia Ai and Wenbing Tao}, doi = {10.1016/j.neucom.2024.129277}, journal = {Neurocomputing}, month = {3}, pages = {129277}, shortjournal = {Neurocomputing}, title = {A hybrid algorithm with inlier-guided hough voting for point cloud registration}, volume = {620}, year = {2025}, }
• (2025). BDMUIE: Underwater image enhancement based on bayesian diffusion model. NEUCOM, 620, 129274. (www)

  @article{ , abstract = {Existing deep learning-based underwater image enhancement (UIE) methods primarily rely on data-driven approaches that map underwater images to their enhanced counterparts. However, these methods often neglect the role of prior information in improving image quality. Such prior information, derived from natural domain images with balanced color distributions and rich structural details, can guide the enhancement process by reducing the domain gap between underwater and natural images. To address this limitation, we introduce BDMUIE, an innovative UIE framework based on Bayesian diffusion models. By incorporating prior knowledge from natural domain images, BDMUIE modifies the reverse diffusion process to effectively reduce the domain gap between underwater and natural images. A novel multi-scale fusion mechanism is introduced, uniquely leveraging encoder features across diffusion branches. This mechanism seamlessly integrates prior information from natural domain images with features from the underwater enhancement branch, resulting in improved detail preservation and color restoration. Additionally, the ECAG block enhances the model’s ability to remove noise while supporting the fusion process to retain structural and textural details. Extensive experiments on widely used underwater image datasets, as well as application experiments, demonstrate that BDMUIE achieves superior performance compared to state-of-the-art methods, with significant improvements in both visual quality and quantitative metrics, including PSNR and SSIM.}, archive = {J_NEUCOM}, author = {Lingfeng Chen and Zhihan Xu and Chao Wei and Yuanxin Xu}, doi = {10.1016/j.neucom.2024.129274}, journal = {Neurocomputing}, month = {3}, pages = {129274}, shortjournal = {Neurocomputing}, title = {BDMUIE: Underwater image enhancement based on bayesian diffusion model}, volume = {620}, year = {2025}, }
• (2025). Enhancing multimodal translation: Achieving consistency among visual information, source language and target language. NEUCOM, 620, 129269. (www)

  @article{ , abstract = {Multimodal machine translation refers to the task of using information from images, videos, etc., to assist in text translation. Numerous studies have demonstrated that incorporating visual representations as additional encoded information can enhance the dependency relations of contextual sequential vectors in autoregressive translation tasks. However, such methods typically only simulate the consistency between visual annotations and the source language, the contextual consistency among the source language, target language and visuals was not given enough attention. To fully utilize the consistent information, this paper introduces a novel multimodal neural machine translation approach. The encoder and decoder are integrated with visual features to ensure visual–textual consistency during the translation process. In the encoder, a cross-modal association mechanism utilizing graph neural networks is designed to effectively adapt textual and visual information. In the decoder, a structured multimodal graph is developed, directly mapping and merging visual and textual information. Experimental results indicate that, compared to the baseline, the method proposed in this paper significantly enhances the translation performance for English-German and English-French language pairs. Moreover, ablation experiments further validate the efficacy of this approach in improving translation quality.}, archive = {J_NEUCOM}, author = {Xiayang Shi and Xiaona Yang and Pei Cheng and Yingjie Zhou and Jingjing Liu}, doi = {10.1016/j.neucom.2024.129269}, journal = {Neurocomputing}, month = {3}, pages = {129269}, shortjournal = {Neurocomputing}, title = {Enhancing multimodal translation: Achieving consistency among visual information, source language and target language}, volume = {620}, year = {2025}, }
• (2025). Transformer-based monocular depth estimation with hybrid attention fusion and progressive regression. NEUCOM, 620, 129268. (www)

  @article{ , abstract = {Monocular depth estimation has gained prominence due to its versatile applications and minimal system requirements, despite the inherent challenges posed by its ill-posed nature. Recent advancements in Transformer techniques have significantly propelled this field forward. This paper introduces a novel Transformer-based encoder-decoder network to enhance monocular depth estimation, presenting two key improvements compared to prior approaches. Firstly, a hybrid attention fusion module is designed in the decoding stage for feature fusion. This module employs a parallel hybrid structure design to fully leverage the complementary advantages of Transformers and convolution methods, effectively capturing both local contextual and global consistency of depth maps. Additionally, to mitigate fine-grained depth loss during decoder upsampling, a deep progressive regression module is proposed. This module gradually predicts depth maps of varying resolutions by enhancing interactions between adjacent prediction branches. Experimental results on challenging indoor and outdoor benchmark datasets affirm the superior performance of the proposed method, showing that it effectively addresses some of the limitations of existing approaches.}, archive = {J_NEUCOM}, author = {Peng Liu and Zonghua Zhang and Zhaozong Meng and Nan Gao}, doi = {10.1016/j.neucom.2024.129268}, journal = {Neurocomputing}, month = {3}, pages = {129268}, shortjournal = {Neurocomputing}, title = {Transformer-based monocular depth estimation with hybrid attention fusion and progressive regression}, volume = {620}, year = {2025}, }
• (2025). Improved image denoising via self-supervised weickert operator learning and plug-and-play learned primal dual. NEUCOM, 620, 129267. (www)

  @article{ , abstract = {Hybrid models are a class of powerful techniques that have earned considerable interest recently, and this is because of their ability to handle complex and highly image-denoising tasks. This paper introduces a new hybrid model for image denoising, governed by a reaction–diffusion learning equation. We are interested in a self-supervised Weickert operator approximated by an artificial neural network approach. We express this inverse problem in the form of an optimal control approach; furthermore, the existence of a solution for the state and control problem was discussed. Representative experiments and comparisons are performed to ensure the effectiveness of the proposed model. This is reinforced by the plug-and-play learned Primal-Dual to achieve the denoising performance.}, archive = {J_NEUCOM}, author = {Imane El Malki and Abdelmajid El Hakoume and Amine Laghrib and Aissam Hadri}, doi = {10.1016/j.neucom.2024.129267}, journal = {Neurocomputing}, month = {3}, pages = {129267}, shortjournal = {Neurocomputing}, title = {Improved image denoising via self-supervised weickert operator learning and plug-and-play learned primal dual}, volume = {620}, year = {2025}, }
• (2025). Hypergraph convolutional network based weakly supervised point cloud semantic segmentation with scene-level annotations. NEUCOM, 620, 129264. (www)

  @article{ , abstract = {Point cloud segmentation with scene-level annotations is a promising but challenging task. Currently, the most popular way is to employ the class activation map (CAM) to locate discriminative regions and then generate point-level pseudo labels from scene-level annotations. However, these methods always suffer from the point imbalance among categories, as well as the sparse and incomplete supervision from CAM. In this paper, we propose a novel weighted hypergraph convolutional network-based method, called WHCN, to confront the challenges of learning point-wise labels from scene-level annotations. Firstly, in order to simultaneously overcome the point imbalance among different categories and reduce the model complexity, superpoints of a training point cloud are generated by exploiting the geometrically homogeneous partition. Then, a hypergraph is constructed based on the high-confidence superpoint-level seeds which are converted from scene-level annotations. Secondly, the WHCN takes the hypergraph as input and learns to predict high-precision point-level pseudo labels by label propagation. Besides the backbone network consisting of spectral hypergraph convolution blocks, a hyperedge attention module is learned to adjust the weights of hyperedges in the WHCN. Finally, a segmentation network is trained by these pseudo point cloud labels. Experimental results on the scanNet, S3DIS, Semantic3D, and ShapeNet Part benchmarks demonstrate that the proposed WHCN is effective to predict the point labels with scene annotations, which outperforms the state-of-the-art by 3.5% to 36.1% mIou. The source code is available at https://github.com/VCG-NJUST/WHCN .}, archive = {J_NEUCOM}, author = {Zhuheng Lu and Peng Zhang and Yuewei Dai and Weiqing Li and Zhiyong Su}, doi = {10.1016/j.neucom.2024.129264}, journal = {Neurocomputing}, month = {3}, pages = {129264}, shortjournal = {Neurocomputing}, title = {Hypergraph convolutional network based weakly supervised point cloud semantic segmentation with scene-level annotations}, volume = {620}, year = {2025}, }
• (2025). Neural observer-based fixed-time formation control of multiagent systems. NEUCOM, 620, 129263. (www)

  @article{ , abstract = {In this paper, the formation control problem for nonlinear multiagent systems (MASs) with immeasurable velocities in the directed communication topology is investigated under the constraints of external disturbances and dynamic uncertainties. To tackle these issues, a neural observer combining an extended state is constructed, utilizing fixed-time sliding mode and adaptive radial basis function (RBF) neural network to significantly enhance the estimation accuracy and convergence rate. Subsequently, a fixed-time formation control strategy is designed based on the feedback information from the neural observer. This strategy also incorporates a centralized event-triggered mechanism to prevent Zeno behavior and reduce the frequency of controllers’ updates. The global fixed-time convergence of the closed-loop MASs is proven through Lyapunov theory. Finally, comparative simulations involving six omnidirectional robots substantiate the effectiveness of the proposed control algorithm.}, archive = {J_NEUCOM}, author = {Zihang Guo and Shuangsi Xue and Junkai Tan and Huan Li and Dapeng Yan and Hui Cao}, doi = {10.1016/j.neucom.2024.129263}, journal = {Neurocomputing}, month = {3}, pages = {129263}, shortjournal = {Neurocomputing}, title = {Neural observer-based fixed-time formation control of multiagent systems}, volume = {620}, year = {2025}, }
• (2025). Source-free domain adaptation framework based on confidence constrained mean teacher for fundus image segmentation. NEUCOM, 620, 129262. (www)

  @article{ , abstract = {Unsupervised domain adaptation (UDA) has been gradually applied in fundus image segmentation, mitigating the challenge of insufficient data annotation by transferring pre-trained knowledge from source domain to target domain. However, UDA requires access to source data, which is usually restricted due to privacy and security concerns. Consequently, Source-Free Domain Adaptation (SFDA), particularly mean teacher-based methods, has garnered widespread attention. The mean teacher primarily includes two components: student model ( f s ), which updates parameters through backpropagation, and teacher model ( f t ), which is the exponential moving average (EMA) of f s . During this process, the EMA parameter heavily influences the efficiency and performance of f t . However, current methods employ fixed EMA parameter, i.e., f t can only learn knowledge from f s based on a fixed proportion, which affects the effectiveness of feature extraction. To address this issue, this study proposes a Confidence Constrained Mean Teacher (CCMT) framework, which dynamically adjusts the exponential moving parameter by measuring the outputs difference between f t and f s , enables the model to actively adjust the learning progress and direction, thereby enhancing the update efficiency of the model and improving its performance. Additionally, we introduce JS divergence to constrain the output distributions of f t and f s , ensuring their consistency, since f s relies on the outputs of f t as pseudo-labels. Experimental on two publicly available benchmark datasets demonstrate the advantages of CCMT in achieving accurate and robust segmentation of the fundus optic disc and optic cup in source-free settings, outperforming the state-of-the-art result.}, archive = {J_NEUCOM}, author = {Yanqin Zhang and Ding Ma and Xiangqian Wu}, doi = {10.1016/j.neucom.2024.129262}, journal = {Neurocomputing}, month = {3}, pages = {129262}, shortjournal = {Neurocomputing}, title = {Source-free domain adaptation framework based on confidence constrained mean teacher for fundus image segmentation}, volume = {620}, year = {2025}, }
• (2025). Novel source domain filtering dual classifier network with adaptive pseudo label refinement for partial domain fault diagnosis. NEUCOM, 620, 129260. (www)

  @article{ , abstract = {A novel source domain filtering dual classifier network with adaptive pseudo label refinement is proposed to address the issue of negative transfer in partial domain adaptation for fault diagnosis of rotating machinery. Existing partial domain adaptation methods often rely on weighting strategies to mitigate the impact of irrelevant source domain data, but these approaches suffer from two major issues: (1) Including all fault class samples during training leads to premature domain shifts, complicating subsequent fault diagnosis; (2) The effectiveness of weighting strategies heavily depends on the quality of pseudo labels, which is challenging to control. To overcome these challenges, a source domain filter is first designed to perform coarse filtering of irrelevant outliers using filtering coefficients, thereby preventing them from contaminating future training. A dual classifier module is then employed for more refined filtering, reducing the influence of outliers on the model. Finally, a pseudo label refinement method based on a truncated Gaussian distribution is proposed to improve the quality of pseudo labels. Experiments conducted on two test benches demonstrate that the proposed method surpasses state-of-the-art techniques by over 4% in accuracy, highlighting its superior capabilities in partial domain transfer.}, archive = {J_NEUCOM}, author = {Xiaolin Liu and Fuzheng Liu and Tongzhuo Han and Longqing Fan and Xiangyi Geng and Mingshun Jiang and Faye Zhang}, doi = {10.1016/j.neucom.2024.129260}, journal = {Neurocomputing}, month = {3}, pages = {129260}, shortjournal = {Neurocomputing}, title = {Novel source domain filtering dual classifier network with adaptive pseudo label refinement for partial domain fault diagnosis}, volume = {620}, year = {2025}, }
• (2025). Spatial and channel enhanced self-attention network for efficient single image super-resolution. NEUCOM, 620, 129258. (www)

  @article{ , abstract = {In recent years, the widespread adoption of the swin transformer model in computer vision has significantly enhanced the quality of super-resolution reconstructions. However, existing single-image super-resolution models, while effective in capturing spatial details, often overlook interactions across channels, limiting their ability to fully leverage the depth of feature information. This deficiency hampers their performance, leading to suboptimal reconstruction quality, as the models cannot adequately capture global dependencies across both spatial and channel dimensions. Consequently, these models tend to produce less accurate high-frequency details, which are critical for enhancing image clarity and preserving fine structures. To address this issue, we introduce the Spatial and Channel Enhanced Self-Attention Network (SCESN), designed to integrate both spatial and channel attention mechanisms, thereby enhancing feature utilization and delivering superior reconstruction performance. In this work, a global self-attention module (GSM) was designed to perform deep feature extraction by utilizing global contextual information in the images. Specifically, the GSM incorporates spatial self-attention block (SSB), channel self-attention block (CSB), and spatial and channel enhanced blocks (SCEB). SCEB boosts the interaction between spatial information from SSBs and channel information from CSBs, thereby maximizing feature utilization and enhancing network performance. Moreover, efficient convolution blocks are integrated to efficiently enhance the model’s deep spatial feature extraction capabilities. Quantitative evaluations across multiple test sets, along with visual comparisons, demonstrate that the proposed method achieves superior reconstruction results with fewer parameters compared to existing methodologies.}, archive = {J_NEUCOM}, author = {Xiaogang Song and Yuping Tan and Xinchao Pang and Lei Zhang and Xiaofeng Lu and Xinhong Hei}, doi = {10.1016/j.neucom.2024.129258}, journal = {Neurocomputing}, month = {3}, pages = {129258}, shortjournal = {Neurocomputing}, title = {Spatial and channel enhanced self-attention network for efficient single image super-resolution}, volume = {620}, year = {2025}, }
• (2025). Transformer-based neural architecture search for effective visible-infrared person re-identification. NEUCOM, 620, 129257. (www)

  @article{ , abstract = {Visible-infrared person re-identification (VI-reID) is a complex task in security and video surveillance that aims to identify and match a person captured by various non-overlapping cameras. In recent years, there has been a notable advancement in reID owing to the development of transformer-based architectures. Although many existing methods emphasize on learning both modality-specific and shared features, challenges remain in fully exploiting the complementary information between infrared and visible modalities. Consequently, there is still opportunity to increase retrieval performance by effectively comprehending and integrating cross-modality semantic information. These designs often have problems with model complexity and time-consuming processes. To tackle these issues, we employ a novel transformer-based neural architecture search (TNAS) deep learning approach for effective VI-reID. To alleviate modality gaps, we first introduce a global–local transformer (GLT) module that captures features at both global and local levels across different modalities, contributing to better feature representation and matching. Then, an efficient neural architecture search (NAS) module is developed to search for the optimal transformer-based architecture, which further enhances the performance of VI-reID. Additionally, we introduce distillation loss and modality discriminative (MD) loss to examine the potential consistency between different modalities to promote intermodality separation between classes and intramodality compactness within classes. Experimental results on two challenging benchmark datasets illustrate that our developed model achieves state-of-the-art results, outperforming existing VI-reID methods.}, archive = {J_NEUCOM}, author = {Prodip Kumar Sarker}, doi = {10.1016/j.neucom.2024.129257}, journal = {Neurocomputing}, month = {3}, pages = {129257}, shortjournal = {Neurocomputing}, title = {Transformer-based neural architecture search for effective visible-infrared person re-identification}, volume = {620}, year = {2025}, }
• (2025). Spatio-temporal knowledge distilled video vision transformer (STKD-VViT) for multimodal deepfake detection. NEUCOM, 620, 129256. (www)

  @article{ , abstract = {The widespread circulation of manipulated videos using deepfake techniques has raised concerns about the authenticity of multimedia content. In response, deepfake detection techniques have made significant strides in specific scenarios. However, most of the existing methods are unimodal and focus only on extracting traditional spatial features, due to which they struggle to accurately identify modern deepfakes. This work introduces the STKD-VViT model for detecting deepfakes across multiple modalities while employing spatiotemporal features. STKD-VViT combines the strengths of the Video vision transformer and the Vision transformer to process visual and audio streams. The Video vision transformer employs a multi-head attention mechanism and tubelet embedding to extract the video’s spatial and temporal features. Alternatively, the vision transformer extracts the salient features from the mel-spectrograms of audio files. Furthermore, STKD-VViT leverages the knowledge distillation technique to reduce the number of FLOPs and the model’s parameters. Experimental results on the benchmark FakeAVCeleb dataset demonstrate that STKD-VViT achieves a testing accuracy of 97.49% for video stream data, 98.65% for audio stream data and 96.0% when both streams are combined using score-level fusion, surpassing other state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Shaheen Usmani and Sunil Kumar and Debanjan Sadhya}, doi = {10.1016/j.neucom.2024.129256}, journal = {Neurocomputing}, month = {3}, pages = {129256}, shortjournal = {Neurocomputing}, title = {Spatio-temporal knowledge distilled video vision transformer (STKD-VViT) for multimodal deepfake detection}, volume = {620}, year = {2025}, }
• (2025). MCMTNet: Advanced network architectures for EEG-based motor imagery classification. NEUCOM, 620, 129255. (www)

  @article{ , abstract = {Brain–computer interface (BCI) technology converts electroencephalogram (EEG) signals into control commands to help patients with motor disorders, such as stroke and amyotrophic lateral sclerosis, interact with external devices and improve their quality of life. However, the decoding accuracy of motor imagery tasks remains a significant challenge for BCI. To address this problem, a new network model, MCMTNet, is proposed, which consists of a multi-domain convolution block, a multi-head attention block, and a temporal convolutional network block. The MCMTNet model directly takes raw EEG signals as input, eliminating the need for extensive preprocessing. The multi-domain convolution block effectively filters, extracts, and down-samples the original signal, the multi-head attention block enhances the inter-scale correlation of the EEG, and the temporal convolution network block improves the coherence between EEG signals. Experimental results show that the MCMTNet model achieves a classification accuracy of 86.37% on the BCI Competition IV 2a dataset, outperforming state-of-the-art models. To demonstrate the generalization ability of MCMTNet, a high gamma dataset is selected as a control, achieving a classification accuracy of 96.78%. These results indicate that MCMTNet performs well in motor imagery decoding tasks.}, archive = {J_NEUCOM}, author = {Yingjie Yang and Xiu Zhang and Xin Zhang and Changyi Yu}, doi = {10.1016/j.neucom.2024.129255}, journal = {Neurocomputing}, month = {3}, pages = {129255}, shortjournal = {Neurocomputing}, title = {MCMTNet: Advanced network architectures for EEG-based motor imagery classification}, volume = {620}, year = {2025}, }
• (2025). Towards real-time detection of underwater target with pruning lightweight deep learning method in side-scan sonar images. NEUCOM, 620, 129254. (www)

  @article{ , abstract = {Real-time detection of underwater targets through side-scan sonar (SSS) images is challenging due to the complex underwater environment and heavy computational load. This paper proposes a pruning lightweight deep learning network for real-time underwater target detection. A feature extraction backbone network combining the receptive field convolution and the omni-dimensional dynamic convolution is designed to extract the diversified features of sparse and unclear targets in SSS images effectively. Furthermore, a lightweight feature information aggregation network is developed. It consists of a lightweight feature aggregation architecture, a feature information filtering and capturing module, and an upsampling module. On this basis, reliable pruning and knowledge distillation algorithms are introduced to reduce further the computational burden and parameter complexity for real-time detection. Experimental studies are conducted on the shipwreck SSS images dataset and public acoustic images dataset, which show that the developed lightweight network achieves accurate and stable detection of underwater targets. On the shipwreck SSS images dataset, the average accuracy mAP50 of underwater target detection reached 87.6% by applying the proposed model, whereas the baseline model is 83.9%. The number of parameters of the network is reduced by 56.65% compared with the baseline model, and the floating-point operations (FLOPs) are reduced by 68.29%. In addition, the dataset includes 851 SSS images of shipwrecks is created and available at https://gitee.com/nwpu-r/underwater-shipwreck-dataset .}, archive = {J_NEUCOM}, author = {Rui Tang and Yimin Chen and Jian Gao and Yazhou Wang and Shaowen Hao}, doi = {10.1016/j.neucom.2024.129254}, journal = {Neurocomputing}, month = {3}, pages = {129254}, shortjournal = {Neurocomputing}, title = {Towards real-time detection of underwater target with pruning lightweight deep learning method in side-scan sonar images}, volume = {620}, year = {2025}, }
• (2025). Weighted tensor-based consistent anchor graph learning for multi-view clustering. NEUCOM, 620, 129253. (www)

  @article{ , abstract = {The anchor graph-based multi-view subspace clustering (MVSC) methods have shown promising performance when dealing with large-scale data. However, the consistent information is not fully explored when learning anchor graphs and exploring higher-order relationships, affecting the clustering performance. To this end, we propose a novel multi-view clustering (MVC) model called weighted tensor-based consistent anchor graph learning (WTCAGL). Specifically, we first learn the embedding anchor graphs by projecting the original data into the embedding space. Meanwhile, we learn the consistent anchor graph by concatenating features of different views, which can well capture the consistent relationship between different views to fully explore the cross-view consistent information. Furthermore, we impose the weighted low-rank tensor constraint on the embedding anchor graphs and the consistent anchor graph, which can capture the higher-order relationships between the embedding anchor graph and the consistent anchor graph. Experiments on multi-view datasets validate the effectiveness of the proposed WTCAGL method.}, archive = {J_NEUCOM}, author = {Guanghao Du and Peng Song and Yuanbo Cheng and Zhaowei Liu and Yanwei Yu and Wenming Zheng}, doi = {10.1016/j.neucom.2024.129253}, journal = {Neurocomputing}, month = {3}, pages = {129253}, shortjournal = {Neurocomputing}, title = {Weighted tensor-based consistent anchor graph learning for multi-view clustering}, volume = {620}, year = {2025}, }
• (2025). Different-layer control of robotic manipulators based on a novel direct-discretization RNN algorithm. NEUCOM, 620, 129252. (www)

  @article{ , abstract = {In this paper, the development of discrete-time recurrent neural network (RNN) algorithm is different from previous studies that require the derivation processes in the continuous-time environment. Specifically, the problem of discrete time-dependent different-layer control (DTDDLC) of robotic manipulators is investigated. To solve the DTDDLC problem, a discrete-time RNN algorithm based on direct-discretization technique is proposed, and thus named direct-discretization RNN algorithm for convenience. For comparative purposes, four indirect-discretization RNN algorithms are also developed to solve the same problem. Theoretical analyses and results manifest the convergence performance of the proposed direct-discretization RNN algorithm. Finally, numerical experiments based on a five-link planar robotic manipulator and a UR3 robotic manipulator substantiate the effectiveness and superiority of the proposed direct-discretization RNN algorithm.}, archive = {J_NEUCOM}, author = {Jinjin Guo and Zhanhao Xiao and Jianhua Guo and Xianglei Hu and Binbin Qiu}, doi = {10.1016/j.neucom.2024.129252}, journal = {Neurocomputing}, month = {3}, pages = {129252}, shortjournal = {Neurocomputing}, title = {Different-layer control of robotic manipulators based on a novel direct-discretization RNN algorithm}, volume = {620}, year = {2025}, }
• (2025). Gaussian-based R-CNN with large selective kernel for rotated object detection in remote sensing images. NEUCOM, 620, 129248. (www)

  @article{ , abstract = {Accurate object detection in remote sensing images is essential for applications such as environmental monitoring, urban planning, and disaster management. However, the detection of small rotated objects remains a major challenge, limiting advancements in remote sensing object detection. The performance of generic object detectors tends to drastically deteriorate on small rotated object detection tasks. Generic object detectors often struggle with tasks involving small rotated objects due to issues such as feature misalignment, angle periodicity, and metric-loss inconsistency, arising from arbitrary orientations and low-quality feature representations caused by their small sizes. In response to these obstacles, we propose a Gaussian-based R-CNN, providing a unified solution to the complexities of rotated object detection. Our approach integrates Oriented Region Proposal Network and Rotated Region of Interest Align to extract rotation-invariant features. Additionally, Gaussian Modeling is introduced to transform oriented bounding boxes into Gaussian distributions, effectively avoiding the regression of the angle parameters. Moreover, we introduce a Dynamic Center-Aligned Gaussian Product Loss (DCAGPL) to accurately measure the similarity between Gaussian distributions, approximating Rotated Intersection over Union loss. Furthermore, our method incorporates large selective kernel attention, dynamically adjusting its large spatial receptive field to enhance the modeling of the surrounding context for various objects, particularly crucial for the detection of small objects. Extensive experiments validate the effectiveness of our proposed methods. Without any additional tricks, the Gaussian-based R-CNN with ResNet-50 achieves state-of-the-art detection accuracy on two commonly used datasets for rotated object detection: DOTA (75.99% mAP) and HRSC2016-MS (68.02% mAP). Furthermore, when integrated with the large selective kernel mechanism, it achieves mAP scores of 76.77% and 71.95%, respectively, demonstrating its superior performance and versatility in addressing the challenges of detecting small rotated objects in remote sensing images.}, archive = {J_NEUCOM}, author = {Xiao Yang and Ahmad Sufril Azlan Mohamed}, doi = {10.1016/j.neucom.2024.129248}, journal = {Neurocomputing}, month = {3}, pages = {129248}, shortjournal = {Neurocomputing}, title = {Gaussian-based R-CNN with large selective kernel for rotated object detection in remote sensing images}, volume = {620}, year = {2025}, }
• (2025). Gauss-newton method conducted kernel adaptive nonlinear post-distorter with sparse structure and optimized width in visible light communication systems. NEUCOM, 620, 129247. (www)

  @article{ , abstract = {Visible light communication has been regarded as a viable supplement to wireless transmissions, but the nonlinear distortion induced by the light emitting diode limits the performance, so the impairment has to be eliminated, where the reproducing kernel Hilbert space (RKHS) schemes have found applications. While initially proposed from the second-order derivatives, Gauss-Newton (GN) algorithm employs the iterative approximation to the inverse of the Hessian matrix in Newton scheme to reduce the complexity. In this paper, aiming at mitigating the nonlinearity, we propose the RKHS kernel GN (KGN) post-distortion scheme, which promises high estimation accuracy and fast convergence. Since the implementation of KGN relies on the dictionary of observations, the computational complexity could be high with a large dictionary size, so the post-distorter with sparse structure is desirable. Therefore, we investigate the online surprise criterion to the post-distorter, and we also bring the randomized sketches (RS) approach to the post-distorter, developing the new KGN-RS method. An appropriate kernel width is crucial for the kernel-based methods, which largely limits the approximation ability. Aiming at finding the optimal width, we further investigate the hyperparameter optimization schemes, where particle swarm optimization, newton, Nelder-Mead, stochastic gradient, differential evolution and covariance matrix adaptation evolutionary strategy are firstly employed for optimizing the width. Numerical results show that the proposed RKHS distorter can converge faster than conventional polynomial based method, and supply better bit error rate performance.}, archive = {J_NEUCOM}, author = {Jieling Wang and Xinzhi Wang and Zihan Kang and Ba-Zhong Shen}, doi = {10.1016/j.neucom.2024.129247}, journal = {Neurocomputing}, month = {3}, pages = {129247}, shortjournal = {Neurocomputing}, title = {Gauss-newton method conducted kernel adaptive nonlinear post-distorter with sparse structure and optimized width in visible light communication systems}, volume = {620}, year = {2025}, }
• (2025). Temporal action localization with state-sensitive mamba and centroid sequences enhancement. NEUCOM, 620, 129246. (www)

  @article{ , abstract = {The temporal action localization task aims to identify and localize human behaviors in unedited videos. However, most previous studies have employed sampling processing and local information aggregation to reduce the computational burden, which results in significant information loss and a weak understanding of contextual information. In this paper, we improve localization accuracy by extending the sequence length and employing global information aggregation, which enhances the model’s ability to capture long-range dependencies and contextual information. To this end, we first propose a novel State-Sensitive Mamba module to replace the self-attention mechanism. Based on its linear computational cost and high sensitivity to state boundaries, we extend the sequence length to five times that of the previous model, which helps to preserve information integrity and leads to a more comprehensive information capture. Secondly, in order to overcome the noise and boundary ambiguity introduced by the extended sequences, we propose a centroid sequence enhancement strategy. This strategy optimizes boundary localization by using central features that are closer to the motion semantics. Finally, extensive experimental results on the THUMOS’14, ActivityNet1.3 and FineAction datasets demonstrate the superior performance of the proposed method.}, archive = {J_NEUCOM}, author = {Peng Wang and Shoupeng Lu and Cheng Dai and Shengxin Dai and Bing Guo}, doi = {10.1016/j.neucom.2024.129246}, journal = {Neurocomputing}, month = {3}, pages = {129246}, shortjournal = {Neurocomputing}, title = {Temporal action localization with state-sensitive mamba and centroid sequences enhancement}, volume = {620}, year = {2025}, }
• (2025). Consistency perception network for 360° omnidirectional salient object detection. NEUCOM, 620, 129243. (www)

  @article{ , abstract = {With the gradual popularization of panoramic cameras and the rapid development of computer vision technology, the research on salient object detection (SOD) in 360 ° omnidirectional images has attracted great attention. Different from traditional 2D images, 360 ° omnidirectional images generally have some drawbacks such as projection distortion, complex scenes, and small objects. To address these knotty issues, in this paper, we propose a novel C onsistency P erception Net work (CPNet) for 360 ° omnidirectional salient object detection, which can reliably localize salient regions in 360 ° omnidirectional images. We adopt a split-splice strategy on equirectangular 360 ° images to solve the distortion problem caused by the discontinuities of objects on the projection boundary, which restores the structure of objects and ensures the integrity of scene to the greatest extent. Inspired by the human visual perception mechanism, we deploy a bidirectional scale-aware module (BSM), which uses different convolutional dilation rates to simulate different receptive fields, performs hierarchical perceptual learning in series and conducts bidirectional guidance positioning in parallel. In addition, a consistent context-aware module (CCM) is designed to facilitate consistent learning of salient regions from different attention perspectives. It not only completely guarantees the global uniformity, but also accurately preserves the local details. Under the gradual feedback and guidance of coarse prediction maps, the edge sharpening module (ESM) can depict the details of salient objects more accurately, thereby generating high-quality saliency maps with clear boundaries. Extensive experiments on three public 360 ° SOD datasets demonstrate that the proposed CPNet achieves comparable and competitive performance in terms of effectiveness and efficiency when compared with the cutting-edge SOD methods.}, archive = {J_NEUCOM}, author = {Hongfa Wen and Zunjie Zhu and Xiaofei Zhou and Jiyong Zhang and Chenggang Yan}, doi = {10.1016/j.neucom.2024.129243}, journal = {Neurocomputing}, month = {3}, pages = {129243}, shortjournal = {Neurocomputing}, title = {Consistency perception network for 360° omnidirectional salient object detection}, volume = {620}, year = {2025}, }
• (2025). AMacP: An adaptive optimization algorithm for deep neural network. NEUCOM, 620, 129242. (www)

  @article{ , abstract = {Stochastic gradient-based optimizers are used to train convolutional neural networks (CNNs). Due to its adaptable momentum, the Adam optimizer has recently gained a lot of attention because it addresses the SGD’s dying gradient problem but ignores local gradient changes. In order to manage step size, angularGrad calculates a score based just on gradient angular values from prior iterations, and the diffGrad optimizer adjusts step size according to the distinction between the gradients in the current and the recent past. Existing optimizers, however, are still unable to effectively leverage optimization curve information, take advantage of trends in parameter updating to generate adaptive terms, or account for the 1st and 2nd moment for parameter updates. The novel aMacP optimizer proposed in this research takes into account the average of both momentums and also the average of two consecutive parameters for adaptively changing the step size. It is the first effort in the domain to use the average of two momentum variables (the exponential decay average (EDA) of past gradients and the EDA of prior squared gradients) as the denominator for parameter updating to enhance the optimization process. With a more precise step size, the optimization steps become smoother. Yet, rigorous tests comparing state-of-the-art approaches to benchmark datasets show that aMacP performs better. Additionally, it is demonstrated that aMacP performs consistently well when training CNN with various activation functions. Using the Rosenbrock function and Rastrigin function to assess the converged path, these appeared that the suggested aMacP optimizer provides superior results due to its assortment of frequently employed optimization techniques. Additionally, the suggested approach produces improved accuracy for a range of NLP applications and mAP for object detection.}, archive = {J_NEUCOM}, author = {Shubhankar Bhakta and Utpal Nandi and Chiranjit Changdar and Bachchu Paul and Tapas Si and Rajat Kumar Pal}, doi = {10.1016/j.neucom.2024.129242}, journal = {Neurocomputing}, month = {3}, pages = {129242}, shortjournal = {Neurocomputing}, title = {AMacP: An adaptive optimization algorithm for deep neural network}, volume = {620}, year = {2025}, }
• (2025). QT-TextSR: Enhancing scene text image super-resolution via efficient interaction with text recognition using a query-aware transformer. NEUCOM, 620, 129241. (www)

  @article{ , abstract = {Scene text image super-resolution (STISR) has obtained widespread attention in recent years due to its ability to enhance text recognition performance. Many previous methods proposed to incorporate text prior knowledge into the super-resolution architecture for reconstructing high-quality text images. However, these text priors are typically derived from pretrained text recognition models, and the inaccurate recognition feedback will hinder overall performance. In this paper, we propose a novel model, QT-TextSR, which promotes scene text image super-resolution by introducing efficient interaction with text recognition to release the inaccurate text feedback through a Query-aware Transformer. Specifically, QT-TextSR decomposes scene text image super-resolution and scene text recognition into different sets of queries within a Vision-Language Cooperation Module, explicitly modeling discriminative and interactive features between text recognition and text image super-resolution tasks. By employing two separate yet simultaneous projection heads on the corresponding features, QT-TextSR can recover the low-quality text image meanwhile obtain the recognition results. Additionally, to mitigate the limitations caused by recognition errors and enhance text structure preservation, we introduce a strong texture prior through self-supervised pre-training, leveraging visual cues more effectively. Experiments on public dataset, TextZoom demonstrate that our QT-TextSR significantly outperforms previous state-of-the-art methods in the metrics of Recognition Accuracy (68% v s . 65.5%), PSNR (22.51 v s . 22.10), and SSIM (0.7960 v s . 7930). The code for QT-TextSR is available at https://github.com/lcy0604/QT-TextSR .}, archive = {J_NEUCOM}, author = {Chongyu Liu and Qing Jiang and Dezhi Peng and Yuxin Kong and Jiaixin Zhang and Longfei Xiong and Jiwei Duan and Cheng Sun and Lianwen Jin}, doi = {10.1016/j.neucom.2024.129241}, journal = {Neurocomputing}, month = {3}, pages = {129241}, shortjournal = {Neurocomputing}, title = {QT-TextSR: Enhancing scene text image super-resolution via efficient interaction with text recognition using a query-aware transformer}, volume = {620}, year = {2025}, }
• (2025). A multi-source classification framework with invariant representation reconstruction for dual-target RSVP-BCI tasks in cross-subject scenario. NEUCOM, 620, 129239. (www)

  @article{ , abstract = {The Rapid Serial Visual Presentation (RSVP) is a widely used paradigm for target detection tasks in Brain-Computer Interface (BCI) by decoding Electroencephalogram (EEG) signals. One major issue concerns the time-consuming calibration in cross-subject scenarios, which worsens in dual-target RSVP-BCI tasks. A new method is desperately needed to detect two targets further with less calibration time. This paper proposed a novel framework named Cross-subject Invariant Representation Extraction-Targeted Stacked Convolutional Autoencoder (CS-IRE-TSCAE) based on reconstructing the invariant representation. After filtering the source subjects, the CS-TSCAE alleviates the subject-dependent effect by reconstructing the invariant representation generated by CS-IRE. It was validated on the ERP datasets from the BCI Controlled Robot Contest 2022. The experimental result showed that CS-IRE-TSCAE obtained the highest Recall, F1 and Average ACC with significant differences both in subject-dependent and inter-subject experiments. It demonstrated that CS-IRE-TSCAE achieved a higher classification performance for dual-target RSVP with less calibration time. Our framework drives the application development of target detection in RSVP-BCI by facilitating multiple target detection in cross-subject scenarios, which has practical significance, especially in fast-deployment scenarios.}, archive = {J_NEUCOM}, author = {Hongying Chen and Dan Wang and Meng Xu and Jiaming Chen and Yueqi Zhang and Yuanfang Chen}, doi = {10.1016/j.neucom.2024.129239}, journal = {Neurocomputing}, month = {3}, pages = {129239}, shortjournal = {Neurocomputing}, title = {A multi-source classification framework with invariant representation reconstruction for dual-target RSVP-BCI tasks in cross-subject scenario}, volume = {620}, year = {2025}, }
• (2025). CrossCode2Vec: A unified representation across source and binary functions for code similarity detection. NEUCOM, 620, 129238. (www)

  @article{ , abstract = {Code similarity detection identifies code by analyzing similarities in syntax, semantics, and structure, which includes types of tasks: source-to-source, binary-to-binary, and source-to-binary. Due to encoding and representation disparities between source and binary code, existing methods have mainly focused on individual tasks, without providing a universal solution. Additionally, current source-to-binary tasks only achieve one-to-one matching between source code and binary functions, neglecting the one-to-many relationship inherent between source code and its cross-compiled binaries. In this paper, we propose CrossCode2Vec, a unified framework for representing code in both source and binary functions, which aims to bridge the gap in original coding features and provide a standardized similarity measurement across three code similarity detection tasks. For source code and its corresponding compiled binary, we first design an enhanced Abstract Path Context data preprocessing method, construct an abstract syntax tree (AST) from both source code functions and decompiled binary functions, and implement the function embedding followed by the pre-trained Word2vec model. Then we propose a task-specific data sampling strategy. We establish a one-to-one correspondence between source and binary functions through symbol tables and create a one-to-many relationship between source functions and their cross-compiled binaries based on sampling rules. Finally, we employ a hierarchical LSTM-attention network to facilitate the representation and similarity measurement of functions. We conduct both extrinsic and intrinsic evaluations to confirm the effectiveness of CrossCode2Vec in code representation and code similarity tasks, validating its superiority in model architecture and data processing methods. CrossCode2Vec demonstrates stable and exceptional performance across multiple experiments, reinforcing its ability to bridge the gap between source and binary code representations while effectively measuring their similarities.}, archive = {J_NEUCOM}, author = {Gaoqing Yu and Jing An and Jiuyang Lyu and Wei Huang and Wenqing Fan and Yixuan Cheng and Aina Sui}, doi = {10.1016/j.neucom.2024.129238}, journal = {Neurocomputing}, month = {3}, pages = {129238}, shortjournal = {Neurocomputing}, title = {CrossCode2Vec: A unified representation across source and binary functions for code similarity detection}, volume = {620}, year = {2025}, }
• (2025). A local learning approach for sequential hybrid systems in time series forecasting. NEUCOM, 620, 129235. (www)

  @article{ , abstract = {Time series forecasting is an important task in several fields, such as finances, healthcare, government, and resources. Despite its importance, real-world data are often comprised of linear and nonlinear patterns that may not be mapped by a single forecasting method. In this study, a dynamic selection approach is developed for a sequential hybrid system to reduce the risk of selecting an inappropriate model, reduce variance error, and allow a better generalization to specific time patterns. This was achieved by first computing the linear prediction of the time series and the nonlinear prediction of its residue; then, creating a region of competence around each query point (pair of linear and nonlinear predictions to be combined) using KNN; and, finally, applying local learning in the defined region to combine the linear and nonlinear forecasts with a dynamically selected Kernel Ridge Regression model, to perform more accurate forecasts for a given test instance as it can be a promising alternative to improve accuracy in residual-based predictive systems. In order to test the proposed model’s performance, experiments with eleven other literature models and ten relevant datasets were conducted, and the results placed the proposed model as the best-performing one with a 1.10 mean rank and the lowest MSE values in nine out of the ten datasets used. For Pollution, Gas, Lake Erie, Electricity, Pigs, Red Wine, Car Sales, Sunspot, Traffic, and Nordic datasets, it obtained a 26.23% to 99.49% performance gain over ARIMA regarding MSE metric, and 6.99% to 99.72% over other tested models, with exception of a single database in which other model obtained the best performance.}, archive = {J_NEUCOM}, author = {Marie Chantelle Cruz Medina and João Fausto L. de Oliveira}, doi = {10.1016/j.neucom.2024.129235}, journal = {Neurocomputing}, month = {3}, pages = {129235}, shortjournal = {Neurocomputing}, title = {A local learning approach for sequential hybrid systems in time series forecasting}, volume = {620}, year = {2025}, }
• (2025). Cross dropout based dynamic learning for blind super resolution. NEUCOM, 620, 129234. (www)

  @article{ , abstract = {The current explicit modeling methods of single image super resolution (SISR) have limited utilization of implicit information within images, and are not effective for some complex degradation images. These often result in over-smoothed or visually uncomfortable artifacts. Meanwhile, kernel-based methods are sensitive to minor estimation errors. To alleviate these issues, in this paper we propose a cross dropout based dynamic network (CDDNet) for multi-degradation blind super resolution. Particularly, our proposed CDDNet models the degradation of low-resolution (LR) images using degradation weights as the global attention and includes a gyroscope structure as the local attention mechanism to further improve performance. It can be seen as performing a sparsity process on the feature map, which means that activation values at random locations are absent or weakened. And it can improve the ability of CDDNet to handle multiple deep degradation. In order to enhance the ability of Generative Adversarial Network (GAN) to handle complex two-stage degradation, we combine a GAN loss, an artif loss and a pixel loss to form an integrate loss suitable for complex blind super-resolution tasks. Experimental results show that our proposed CDDNet is effective, achieving higher reconstruction accuracy and acceptable quality of visual perception, with good results on both synthetic and real-world datasets. Code is available at https://github.com/Aries1213/CDDNet .}, archive = {J_NEUCOM}, author = {Yongsheng Dong and Hongjie Zhou and Lintao Zheng and Xiaohong Wang and Jinwen Ma}, doi = {10.1016/j.neucom.2024.129234}, journal = {Neurocomputing}, month = {3}, pages = {129234}, shortjournal = {Neurocomputing}, title = {Cross dropout based dynamic learning for blind super resolution}, volume = {620}, year = {2025}, }
• (2025). SD-CSMOTE: Over-sampling method based on SNN-DPC and improved SMOTE. NEUCOM, 620, 129233. (www)

  @article{ , abstract = {An over-sampling method, SD-CSMOTE, is proposed to address the problem of intra-class imbalance in data. First, the minority samples are clustered via the shared nearest neighbour-density peak clustering method. The sample density within each cluster is then determined using the kernel density estimation method. Thereafter total number of samples to be generated in each cluster is calculated. Finally, two samples are randomly selected from the cluster on the basis of the cluster centre, and a new sample is created at the centroid of the three samples. The proposed method is compared with 10 different sampling methods through extensive tests on three classifiers and 10 publicly imbalanced datasets. The results demonstrate that the proposed method achieves optimal F1 value, G-mean, and AUC on most of the datasets, and the Friedman rankings are observed to be optimal across multiple classifiers. It is confirmed that the proposed method performs better than other sampling methods in resolving the intra-class imbalance problem. This method also provides a new way to address issues such as small disjuncts and data imbalance.}, archive = {J_NEUCOM}, author = {He Ma and Xu Zhang and Mei Song and Yi Zhu and Wei-Chiang Hong}, doi = {10.1016/j.neucom.2024.129233}, journal = {Neurocomputing}, month = {3}, pages = {129233}, shortjournal = {Neurocomputing}, title = {SD-CSMOTE: Over-sampling method based on SNN-DPC and improved SMOTE}, volume = {620}, year = {2025}, }
• (2025). A photo cartoonization method based on text-to-image diffusion model. NEUCOM, 620, 129221. (www)

  @article{ , abstract = {In modern animation, background scenery such as complex buildings or elaborate structures requires a lot of time and effort to achieve a sense of realism and visual immersion. Various deep learning-based image cartoonization methods have been proposed to reduce these costs by transforming real images into high-quality animated scenes. However, these methods tend to oversimplify high-frequency patterns, producing flat and unrealistic images. Recent diffusion models trained using animated image-text datasets have shown good performance in generating high-quality animated images, but they cannot directly generate animated images from real images. In this paper, we propose a novel diffusion-based image-to-image cartoonization method using three lightweight adapters: cartoon-style adapter, color-structure adapter and semantic adapter. The cartoon-style adapter allows a single model to generate images in a variety of artistic styles. The color-structure adapter ensures that the overall shapes and color tones of the input image are preserved in the cartoonized image, while the semantic adapter ensures that semantic information of the input image for texture and detailed features are preserved. Through extensive experiments on animation backgrounds and real landscape datasets, we show that the proposed method can improve the FID score by up to 38 % and the CLIP-I score by up to 42 % compared to existing cartoonization methods.}, archive = {J_NEUCOM}, author = {Hwyjoon Jeon and Jonghwa Shim and Hyeonwoo Kim and Eenjun Hwang}, doi = {10.1016/j.neucom.2024.129221}, journal = {Neurocomputing}, month = {3}, pages = {129221}, shortjournal = {Neurocomputing}, title = {A photo cartoonization method based on text-to-image diffusion model}, volume = {620}, year = {2025}, }
• (2025). Navigating the artificial intelligence revolution in neuro-oncology: A multidisciplinary viewpoint. NEUCOM, 620, 129220. (www)

  @article{ , abstract = {This comprehensive review presents the role of artificial intelligence (AI), machine learning (ML), and deep learning (DL) in neuro-oncology, focussing their groundbreaking impact on the field. It gives a historical overview of AI's evolution in area of neuro-oncology, representing its progress from basic data management to high-level predictive analytics that now significantly makes personalized treatment approaches. Further, This study examines various ML based methods that emphasize brain tumor classification, genetic profiling, and individualized care plans proposals, which basically reflects the shift towards precision medicine. This paper also highlights, developments in DL particularly in the development of neuro-imaging analysis, which has significantly guided surgical planning and radiation therapy for improved patient outcomes. Through various case studies, the paper presents recent breakthroughs where AI interventions have noticeably enhanced the treatment of neuro-oncological conditions. Moreover, It also addresses the ethical considerations encompassing the use of automated decision-making tools, emphasizing the need for ethical governance in AI deployment. By integrating multidisciplinary perspectives, the study underscores the collaborative effort between clinicians, data scientists, and ethicists, which is very crucial for the ethical and effective implementation of AI in current clinical practice. The main aim of discussion is to equip neurooncologists/sergeons and doctors with understandings into sustainable AI adoption which ensures that AI's path in neuro-oncology aligns with both clinical needs for personalized medicine and ethical standards. This paper concludes by outlining considered methods and frameworks for the future integration of AI in neuro-oncology, predicting trends, and preparing to meet the challenges ahead.}, archive = {J_NEUCOM}, author = {Sanjay Saxena and Soumyaranjan Panda and Ekta Tiwari and Rajesh Singh and Mostafa M. Fouda and Mannudeep K. Kalra and Ketan Kotecha and Luca Saba and Jasjit S. Suri}, doi = {10.1016/j.neucom.2024.129220}, journal = {Neurocomputing}, month = {3}, pages = {129220}, shortjournal = {Neurocomputing}, title = {Navigating the artificial intelligence revolution in neuro-oncology: A multidisciplinary viewpoint}, volume = {620}, year = {2025}, }
• (2025). Improving intermediate reasoning in zero-shot chain-of-thought for large language models with filter supervisor-self correction. NEUCOM, 620, 129219. (www)

  @article{ , abstract = {Chain of Thought (CoT) prompting enables Large Language Models (LLMs) to generate detailed intermediate reasoning steps to solve problems, demonstrating excellent performance across various fields. However, when LLMs encounter complex problems, particularly in multi-step analysis scenarios, they can sometimes suffer from hallucinations, leading to errors in the intermediate results. Our experiments revealed that when the Chain of Thought (CoT) is set from only one perspective, LLMs are more prone to hallucinations and errors in their intermediate reasoning steps. Additionally, we found that the sentences composing the answers generated by LLMs are not equally important; their respective contributions to determining the correctness of an answer vary. To address these issues and enhance the accuracy of intermediate results, we propose a framework called FS-C (Filter Supervisor Self-Correction). This framework automatically analyzes the importance of sentences within the answer by leveraging our proposed Filter and Supervisor modules, enabling targeted self-correction of the generated answer. Experimental results show that FS-C significantly outperforms Zero-Shot-CoT on datasets involving arithmetic reasoning, symbolic reasoning, commonsense reasoning, scientific reasoning and reading comprehension. The implementation code for FS-C can be accessed at: https://github.com/Cansal7159/Filter-Supervisor-Self-Correction .}, archive = {J_NEUCOM}, author = {Jun Sun and Yiteng Pan and Xiaohu Yan}, doi = {10.1016/j.neucom.2024.129219}, journal = {Neurocomputing}, month = {3}, pages = {129219}, shortjournal = {Neurocomputing}, title = {Improving intermediate reasoning in zero-shot chain-of-thought for large language models with filter supervisor-self correction}, volume = {620}, year = {2025}, }
• (2025). Leveraging multi-time-span sequences and feature correlations for improved stock trend prediction. NEUCOM, 620, 129218. (www)

  @article{ , abstract = {Accurate stock trend prediction is critical for informed investment decisions and the stability of financial markets. However, existing methodologies often overlook fine-grained stock price volatility and fail to incorporate a comprehensive spectrum of technical indicators, inadequately capturing the complex interrelationships fundamental to technical analysis. This paper proposes MSFCE, a novel framework for stock market trend prediction that enhances feature correlations across multi-time-span sequences. Specifically, MSFCE designs a multi-scale feature encoder to capture both intraday and daily features, which are processed through a Transformer-based dimensionally adaptive encoder. Furthermore, the framework leverages higher-order interactions among technical indicators via a graph attention network, dynamically modeling their interdependencies to improve prediction robustness in dynamic markets. Extensive experiments on the SSE50 and CSI300 datasets demonstrate that MSFCE significantly outperforms existing state-of-the-art methods, consistently exhibiting superior performance across multiple test periods and market conditions. Its strong prediction accuracy and risk management suggest practical applicability in trading strategies, yielding significant excess returns in empirical backtests.}, archive = {J_NEUCOM}, author = {Yawen Li and Mengyu Zhuang and Junheng Wang and Jinyi Zhou}, doi = {10.1016/j.neucom.2024.129218}, journal = {Neurocomputing}, month = {3}, pages = {129218}, shortjournal = {Neurocomputing}, title = {Leveraging multi-time-span sequences and feature correlations for improved stock trend prediction}, volume = {620}, year = {2025}, }
• (2025). Content-activating for artistic style transfer with ambiguous sketchy content image. NEUCOM, 620, 129217. (www)

  @article{ , abstract = {The term “ambiguous sketchy image” refers to images comprising abstract lines. In such images, unclear semantics and weak semantics often occur. Stylizing such ambiguous sketchy images poses a challenge yet holds significance for certain individuals, such as children or those with drawing difficulties. However, prevailing state-of-the-art (SOTA) techniques heavily rely on semantic information, when an ambiguous sketchy image is utilized as content, these methods encounter the dilemma of collapse. This collapse arises from the disparity between the simplicity of the content constraints and the complexity of the style constraints which represented as imbalanced in algorithm design (called content-style imbalance). To address this issue, we propose a content-activating framework for style transfer (CAStyle) with ambiguous sketchy image. CAStyle composed a hierarchical structure, content-activating structure and a content-activating transformation. Specifically, the hierarchical structure facilitates the generation of style textures across various scales, thereby achieving multi-scale style rendition. The content-activating structure converts single-step style transfer into a multi-step process, effectively spreads the content-style imbalance uniformly across each diffusion step, preventing the occurrence of content-style imbalances. The content-activating transformation reinforces content constraints and alleviates excessive style constraints, ensuring a harmonious fusion of latent content and style. By employing CAStyle, the issue of content-style imbalance in sketchy images is significantly mitigated. In experiments, CAStyle has demonstrated commendable results, outperforming SOTA methods when applied to “ambiguous sketchy content images.” Additionally, we introduce the Sketchy Content Image for Style Transfer (SCIST) dataset, aiming to draw broader attention and foster further research.}, archive = {J_NEUCOM}, author = {Yinqi Chen and Yangting Zheng and Peiwen Li and Weijian Luo and Shuo Kang and Wenbin He}, doi = {10.1016/j.neucom.2024.129217}, journal = {Neurocomputing}, month = {3}, pages = {129217}, shortjournal = {Neurocomputing}, title = {Content-activating for artistic style transfer with ambiguous sketchy content image}, volume = {620}, year = {2025}, }
• (2025). Brainformer: Mimic human visual brain functions to machine vision models via fMRI. NEUCOM, 620, 129213. (www)

  @article{ , abstract = {Human perception plays a vital role in forming beliefs and understanding reality. A deeper understanding of brain functionality will lead to the development of novel deep neural networks. In this work, we introduce a novel framework named Brainformer, a straightforward yet effective Transformer-based framework, to analyze Functional Magnetic Resonance Imaging (fMRI) patterns in the human perception system from a machine-learning perspective. Specifically, we present the Multi-scale fMRI Transformer to explore brain activity patterns through fMRI signals. This architecture includes a simple yet efficient module for high-dimensional fMRI signal encoding and incorporates a novel embedding technique called 3D Voxels Embedding. Secondly, drawing inspiration from the functionality of the brain’s Region of Interest, we introduce a novel loss function called Brain fMRI Guidance Loss. This loss function mimics brain activity patterns from these regions in the deep neural network using fMRI data. This work introduces a prospective approach to transferring knowledge from human perception to neural networks. Our experiments demonstrate that leveraging fMRI information allows the machine vision model to achieve results comparable to State-of-the-Art methods in various image recognition tasks.}, archive = {J_NEUCOM}, author = {Xuan-Bac Nguyen and Xin Li and Pawan Sinha and Samee U. Khan and Khoa Luu}, doi = {10.1016/j.neucom.2024.129213}, journal = {Neurocomputing}, month = {3}, pages = {129213}, shortjournal = {Neurocomputing}, title = {Brainformer: Mimic human visual brain functions to machine vision models via fMRI}, volume = {620}, year = {2025}, }
• (2025). An interpretable bilateral detail optimization deep unfolding network for pansharpening. NEUCOM, 620, 129212. (www)

  @article{ , abstract = {Pansharpening aims to acquire high-resolution multispectral (HRMS) images for many remote sensing applications. Deep unfolding networks (DUNs) have achieved great success in recent years, combining the interpretability of model-based methods with the feature representation ability of deep learning-based methods. However, the existing DUNs suffer from spectral and spatial distortions due to underutilization of the intrinsic properties of source images and severe information loss in the signal flow. To address the above issues, we first propose two source image-guided detail priors to formulate a novel variational optimization model for multispectral pansharpening, which can alleviate information loss across stages and enhance the spectral and spatial fidelity of fusion images. Different from the existing DUNs for multispectral pansharpening, our optimization objective is spectral and spatial detail information, which is easier to optimize and boost the accuracy of the increased depth network. Then this model is solved by the half-quadratic splitting method and unfolded into an interpretable bilateral detail optimization deep unfolding network (BDODUN), which consists of reconstruction modules and prior modules corresponding to the iterative process. In addition, a novel loss function is proposed to enhance our model’s performance and adaptively weight the results of two branches, effectively mitigating both spectral and spatial distortions. Extensive experiments validate that our method is superior to other state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Yufei Ge and Xiaoli Zhang and Xiongfei Li and Siwei Ma}, doi = {10.1016/j.neucom.2024.129212}, journal = {Neurocomputing}, month = {3}, pages = {129212}, shortjournal = {Neurocomputing}, title = {An interpretable bilateral detail optimization deep unfolding network for pansharpening}, volume = {620}, year = {2025}, }
• (2025). Efficient image super resolution via mixed window and dimension interaction. NEUCOM, 620, 129211. (www)

  @article{ , abstract = {It is a difficult task to achieve an efficient balance between Single Image Super Resolution (SISR) model complexity and performance, since the model performance improvement is not that promising from simple stacking of complex modules based on ConvNet or Transformer. By analyzing the evolution of ConvNet and Vision Transformers, we give the following two key insights: (1) Employing long-range and multi-order modeling is important for performance enhancement. (2) Complementary frequency components process is desirable in the SISR task. Based on these, we propose an efficient SISR network called Mixed Window and Dimension Interaction Network (MWDIN). The key module Mixed Window and Dimension Interaction Block employs a parallel structure to capture wide range of window and global dimensional information, and fuses feature information by the bi-directional interaction. Dynamic Channel Aggregation is designed to reduce the redundancy of channel and spatial information. Numerous experiments on mainstream test datasets prove that MWDIN achieves excellent super-resolution results. The code is available at: https://github.com/wsy8086/MWDIN.}, archive = {J_NEUCOM}, author = {Shouyi Wang and Gang Liu and Xiao Liu and Xiangyu Liao and Chao Ren}, doi = {10.1016/j.neucom.2024.129211}, journal = {Neurocomputing}, month = {3}, pages = {129211}, shortjournal = {Neurocomputing}, title = {Efficient image super resolution via mixed window and dimension interaction}, volume = {620}, year = {2025}, }
• (2025). Class-aware graph siamese representation learning. NEUCOM, 620, 129209. (www)

  @article{ , abstract = {Currently, two issues exist in the field of graph Siamese representation learning. First, the strategies for positive sample selection often impose strict constraints on the candidate set, which may result in critical class information being overlooked. Second, traditional data augmentation strategies often alter the topological structure of graphs, which may impact the representations of nodes. Therefore, we propose a class-aware graph Siamese representation learning method named CGSRL to address these issues. Inspired by the Jaccard coefficient, we design a node structural similarity calculation method that considers neighborhood information and average degree. In the selection of positive samples, we comprehensively considered both feature similarity and structural similarity, allowing for the selection of positive samples belonging to the same class over a wider range. Additionally, we craft a feature augmentation method that preserves the raw class information. Specifically, CGSRL enforces a discriminator to extract high-level semantics representing class information from the raw features through carefully designed consistency constraints. In turn, a generator synthesizes augmented features with the same class information as the raw features based on these high-level semantics. To validate the superiority of our proposed method, we conducted experiments in node classification and node clustering on five real-world graph datasets. The results consistently demonstrate that our proposed CGSRL method outperforms state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Chengcheng Xu and Tianfeng Wang and Man Chen and Jun Chen and Zhisong Pan}, doi = {10.1016/j.neucom.2024.129209}, journal = {Neurocomputing}, month = {3}, pages = {129209}, shortjournal = {Neurocomputing}, title = {Class-aware graph siamese representation learning}, volume = {620}, year = {2025}, }
• (2025). MIMNet: Multi-interest meta network with multi-granularity target-guided attention for cross-domain recommendation. NEUCOM, 620, 129208. (www)

  @article{ , abstract = {Cross-domain recommendation (CDR) plays a critical role in alleviating the sparsity and cold-start problem and substantially boosting the performance of recommender systems. Existing CDR methods prefer to either learn a common preference bridge shared by all users or a personalized preference bridge tailored for each user to transfer user preference from the source domain to the target domain. Although these methods significantly improve the recommendation performance, there are still some limitations. First, these methods usually assume a user only has a unique interest, while ignoring the fact that a user may interact with items with different interest preferences. Second, they learn transformed preference representation mainly relies on the source domain signals, while neglecting the rich information available in the target domain. To handle these issues, in this paper, we propose a novel method named Multi-interest Meta Network with Multi-granularity Target-guided Attention (MIMNet) for cross-domain recommendation. To be specific, we employ the capsule network to learn user multiple interests in the source domain, which will be fed into a meta network to generate multiple interest-level preference bridges. Then, we transfer user representations from the source domain to the target domain based on these multi-interest bridges. In addition, we introduce both fine-grained and coarse-grained target signals to aggregate user transformed interest-level representations by incorporating a novel multi-granularity target-guided attention network. We conduct extensive experimental results on three real-world CDR tasks, and the results show that our proposed approach MIMNet consistently outperforms all baseline methods. The source code of MIMNet is released at https://github.com/marqu22/MIMNet .}, archive = {J_NEUCOM}, author = {Xiaofei Zhu and Yabo Yin and Li Wang}, doi = {10.1016/j.neucom.2024.129208}, journal = {Neurocomputing}, month = {3}, pages = {129208}, shortjournal = {Neurocomputing}, title = {MIMNet: Multi-interest meta network with multi-granularity target-guided attention for cross-domain recommendation}, volume = {620}, year = {2025}, }
• (2025). The abnormal delta rhythm mechanism in disorders of consciousness: Intrinsic neuronal properties and network dynamics. NEUCOM, 620, 129206. (www)

  @article{ , abstract = {The brain rhythm is created by the periodic physiological oscillations, facilitating the propagation of the neural information and maintaining the brain function. Disorders of consciousness (DOC) presents a typical rhythm transition from the alpha rhythm (9–12 Hz) to delta rhythm (0.5–4 Hz). However, the underlying rhythm transition mechanism to reveal the pathological mechanism of DOC from the neural and network levels remains unclear. We establish a conductance-based computational network model for thalamo-cortical circuit, and examine thoroughly the effect of membrane capacitance, time constant of T-type calcium current, noise in brain and the synaptic loss on generating of delta brain rhythm. Simulation result shows the neuronal intrinsic properties can vary the neural firing patterns and further affect the brain rhythm. At the neuronal level, the enhanced membrane capacitance and the enlarged time constant of T-type calcium current activation gate variable make the firing pattern switch from the tonic to burst mode. The emergence of burst mode induces the delta rhythm, manifesting as the generation of cyclic saddle-node bifurcation at the dynamic activity. At the network level, the decreased brain noise intensity and the attenuated synaptic weight also play an important role in transmitting brain rhythm from alpha to delta band. The hybrid impact of these two factors with Our findings pave the way for exploring the potential mechanism underlying the generation of abnormal delta rhythm in DOC from the neuronal to network perspectives and provide theoretical basis for possible neural basis of DOC. This work lay the foundation for the stimulation therapy mechanism for DOC.}, archive = {J_NEUCOM}, author = {Jixuan Wang and Bin Deng and Jiang Wang and Yi Yang and Chen Liu}, doi = {10.1016/j.neucom.2024.129206}, journal = {Neurocomputing}, month = {3}, pages = {129206}, shortjournal = {Neurocomputing}, title = {The abnormal delta rhythm mechanism in disorders of consciousness: Intrinsic neuronal properties and network dynamics}, volume = {620}, year = {2025}, }
• (2025). Missing signal reconstruction and aileron fault detection via generative adversarial learning. NEUCOM, 620, 129205. (www)

  @article{ , abstract = {Aileron faults must be detected and identified quickly to improve the fixed-wing aircraft’s safety and reliability. However, the low-frequency monitoring signals of the heterogeneous sensors are easily lost in transmission and sampling. Therefore, incomplete data will overlap fault features, leading to detection mistakes. This study proposes a generative adversarial learning framework to reconstruct missing signals based on the rest signals called Dual-hierarchical-variational Graph Autoencoder/ Generative Adversarial Network (DHVGAE/GAN). A pluggable reconstruction module is designed to enhance the generalization performance of the detection model. A dual spatiotemporal encoder (DSE) is constructed by a local graph attention network (LGAT) and Transformer-based feature extractors (TFEs) to extract common spatiotemporal features of missing and rest signals. A shared hierarchical variational autoencoder (HVAE) maps them to multiple conditionally independent latent distributions and generates reconstructed signals. The historical prior information of missing signals is transferred to the reconstruction module through adversarial learning. The detection model maintains high accuracy for data with missing signals relying on the recovered signals. The experimental results of the pre-launch and flight datasets verify the feasibility and effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Yi He and Lifu Du and Wei Chen and Fuyang Chen and Yuntao Xu}, doi = {10.1016/j.neucom.2024.129205}, journal = {Neurocomputing}, month = {3}, pages = {129205}, shortjournal = {Neurocomputing}, title = {Missing signal reconstruction and aileron fault detection via generative adversarial learning}, volume = {620}, year = {2025}, }
• (2025). ARNN: Attentive recurrent neural network for multi-channel EEG signals to identify epileptic seizures. NEUCOM, 620, 129203. (www)

  @article{ , abstract = {Electroencephalography (EEG) is a widely used tool for diagnosing brain disorders due to its high temporal resolution, non-invasive nature, and affordability. Manual analysis of EEG is labor-intensive and requires expertise, making automatic EEG interpretation crucial for reducing workload and accurately assessing seizures. In epilepsy diagnosis, prolonged EEG monitoring generates extensive data, often spanning hours, days, or even weeks. While machine learning techniques for automatic EEG interpretation have advanced significantly in recent decades, there remains a gap in its ability to efficiently analyze large datasets with a balance of accuracy and computational efficiency. To address the challenges mentioned above, an Attention Recurrent Neural Network (ARNN) is proposed that can process a large amount of data efficiently and accurately. This ARNN cell recurrently applies attention layers along a sequence and has linear complexity with the sequence length and leverages parallel computation by processing multi-channel EEG signals rather than single-channel signals. In this architecture, the attention layer is a computational unit that efficiently applies self-attention and cross-attention mechanisms to compute a recurrent function over a wide number of state vectors and input signals. This framework is inspired in part by the attention layer and long short-term memory (LSTM) cells, but it scales this typical cell up by several orders to parallelize for multi-channel EEG signals. It inherits the advantages of attention layers and LSTM gate while avoiding their respective drawbacks. The model’s effectiveness is evaluated through extensive experiments with heterogeneous datasets, including the CHB-MIT and UPenn and Mayo’s Clinic datasets. The empirical findings suggest that the proposed model outperforms baseline methods such as LSTM, Vision Transformer (ViT), Compact Convolution Transformer (CCT), and R-Transformer (RT), showcasing superior performance with faster processing capabilities across EEG datasets. The code has been made publicly accessible at https://github.com/Salim-Lysiun/ARNN .}, archive = {J_NEUCOM}, author = {Salim Rukhsar and Anil K. Tiwari}, doi = {10.1016/j.neucom.2024.129203}, journal = {Neurocomputing}, month = {3}, pages = {129203}, shortjournal = {Neurocomputing}, title = {ARNN: Attentive recurrent neural network for multi-channel EEG signals to identify epileptic seizures}, volume = {620}, year = {2025}, }
• (2025). SE-MSResNet: A lightweight squeeze-and-excitation multi-scaled ResNet with domain generalization for sleep apnea detection. NEUCOM, 620, 129201. (www)

  @article{ , abstract = {Sleep apnea (SA) is a common contributing factor to many sleep-related and cardiovascular diseases. The non-invasive sleep apnea monitoring method based on a single-lead Electrocardiogram (ECG) is essential for its prevention and treatment. Although previous attempts to detect SA have achieved high classification performance, two challenges still need to be addressed: (1) the non-stationary nature of ECG signals makes it difficult to effectively design a deep feature-extracting algorithm to extract typical sleep apnea-related features from ECG segments. (2) generalizing deep learning-based SA detection models in real clinical scenarios is challenging due to the individual differences among subjects. This paper presents a lightweight squeeze-and-excitation multi-scaled ResNet model named SE-MSResNet with domain generalization for SA detection. First, our work introduces a multi-scaled residual block (MSResBlock) to extract temporal features from RR intervals and R-peak amplitude sequences derived from single-channel ECG signals. Next, a squeeze-and-excitation attention module is applied to fuse the features extracted from the two above ECG-derived signals. Finally, we propose a jointly shared feature training strategy based on domain adversarial generalization and form a unified training framework. It aims to minimize the feature distribution discrepancy between the source domains (training subjects) and enhance common sleep apnea-related features, thus improving the proposed model’s generalization performance for unseen subjects. Experimental results on the publicly available dataset validate the effectiveness of our proposed model, surpassing the performance of state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Yuxuan Zhao and Haitao He and Qian Wang and Lu Yu and Jiadong Ren}, doi = {10.1016/j.neucom.2024.129201}, journal = {Neurocomputing}, month = {3}, pages = {129201}, shortjournal = {Neurocomputing}, title = {SE-MSResNet: A lightweight squeeze-and-excitation multi-scaled ResNet with domain generalization for sleep apnea detection}, volume = {620}, year = {2025}, }
• (2025). CoFM: Alternating convolution and frequency module is also strong for visual recognition. NEUCOM, 620, 129200. (www)

  @article{ , abstract = {In recent years, Transformer-based and CNN-Transformer models have achieved remarkable success in various vision tasks. There have been both multi-head self-attention (MHSA) based models similar to the original Vision Transformer (ViT) and Fourier transform (FT) based models. Some studies suggest that combining CNNs and ViT can achieve a good balance between performance and computational cost, and also demonstrate that spectral layers can serve as an effective replacement to MHSA layers. In this paper, we investigate both CNNs and spectral layers provides outstanding performance. We thus propose a new CNN-based network based on CNNs and FT. We constructed the Discrete Fourier Transform Module (DFTM) with Convolutional Spectrum Encoding (CSE) and Spectrum Self-Calibration (SSC) to learn long-range dependencies and achieve global information mixing. Our CoFM-B achieves 84.8% top-1 accuracy on ImageNet 1K, surpassing Swin, CSwin, ConvNeXt and other state-of-the-art models, while maintaining lower computational costs and FLOPs. It is worth mentioning that out CoFM-B achieves 52.9% box AP and 46.0% mask AP on MS-COCO, outperforming ConvNeXt-B and Swin-B by 0.2% and 1.0% in term of box AP and 0.4% and 1.0% in term of mask AP respectively, additionally, achieves 50.2% mIoU on ADE20K, which is also better than ConvNeXt-B and Swin-B as the backbone.}, archive = {J_NEUCOM}, author = {Jiayong Bao and Jiangshe Zhang and Chunxia Zhang}, doi = {10.1016/j.neucom.2024.129200}, journal = {Neurocomputing}, month = {3}, pages = {129200}, shortjournal = {Neurocomputing}, title = {CoFM: Alternating convolution and frequency module is also strong for visual recognition}, volume = {620}, year = {2025}, }
• (2025). Self-supervised visual learning in the low-data regime: A comparative evaluation. NEUCOM, 620, 129199. (www)

  @article{ , abstract = {Self-Supervised Learning (SSL) is a valuable and robust training methodology for contemporary Deep Neural Networks (DNNs), enabling unsupervised pretraining on a ‘pretext task’ that does not require ground-truth labels/annotation. This allows efficient representation learning from massive amounts of unlabeled training data, which in turn leads to increased accuracy in a ‘downstream task’ by exploiting supervised transfer learning. Despite the relatively straightforward conceptualization and applicability of SSL, it is not always feasible to collect and/or to utilize very large pretraining datasets, especially when it comes to real-world application settings. In particular, in cases of specialized and domain-specific application scenarios, it may not be achievable or practical to assemble a relevant image pretraining dataset in the order of millions of instances or it could be computationally infeasible to pretrain at this scale, e.g., due to unavailability of sufficient computational resources that SSL methods typically require to produce improved visual analysis results. This situation motivates an investigation on the effectiveness of common SSL pretext tasks, when the pretraining dataset is of relatively limited/constrained size. This work briefly introduces the main families of modern visual SSL methods and, subsequently, conducts a thorough comparative experimental evaluation in the low-data regime, targeting to identify: (a) what is learnt via low-data SSL pretraining, and (b) how do different SSL categories behave in such training scenarios. Interestingly, for domain-specific downstream tasks, in-domain low-data SSL pretraining outperforms the common approach of large-scale pretraining on general datasets. Grounded on the obtained results, valuable insights are highlighted regarding the performance of each category of SSL methods, which in turn suggest straightforward future research directions in the field.}, archive = {J_NEUCOM}, author = {Sotirios Konstantakos and Jorgen Cani and Ioannis Mademlis and Despina Ioanna Chalkiadaki and Yuki M. Asano and Efstratios Gavves and Georgios Th. Papadopoulos}, doi = {10.1016/j.neucom.2024.129199}, journal = {Neurocomputing}, month = {3}, pages = {129199}, shortjournal = {Neurocomputing}, title = {Self-supervised visual learning in the low-data regime: A comparative evaluation}, volume = {620}, year = {2025}, }
• (2025). TACL: A trusted action-enhanced curriculum learning approach to multimodal affective computing. NEUCOM, 620, 129195. (www)

  @article{ , abstract = {Previous studies on Multimodal Affective Computing (MAC) predominantly focus on leveraging language, acoustic, and facial information to identify human’s affective states, which largely ignore the dynamic and temporal action information, despite such information being crucial for precisely inferring the affective states. In this way, this paper first attempts to consider the action information for MAC and further argues that exploiting the action information faces two key challenges, i.e., credibility and sparsity challenges. To this end, this paper proposes a new T rusted A ction-enhanced C urriculum L earning (TACL) approach to incorporate the action information for boosting MAC. Specifically, this approach designs two main components, i.e., the Trusted Curriculum Learning block and the Action-enhanced Vision Regulator, to address the above credibility challenge and sparsity challenge. Furthermore, a high-quality action-enhanced video dataset is constructed to evaluate TACL and detailed evaluations show the great advantage of TACL over the state-of-the-art baselines. Particularly, an interesting finding is observed that action information is more conducive to facilitating the recognition of negative emotions, which aligns with the intuition that humans prefer using actions more when expressing negative emotions.}, archive = {J_NEUCOM}, author = {Tan Yu and Jingjing Wang and Jiamin Luo and Jiawen Wang and Guodong Zhou}, doi = {10.1016/j.neucom.2024.129195}, journal = {Neurocomputing}, month = {3}, pages = {129195}, shortjournal = {Neurocomputing}, title = {TACL: A trusted action-enhanced curriculum learning approach to multimodal affective computing}, volume = {620}, year = {2025}, }
• (2025). Dynamic weight reinforcement learning method considering multiple factors in mobile edge computing system. NEUCOM, 620, 129194. (www)

  @article{ , abstract = {The accelerated advancement of Mobile Edge Computing (MEC) has facilitated significant progressions in digital medical diagnostic services. However, the multi-region, multi-user, and multi-priority characteristics in medical diagnostic services within MEC systems make the task of offloading problematic. On the one hand, the task offloading problem involves multiple conflicting objectives, such as minimizing delay costs, server load balancing, and user fairness. Some researchers have explored these issues and proposed reinforcement learning-based methods to tackle such problems. However, existing work often characterizes objectives as linear scalarizations of multiple objectives, overlooking their conflicts. On the other hand, the preferences for various objectives in different regions cannot be predicted in advance and may indeed differ, making it challenging to handle in existing research. To address these challenges, we propose a multi-objective, multi-agent reinforcement learning approach. In this approach, the reward at each step is a vector, with each scalar corresponding to an objective. Furthermore, we propose a multi-agent tournament selection method to identify historically significant preferences. This mechanism considers the strategies of other agents while preserving the policies previously learned by the current agent. The objective is to achieve cooperative scheduling, allowing agents to synchronize their decisions based on their historical preferences and those of other agents. Simulation results demonstrate that the proposed algorithm outperforms several baseline algorithms across various performance metrics.}, archive = {J_NEUCOM}, author = {Shihua Li and Yanjie Zhou and Xiangqian Liu and Ning Wang and Junqi Wang and Bing Zhou and Zongmin Wang}, doi = {10.1016/j.neucom.2024.129194}, journal = {Neurocomputing}, month = {3}, pages = {129194}, shortjournal = {Neurocomputing}, title = {Dynamic weight reinforcement learning method considering multiple factors in mobile edge computing system}, volume = {620}, year = {2025}, }
• (2025). Multi-factor embedding GNN-based traffic flow prediction considering intersection similarity. NEUCOM, 620, 129193. (www)

  @article{ , abstract = {Existing studies on traffic flow prediction primarily rely on on-board devices to collect vehicle trajectory data, which can potentially infringe upon the privacy of users and limit the applicability of the method. Additionally, traffic flow prediction remains challenging due to the complex spatial and temporal dependencies within real-world traffic networks. To address these limitations, this paper introduces a framework for analyzing discrete vehicle trajectory data at urban intersections. By incorporating various external physical factors into traffic flow prediction, this framework derives embedding vectors from vehicle trajectory sequences and road network topology, modeling their spatio-temporal dependencies using Skip-Gram and GraphSAGE, respectively. Additionally, the intersection similarity is introduced to capture and integrate traffic flow patterns between the target intersection and similar intersections. A Spatio-Temporal Graph Convolutional Neural Network (ST-GCN) algorithm, which combines Graph Convolutional Networks (GCN) with Long Short-Term Memory (LSTM), is developed to achieve precise traffic flow prediction. Extensive experiments on a real-world traffic flow dataset from Qingdao, China, validate that the proposed method outperforms state-of-the-art baseline methods.}, archive = {J_NEUCOM}, author = {Ruirui Zhong and Bingtao Hu and Fei Wang and Yixiong Feng and Zhiwu Li and Xiuju Song and Yong Wang and Shanhe Lou and Jianrong Tan}, doi = {10.1016/j.neucom.2024.129193}, journal = {Neurocomputing}, month = {3}, pages = {129193}, shortjournal = {Neurocomputing}, title = {Multi-factor embedding GNN-based traffic flow prediction considering intersection similarity}, volume = {620}, year = {2025}, }
• (2025). Domain adaptation for textual adversarial defense via prompt-tuning. NEUCOM, 620, 129192. (www)

  @article{ , abstract = {Most existing systems based on deep learning are vulnerable to adversarial attacks, i.e., strategically modified samples. These samples are generated by some imperceptible perturbations on different levels (sentence/word/character), which can fool deep models to give erroneous predictions. Recently, fine-tuning Pre-trained Language Models (PLMs) methods have achieved tremendous success in a series of downstream Natural Language Processing (NLP) tasks. However, recent studies indicate these methods need extensive adversarial examples for fine-tuning. In this paper, we propose a novel domain adaptation method for textual adversarial defense via prompt-tuning. Our method simulates original and adversarial text using source and target domains, then employs a prompt-tuning model to minimize domain deviations, treating textual adversarial defense as a cross-domain text classification task. To avoid noise and high time complexity from introducing external knowledge, we extract knowledge from the target domain and use hierarchical clustering and optimization strategies to fine-tune expansion words for the verbalizer in prompt-tuning. Experimental results show that the obvious improvement is obtained compared with other SOTA PLMs methods on textual adversarial defense.}, archive = {J_NEUCOM}, author = {Zhenglong Li and Yi Zhu and Chenqi Hua and Yun Li and Yunhao Yuan and Jipeng Qiang}, doi = {10.1016/j.neucom.2024.129192}, journal = {Neurocomputing}, month = {3}, pages = {129192}, shortjournal = {Neurocomputing}, title = {Domain adaptation for textual adversarial defense via prompt-tuning}, volume = {620}, year = {2025}, }
• (2025). Markowitz random forest: Weighting classification and regression trees with modern portfolio theory. NEUCOM, 620, 129191. (www)

  @article{ , abstract = {Tree-based ensembles such as random forest (RF) are essential methods for supervised learning. Whereas traditional RFs assign equal weights to their trees, significant evidence suggests that tree weighting schemes can enhance predictive performance. Previous works have focused solely on the predictive performance of individual trees, assigning greater weights to high-performing trees. However, the predictive power of RF arises not only from high performing trees but also from tree variety, a factor that has not been considered before. In this paper we propose Markowitz RF, a tree weighting method that considers both tree performance and variety, using a tree covariance matrix for risk regularization. Our method is formulated as an adapted optimization process inspired by financial mathematics. It can be applied to binary and multi-class classification, as well as regression tasks. Our experiments on 15 benchmark datasets indicate that MRF can significantly outperform previously proposed tree weighting approaches and other learning methods in terms of Precision-Recall AUC, mean absolute error and F 1 _ m a c r o .}, archive = {J_NEUCOM}, author = {Eleftherios Kouloumpris and Ioannis Vlahavas}, doi = {10.1016/j.neucom.2024.129191}, journal = {Neurocomputing}, month = {3}, pages = {129191}, shortjournal = {Neurocomputing}, title = {Markowitz random forest: Weighting classification and regression trees with modern portfolio theory}, volume = {620}, year = {2025}, }
• (2025). Understanding LLMs: A comprehensive overview from training to inference. NEUCOM, 620, 129190. (www)

  @article{ , abstract = {The introduction of ChatGPT has led to a significant increase in the utilization of Large Language Models (LLMs) for addressing downstream tasks. There is an increasing focus on cost-efficient training and deployment within this context. Low-cost training and deployment of LLMs represent the future development trend. This paper reviews the evolution of LLMs training techniques and inference deployment technologies aligned with this emerging trend. The objective is to provide researchers with a guide for integrating LLMs into their work. The discussion on training includes various aspects, including data preprocessing, training architecture, pre-training tasks, parallel training, and relevant content related to model fine-tuning. On the inference side, the paper covers topics such as model compression, parallel computation, memory scheduling, and structural optimization. It also explores LLMs’ utilization and provides insights into their future development.}, archive = {J_NEUCOM}, author = {Yiheng Liu and Hao He and Tianle Han and Xu Zhang and Mengyuan Liu and Jiaming Tian and Yutong Zhang and Jiaqi Wang and Xiaohui Gao and Tianyang Zhong and Yi Pan and Shaochen Xu and Zihao Wu and Zhengliang Liu and Xin Zhang and Shu Zhang and Xintao Hu and Tuo Zhang and Ning Qiang and Tianming Liu and Bao Ge}, doi = {10.1016/j.neucom.2024.129190}, journal = {Neurocomputing}, month = {3}, pages = {129190}, shortjournal = {Neurocomputing}, title = {Understanding LLMs: A comprehensive overview from training to inference}, volume = {620}, year = {2025}, }
• (2025). Potential distribution adjustment and parametric surrogate gradient in spiking neural networks. NEUCOM, 620, 129189. (www)

  @article{ , abstract = {In recent years, there has been significant research done on the use of surrogate gradients (SG) in spiking neural networks (SNNs) to deal with the challenges of network training with spiking activation functions that are non-differentiable at the threshold and have zero-valued derivatives for other membrane potentials. Due to the lack of well-recognized SG selection rules, most SGs are usually chosen intuitively. As such, the issue of whether a SG is optimal or not remains an open question. To facilitate adjustment of the SG for efficient training, we propose a parametric surrogate gradient (PSG) strategy which can be iteratively updated. Through parametric control and calibration of the SG, one can derive appropriate SG for SNNs training. It is commonly known that neural potential distribution tends to deviate unpredictably due to quantization error. This paper will further evaluate such potential shift and propose methodology for potential distribution adjustment (PDA) to minimize the loss of undesired pre-activations. Experimental results demonstrate that the proposed methods exhibit excellent computational efficiency and robustness to hardware noise. These methods can be readily integrated with backpropagation through time (BPTT) algorithm and help modulated SNNs to achieve state-of-the-art performance on various dataset with fewer timesteps.}, archive = {J_NEUCOM}, author = {Siqi Wang and Tee Hiang Cheng and Meng-Hiot Lim}, doi = {10.1016/j.neucom.2024.129189}, journal = {Neurocomputing}, month = {3}, pages = {129189}, shortjournal = {Neurocomputing}, title = {Potential distribution adjustment and parametric surrogate gradient in spiking neural networks}, volume = {620}, year = {2025}, }
• (2025). Uncertainty embedding of attribute networks based on multi-view information fusion and multi-order proximity preservation. NEUCOM, 620, 129188. (www)

  @article{ , abstract = {The attribute network embedding process aims to define the low-dimensional continuous vector representation of nodes by fusing the network structure view and the node attribute view, and to preserve the multi-order proximity information of the network in the embedded latent space. However, attribute network contains multi-view information with nonlinear relationships, uncertainty in node representation caused by complex interactions between nodes, and multi-order proximity information in the network structure and node attributes. The reported research has not comprehensively considered the mentioned problems, resulting in poor embedding effect of attribute network. This paper proposes a framework of a ttribute n etwork u ncertainty e mbedding based on multi- v iew i nformation f usion and multi-order p roximity p reservation (ANUE-VIFPP). ANUE-VIFPP comprehensively extracts the information of network structure view, attribute view and structure-attribute view, which uses the combination of self-attention mechanism and nonlinear layer to fuse the highly nonlinear coupled multi-view information effectively. The normalization flow (NF) algorithm is introduced into variational auto-encoders (VAE) to establish a flexible distribution for characterizing the uncertainty of node representation. A multi-conditional loss function constraint model is developed to effectively preserve the low and high order proximity information of network structure and node attributes in the embedded hidden space. Finally, experimental results obtained on three commonly used real-world datasets demonstrate that ANUE-VIFPP performs well in node classification, link prediction, and network visualization tasks, outperforming some state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Xin Yang and Xiangang Cao and Jiangbin Zhao and Yong Duan and Fuyuan Zhao}, doi = {10.1016/j.neucom.2024.129188}, journal = {Neurocomputing}, month = {3}, pages = {129188}, shortjournal = {Neurocomputing}, title = {Uncertainty embedding of attribute networks based on multi-view information fusion and multi-order proximity preservation}, volume = {620}, year = {2025}, }
• (2025). A similarity measure based on subspace distance for spectral clustering. NEUCOM, 620, 129187. (www)

  @article{ , abstract = {The performance of Spectral Clustering (SC) relies heavily on the choice of similarity matrix used to compute pairwise similarities between data points, especially when handling data distributed across multiple subspaces. Despite the effectiveness of subspace learning methods in identifying clusters within high-dimensional data, their integration into SC is often limited. Specifically, a majority of SC techniques rooted in subspace learning either lack efficient similarity metrics or encounter difficulties in uncovering clusters within datasets that share common subspaces. To address these concerns, this paper introduces a novel similarity metric, termed Similarity Measure based on the Distance of Subspaces (SMDS). The proposed SMDS criterion yields three key advantages. Firstly, SMDS involves identifying the local neighborhood of each sample, which typically exerts a stronger influence than global factors. Secondly, it employs subspace learning, leveraging the fact that estimating small linear subspaces is computationally more tractable than handling larger and more complex ones. Thirdly, it introduces a novel subspace clustering approach by establishing a similarity matrix based on subspace distance. This property effectively addresses the challenges posed by overlapping subspaces and facilitates their merging. Moving forward, this novel SMDS similarity matrix is then utilized within SC, leading to the proposal of SC-SMDS, a new method tailored for clustering tasks. The SC-SMDS method is evaluated through various experiments on a number of real-world benchmark datasets, demonstrating its superior performance over several competing clustering methods.}, archive = {J_NEUCOM}, author = {Nadimeh Naseri and Mahdi Eftekhari and Farid Saberi-Movahed and Mehdi Radjabalipour and Lluis A. Belanche}, doi = {10.1016/j.neucom.2024.129187}, journal = {Neurocomputing}, month = {3}, pages = {129187}, shortjournal = {Neurocomputing}, title = {A similarity measure based on subspace distance for spectral clustering}, volume = {620}, year = {2025}, }
• (2025). A semi-supervised multiscale generalized-VAE framework for one-class classification. NEUCOM, 620, 129172. (www)

  @article{ , abstract = {Deep-learning based approaches for unsupervised anomaly detection typically learn either a generative model of the inlier class or a decision boundary to encapsulate the inlier class. In addition to the training data from the inlier class, the availability of a small amount of training data from the outlier class can aid in refining the classifier model using principles of semi-supervised learning. This paper proposes a novel end-to-end deep semi-supervised variational framework for one-class classification of images, leveraging data-adaptive generalized-Gaussian (GG) models leading to effective modeling of distributions in both latent space and image space. The framework proposes a novel variational encoder that models a distribution on a multiscale (here, “scale” refers to spatial resolution) latent-space encoding, together with a generalized reparameterization scheme for the GG model’s sampling at each such scale. While the multiscale latent-space helps effective feature learning at coarse and fine spatial scales, the semi-supervision helps tune the feature learning to improve separability between the inliers and the outliers. Results on several publicly available industrial-imaging and medical-imaging datasets show the benefits of our framework’s novel components over existing approaches.}, archive = {J_NEUCOM}, author = {Renuka Sharma and Suyash P. Awate}, doi = {10.1016/j.neucom.2024.129172}, journal = {Neurocomputing}, month = {3}, pages = {129172}, shortjournal = {Neurocomputing}, title = {A semi-supervised multiscale generalized-VAE framework for one-class classification}, volume = {620}, year = {2025}, }
• (2025). A compact version of EfficientNet for skin disease diagnosis application. NEUCOM, 620, 129166. (www)

  @article{ , abstract = {EfficientNet is a convolutional neural network architecture that resulted from a neural architecture search using the AutoML MNAS framework. This process optimized both accuracy and efficiency, yielding a model that competes favorably with state-of-the-art approaches while having significantly lower parameter computation burdens on the ImageNet challenge. However, the execution of EfficientNet does not meet the speed expectations, particularly on mobile devices. This paper introduces a compact version of EfficientNet, which maintains accuracy similar on the ImageNet dataset but operates nearly twice as fast as the origin, making it more suitable for mobile devices. Furthermore, we assess the proposed model’s performance on skin cancer diagnosis using the HAM10000 dataset. To enhance the dataset, we include a new class of normal skin images, thereby enabling the trained models to be effectively applied in real-world scenarios. The primary objective of our research is to advance skin cancer diagnosis by offering a more accurate and efficient model. We believe our contributions can pave the way for improved medical applications, benefiting patients and medical practitioners alike.}, archive = {J_NEUCOM}, author = {Hoang Van Thanh and Nguyen Duy Quang and Tu Minh Phuong and Kang-Hyun Jo and Van-Dung Hoang}, doi = {10.1016/j.neucom.2024.129166}, journal = {Neurocomputing}, month = {3}, pages = {129166}, shortjournal = {Neurocomputing}, title = {A compact version of EfficientNet for skin disease diagnosis application}, volume = {620}, year = {2025}, }
• (2025). Bi-orientated rectification few-shot segmentation network based on fine-grained prototypes. NEUCOM, 620, 129160. (www)

  @article{ , abstract = {Most existing prototype-based methods typically extract a single target class prototype from the support image, leading to false activation or information loss. In this paper, we propose a bi-orientated rectification few-shot segmentation network based on fine-grained prototypes. Our approach consists of a fine-grained prototype generation module and a bi-orientated rectification module, addressing the limitations of current methods that only extract general target prototype. By integrating support features with the support mask and the general target prototype, we extract fine-grained prototypes that capture comprehensive information about both target and non-target regions of the support image. Furthermore, the network develops a bi-orientated rectification mechanism that significantly enhances the quality of query feature activation, consisting of an activation for the target region and an inhibition for the non-target region. By fusing fine-grained foreground prototypes with general foreground prototype, the network activates and compensates the target region of the query image. Meanwhile, the network inhibits and eliminates the non-target region of the query image by fusing fine-grained background prototypes with general background prototype. Extensive experiments on benchmark datasets (PASCAL-5 i and COCO-20 i ) validate the effectiveness of our approach. Particularly using ResNet50 as the backbone network, our network outperforms the baseline by 2.15% in terms of mIoU on the COCO-20 i dataset.}, archive = {J_NEUCOM}, author = {Aiping Yang and Zijia Sang and Yaran Zhou and Jiale Cao and Liping Liu}, doi = {10.1016/j.neucom.2024.129160}, journal = {Neurocomputing}, month = {3}, pages = {129160}, shortjournal = {Neurocomputing}, title = {Bi-orientated rectification few-shot segmentation network based on fine-grained prototypes}, volume = {620}, year = {2025}, }
• (2025). Data-driven adaptive formation control based on preview mechanism for networked multi-robot systems with communication delays. NEUCOM, 620, 129151. (www)

  @article{ , abstract = {This study addresses the formation control task of networked multi-robot systems operating in an edge-end network architecture with communication delays. We propose a data-driven distributed adaptive formation controller DDAPM for stable and precise formation performance. Implemented at each edge-node, the DDAPM controller includes a single closed-loop angular velocity control module and a double closed-loop linear velocity control module connected in parallel. It integrates a preview mechanism to capture the intentions of robots and delay compensators to mitigate the adverse effects of communication delays. We analyze the stability and consensus of the NMRS utilizing DDAPM controllers. Numerical simulations are conducted to validate the feasibility, effectiveness, and adaptability of the DDAPM. Comparative simulations between the DDAPM and other controllers further substantiate its advantages. Additionally, an ablation study is performed to highlight the effectiveness and necessity of delay compensators and adaptive parameter pseudo-partial derivatives. Finally, we discuss the study’s limitations and propose potential directions for future research.}, archive = {J_NEUCOM}, author = {Chenzhuolei Chao and Haoran Tan and Xueming Zhang and Gang Wang and You Wu and Yaonan Wang}, doi = {10.1016/j.neucom.2024.129151}, journal = {Neurocomputing}, month = {3}, pages = {129151}, shortjournal = {Neurocomputing}, title = {Data-driven adaptive formation control based on preview mechanism for networked multi-robot systems with communication delays}, volume = {620}, year = {2025}, }
• (2025). Investigating 3D object detection using stereo camera and LiDAR fusion with bird's-eye view representation. NEUCOM, 620, 129144. (www)

  @article{ , abstract = {Multi-sensor fusion for 3D object detection is a crucial development in autonomous vehicle technology. Current research primarily explores the combination of monocular cameras with LiDARs. However, there is a notable gap in the integration of stereo cameras with LiDARs. A significant drawback of the fusion between monocular cameras and LiDARs is its vulnerability to sensor failures, which can completely disable the system. This paper presents a novel end-to-end framework called SLBEVFusion, which leverages the effective combination of stereo cameras and LiDARs for 3D object detection. The framework utilizes the rich semantic and depth data from stereo images along with spatial data from LiDAR point clouds to improve detection accuracy. Moreover, it addresses sensor failures by isolating the data streams of each sensor and implementing straightforward feature fusion in Bird's Eye View (BEV) space. Comprehensive testing on the challenging KITTI dataset demonstrates that our approach significantly enhances performance in 3D object detection by merging stereo cameras with LiDARs.}, archive = {J_NEUCOM}, author = {Xin Nie and Lin Zhu and Zhicheng He and Aiguo Cheng and Shengshi Zhong and Eric Li}, doi = {10.1016/j.neucom.2024.129144}, journal = {Neurocomputing}, month = {3}, pages = {129144}, shortjournal = {Neurocomputing}, title = {Investigating 3D object detection using stereo camera and LiDAR fusion with bird's-eye view representation}, volume = {620}, year = {2025}, }
• (2025). ReZG: Retrieval-augmented zero-shot counter narrative generation for hate speech. NEUCOM, 620, 129140. (www)

  @article{ , abstract = {The proliferation of hate speech (HS) on social media poses a serious threat to societal security. Automatic counter narrative (CN) generation, as an active strategy for HS intervention, has garnered increasing attention in recent years. Existing methods for automatically generating CNs mainly rely on re-training or fine-tuning pre-trained language models (PLMs) on human-curated CN corpora. Unfortunately, the annotation speed of CN corpora cannot keep up with the growth of HS targets, while generating specific and effective CNs for unseen targets remains a significant challenge for the model. To tackle this issue, we propose Re trieval-Augmented Z ero-shot G eneration (ReZG) to generate CNs with high-specificity for unseen targets. Specifically, we propose a multi-dimensional hierarchical retrieval method that integrates stance, semantics, and fitness, extending the retrieval metric from single dimension to multiple dimensions suitable for the knowledge that refutes HS. Then, we implement an energy-based constrained decoding mechanism that enables PLMs to use differentiable knowledge preservation, countering, and fluency constraint functions instead of in-target CNs as control signals for generation, thereby achieving zero-shot CN generation. With the above techniques, ReZG can integrate external knowledge flexibly and improve the specificity of CNs. Experimental results show that ReZG exhibits stronger generalization capabilities and outperforms strong baselines with significant improvements of 2.0%+ in the relevance and 4.5%+ in the countering success rate metrics.}, archive = {J_NEUCOM}, author = {Shuyu Jiang and Wenyi Tang and Xingshu Chen and Rui Tang and Haizhou Wang and Wenxian Wang}, doi = {10.1016/j.neucom.2024.129140}, journal = {Neurocomputing}, month = {3}, pages = {129140}, shortjournal = {Neurocomputing}, title = {ReZG: Retrieval-augmented zero-shot counter narrative generation for hate speech}, volume = {620}, year = {2025}, }
• (2025). Numerical solutions for space–time conformable nonlinear partial differential equations via a scientific machine learning technique. NEUCOM, 620, 129134. (www)

  @article{ , abstract = {Several techniques have been proposed to find analytical and numerical solutions to (partial differential equations) PDEs and fractional PDEs. However, exact analytical solutions to fractional PDEs are rarely available, numerical methods are so far applicable. Therefore, accurate methods for finding numerical solutions to fractional PDEs are still under investigation and are currently an active area of research. One of the accurate techniques for finding approximate numerical solutions for PDEs is a physics-informed neural network (PINN) through scientific machine learning. In this paper, we introduce an effective homotopy-preserving physics-informed neural network (HP-PINN) that adopts smooth trainable embedding functional parameters through a neural network using the homotopy technique to solve forward and inverse problems governed by conformable space–time fractional PDEs. A more general homotopy perturbation technique is introduced to obtain the analytical solutions for the debated problems. A numerical comparison shows the effectiveness and accuracy of the proposed HP-PINN.}, archive = {J_NEUCOM}, author = {Hayman Thabet and Subhash Kendre and James Peters}, doi = {10.1016/j.neucom.2024.129134}, journal = {Neurocomputing}, month = {3}, pages = {129134}, shortjournal = {Neurocomputing}, title = {Numerical solutions for space–time conformable nonlinear partial differential equations via a scientific machine learning technique}, volume = {620}, year = {2025}, }
• (2025). Unsupervised spiking neural network based on liquid state machine and self-organizing map. NEUCOM, 620, 129120. (www)

  @article{ , abstract = {The liquid state machine (LSM) is a reservoir computing paradigm and also a type of recurrent spiking neural network, combining the core strengths of both spiking neural networks (SNNs) and recurrent neural networks (RNNs). The architecture and functionality of the LSM are closely related to biological neural systems in many respects. Current research has extensively focused on the optimization of the structure and parameters within the liquid layer of LSM. However, there is a relative scarcity of studies on the readout layer, and its potential and effectiveness warrant further investigation. In this paper, a fully unsupervised spiking neural network approach is proposed, integrating the LSM with a spike self-organizing map (SOM). This integration results in a fully unsupervised SNN capable of performing visual clustering analysis on the readout layer of LSM and enhancing its biological interpretability. To rigorously assess the performance of our proposed method, we selected two well-established datasets for experimental validation: the MNIST dataset for image processing and the FSDD dataset for sound recognition. The experimental results indicate that our unsupervised learning approach achieved classification accuracies of 90.0 % and 88.0 % on these respective datasets. The model proposed in this paper utilizes randomly generated weights in the liquid layer, thereby bypassing the complex process of parameter optimization and simplifying the model's architecture. Additionally, the adoption of unsupervised learning eliminates the need for labeled data, significantly improving computational efficiency. These attributes provide our model with substantial practical utility and significance, particularly in environments where computational resources are limited and labeled data is not readily available.}, archive = {J_NEUCOM}, author = {Youdong Zhang and Lingfei Mo and Xu He and Xiaolin Meng}, doi = {10.1016/j.neucom.2024.129120}, journal = {Neurocomputing}, month = {3}, pages = {129120}, shortjournal = {Neurocomputing}, title = {Unsupervised spiking neural network based on liquid state machine and self-organizing map}, volume = {620}, year = {2025}, }
• (2025). Few-shot cyberviolence intent classification with meta-learning AutoEncoder based on adversarial domain adaptation. NEUCOM, 620, 129089. (www)

  @article{ , abstract = {The phenomenon of cyberviolence has become a critical issue in online security, drawing attention from various stakeholders. A major shortcoming in the previous works is the limitation of using simple methods like ”yes” or ”no” to evaluate cyberviolence utterances, which can significantly restrict netizens’ free speech. Therefore, we provide a novel strategy for detecting cyberviolence utterances based on the user’s real intent. Fine-grained cyberviolence intents are complex, leading to texts that share similar syntactic structures and semantics but differ in intent category. The previous method did not consider this issue. To address this, in this paper, we propose a Meta -learning A uto E ncoder (MetaAE) based on adversarial domain adaptation. The goal is to comprehend and learn the inherent logical rules and important semantic knowledge of cyberviolence utterances, specifically targeting fine-grained cyberviolence intent problems. Specifically, we use the autoencoder structure to help the model implement self-supervised learning. This enables the model to comprehend the inherent logical structure of texts with different intent categories and helps the model learn important semantic knowledge of the text during the encoder compression process. At the same time, to solve the problem of overfitting in small samples and multidomain cyberviolence utterance, we introduce domain adversarial learning to align domain features and enhance model robustness. Experimental results on both a real cyberviolence intent classification dataset and a public dataset demonstrate significant improvements. On 5-way 1-shot and 5-shot Chinese and English cyberviolence datasets, MetaAE improved the accuracy by approximately 7.23%, 8.27%, 7.22%, and 5%, respectively. In the public dataset, MetaAE improved accuracy by approximately 2.53% on 5-way 5-shot. 1}, archive = {J_NEUCOM}, author = {Shun Yang and YaJun Du and ShangYi Du and XianYong Li and XiaoLiang Chen and YanLi Li and ChunZhi Xie and Jia Liu}, doi = {10.1016/j.neucom.2024.129089}, journal = {Neurocomputing}, month = {3}, pages = {129089}, shortjournal = {Neurocomputing}, title = {Few-shot cyberviolence intent classification with meta-learning AutoEncoder based on adversarial domain adaptation}, volume = {620}, year = {2025}, }
• (2025). MPE3: Learning meta-prompt with entity-enhanced semantics for few-shot named entity recognition. NEUCOM, 620, 129031. (www)

  @article{ , abstract = {Recently, prompt-tuning has been proven to be surprisingly effective on few-shot tasks. Intuitively, some studies explore Few-shot Named Entity Recognition (NER) based on prompt-tuning. However, how to properly initialize and effectively learn the prompt under limited training conditions still remains significantly challenging for few-shot NER. To meet these challenges, we propose a novel M eta- P rompt with E ntity- E nhanced semantics for Few-shot N E R, MPE 3 for brevity. Specifically, we first explore the importance of the named entities’ semantics in the few-shot NER task. And we propose to construct prompts with entity-enhanced semantics which contain much useful prior knowledge for identifying named entities. Furthermore, to address the issue of inadequate training more substantially, we aim to train a meta-prompt that can be more effective and adaptive for few-shot NER scenarios. To achieve this, we divide the training data into many source-domain agnostic meta-tasks tailored to the characteristics of the NER problem for training. And we specially design a prompt meta-learner for training these meta-tasks. This training strategy succeeds in guiding prompts to optimize in a better direction for few-shot scenarios by the learned meta-knowledge from each meta-task. We conduct extensive experiments on three NER datasets under two different few-shot settings. Our method outperforms the current state-of-the-art model by 5.60% ∼ 13.34% and 2.41% ∼ 7.34% on average in the two different few-shot settings respectively, which validates the effectiveness and superiority of our model.}, archive = {J_NEUCOM}, author = {Yuwei Xia and Zhao Tong and Liang Wang and Qiang Liu and Shu Wu and Xiaoyu Zhang}, doi = {10.1016/j.neucom.2024.129031}, journal = {Neurocomputing}, month = {3}, pages = {129031}, shortjournal = {Neurocomputing}, title = {MPE3: Learning meta-prompt with entity-enhanced semantics for few-shot named entity recognition}, volume = {620}, year = {2025}, }
• (2025). Meta channel masking for cross-domain few-shot image classification. NEUCOM, 620, 128956. (www)

  @article{ , abstract = {Cross-domain Few-shot Learning (CD-FSL) aims to address the challenges of FSL where significant domain gaps exist between source and target image datasets. Unlike many existing CD-FSL methods that utilize an auxiliary target dataset with a few labeled target images to enhance model generalization, our approach directly tackles the limitations imposed by the reliance on source-specific knowledge. We observe that models trained on unbalanced datasets tend to overfit to source-specific features, which, while effective in the source domain, generalize poorly to the target image domain. To address this, we introduce a novel dropout-based framework named Meta Channel Masking (MCM). This framework attenuates the learning of model channels on the source domain by dynamically masking source feature channels during training. In contrast to traditional dropout techniques that manually set masking probabilities based on statistical assumptions about the source data, our MCM framework employs a meta-learning process that automatically adjusts channel mask probabilities. This adjustment is informed by auxiliary target data, effectively minimizing few-shot loss on the auxiliary target dataset and thereby enhancing the model’s generalization capabilities in the target domain. Our extensive experiments across various image classification benchmark datasets demonstrate that our framework outperforms state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Siqi Hui and Sanping Zhou and Ye Deng and Pengna Li and Jinjun Wang}, doi = {10.1016/j.neucom.2024.128956}, journal = {Neurocomputing}, month = {3}, pages = {128956}, shortjournal = {Neurocomputing}, title = {Meta channel masking for cross-domain few-shot image classification}, volume = {620}, year = {2025}, }
• (2025). Adaptive multi-view subspace clustering algorithm based on representative features and redundant instances. NEUCOM, 620, 128839. (www)

  @article{ , abstract = {The goal of multi-view subspace clustering is to group objects into distinct clusters using information from multiple views, which remains challenging when dealing with diverse data sources. However, the learned consensus matrix of most of the existing approaches may be inaccurate and lead to unsatisfactory clustering performance due to the following two limitations. The first one is that the representative features are treated equally with the other unimportant features, thus resulting in the incapability of capturing the latent information in each view. Second, the redundant information is not considered from the viewpoints, which raises a negative impact caused by certain noise. This paper proposes a novel adaptive multi-view subspace clustering based on representative features and redundant instances (AMC2R) to address these two challenging issues. Specifically, the proposed algorithm focuses on the representative feature, in which a weighted matrix is designed for each view to dynamically obtain a consensus matrix using the information conveyed by each feature. After that, all the views are concatenated together to utilize the complementary information across the views, and a matrix under l 2 , 1 -norm constraint is used to obtain the final consensus matrix to eliminate the negative impact raised by the redundant information. A unified framework is then designed to integrate the above steps. In this way, the representative information and the redundant information can simultaneously interact with one another during the iterations. Experimental results on different datasets demonstrate the effectiveness of the proposed algorithm and have achieved excellent results on Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI).}, archive = {J_NEUCOM}, author = {Zhuoyue Ou and Xiuqin Deng and Lei Chen and Jiadi Deng}, doi = {10.1016/j.neucom.2024.128839}, journal = {Neurocomputing}, month = {3}, pages = {128839}, shortjournal = {Neurocomputing}, title = {Adaptive multi-view subspace clustering algorithm based on representative features and redundant instances}, volume = {620}, year = {2025}, }
• (2025). A versatile feature selection learning method for satellite attitude control system fault diagnosis with limited data. NEUCOM, 619, 129216. (www)

  @article{ , abstract = {Fault diagnosis in satellite attitude control systems is crucial for ensuring the operation and maintenance of satellites. However, current methods face challenges in achieving the desired classification results due to limited fault samples and complex temporal characteristics. To address these problems, this paper proposes a versatile fault diagnosis method called feature selection learning, which divides attitude control system faults into component-level and system-level to create a fault hierarchy. The method considers an overfitted classifier as the base-classifier (BC). It then sifts through a subset of downscaled fusion features derived from the BC to serve as inputs for the feature selection embedder (FSE), aiming to identify the FSE that captures the core fault features. The efficacy of the proposed method in each group of fault diagnosis is confirmed using the satellite attitude control system dataset. The experimental results show that the proposed method achieves a fault diagnosis accuracy of 91.08% under extremely scarce sample conditions, which is an increase of 4.69% compared to the optimal accuracy among the five comparison methods. Furthermore, the proposed method exhibits stable fault diagnosis performance across different datasets, sample sizes, and BCs, demonstrating its applicability in the context of fault hierarchy.}, archive = {J_NEUCOM}, author = {Huaitao Xia and Tao Meng}, doi = {10.1016/j.neucom.2024.129216}, journal = {Neurocomputing}, month = {2}, pages = {129216}, shortjournal = {Neurocomputing}, title = {A versatile feature selection learning method for satellite attitude control system fault diagnosis with limited data}, volume = {619}, year = {2025}, }
• (2025). Vertical NAND flash memory-based matrix-matrix multiplication scheme for energy-efficient attention score computation. NEUCOM, 619, 129210. (www)

  @article{ , abstract = {We introduce a vertical NAND (V-NAND) flash memory-based matrix-matrix multiplication (MMM) scheme optimized for attention score computing in the self-attention module. Recent advances in transformer-based artificial intelligence models have significantly outperformed in various tasks. However, the MMM essential to the computation of self-attention is not well-suited for acceleration using nonvolatile memory-based compute-in-memory systems. Our scheme utilizes the three-terminal characteristics of V-NAND flash memory cells, considering the conductance resolution, current sum loss in the bit-line, and ADC quantization. Through this approach, a vision transformer demonstrates an accuracy comparable to the software baseline, achieving 87.03 % on the CIFAR-100 test dataset image classification with an 8-bit ADC. Furthermore, proposed V-NAND flash memory-based MMM scheme consumes only 12.2 % of the energy compared to the tensor processing unit (TPU), and when using sparse attention, it consumes as little as 0.37 % of the energy compared to the TPU.}, archive = {J_NEUCOM}, author = {Jonghyun Ko and Sung-Ho Park and Jiseong Im and Jeonghyun Kim and Ryun-Han Koo and Yeongheon Yang and Jong-Ho Lee}, doi = {10.1016/j.neucom.2024.129210}, journal = {Neurocomputing}, month = {2}, pages = {129210}, shortjournal = {Neurocomputing}, title = {Vertical NAND flash memory-based matrix-matrix multiplication scheme for energy-efficient attention score computation}, volume = {619}, year = {2025}, }
• (2025). RDMIF: Reverse dictionary model based on multi-modal information fusion. NEUCOM, 619, 129202. (www)

  @article{ , abstract = {In reverse dictionary (RD) research, the goal is to construct the components of a target word vector through the parsing of human-comprehensible definitions. This process emulates the human lexical memory mechanism, where the introduction of visual information is considered an efficient memory strategy. However, due to the scarcity of multimodal datasets, existing studies have largely failed to capitalize on visual information. To bridge this research gap, we have constructed a multimodal dataset and a reverse dictionary model based on multimodal (RDMIF) tailored for common objects. Dataset comprises words, images, definitions, and their examples. RDMIF employs an attention mechanism to selectively focus on key elements within textual and visual information and reveals underlying semantic connections through global intra-modal and inter-modal correlation analysis. The model computes cosine similarity between definitions and target words to evaluate explanatory power. The experimental results indicate that RDMIF improves rank value by 1.2 %, Top1 accuracy by 3.6 %, and Top5 accuracy by 5.2 % over baseline models. Our research offers new perspectives and methodologies for the multimodal RD domain.}, archive = {J_NEUCOM}, author = {Sicheng Tian and Shaobin Huang and Rongsheng Li and Chi Wei}, doi = {10.1016/j.neucom.2024.129202}, journal = {Neurocomputing}, month = {2}, pages = {129202}, shortjournal = {Neurocomputing}, title = {RDMIF: Reverse dictionary model based on multi-modal information fusion}, volume = {619}, year = {2025}, }
• (2025). A novel method for consumer preference extraction based on perceived usefulness and de-neutral sentiment. NEUCOM, 619, 129197. (www)

  @article{ , abstract = {The accurate assessment of consumer preferences has become increasingly crucial for effective business decision-making. However, merchants face significant challenges in accurately analyzing these preferences due to the overwhelming volume of online consumer feedback, the growing complexity of customer preferences, and the rapid evolution of market trends. Current methodologies are often hampered by various errors and biases, such as the misleading influence of irrelevant information within overloaded data, challenges in extracting latent features from big data, and issues related to neutral sentiment factors. Additionally, the usefulness of information is frequently dependent on voting mechanisms, which are susceptible to biases like the early bird effect. To address these challenges, we propose a novel preference extraction method based on perceived useful information. This approach integrates sentiment analysis rooted in deep learning with neutral sentiment processing and combines multi-criteria decision-making to extract hidden features while mitigating the misleading impact of irrelevant word frequencies on preference calculations. Furthermore, our method employs the principle of information entropy to extract information utility, thereby avoiding common biases associated with traditional voting methods. Experimental results demonstrate the superiority of our method across two case studies: for search-based products, the method achieved an F1 score of 89.6 % and an AUC of 79.4 %, while for experience-based products, it recorded 84.6 % and 80.6 %, respectively. The primary contribution of this research lies in providing a systematic approach to uncover product features and accurately analyze consumer preferences, offering valuable insights for business decision-making. This has significant theoretical and practical implications for product development, marketing, and customer service.}, archive = {J_NEUCOM}, author = {Huiran Liu and Zheng Wang and Zhiming Fang}, doi = {10.1016/j.neucom.2024.129197}, journal = {Neurocomputing}, month = {2}, pages = {129197}, shortjournal = {Neurocomputing}, title = {A novel method for consumer preference extraction based on perceived usefulness and de-neutral sentiment}, volume = {619}, year = {2025}, }
• (2025). An embedded feature pyramid network enables bidirectional information flow for object detection and instance segmentation. NEUCOM, 619, 129196. (www)

  @article{ , abstract = {Object detection and instance segmentation are fundamental tasks in computer vision. In the past decade, deep neural networks have become the dominating approach for accomplishing these purposes, among which feature pyramid networks (FPNs) are popular models for simultaneous object detection and instance segmentation because their unique capability in multi-scale feature representation. However, the existing FPNs are often suffered from low utilization of multi-scale information and inadequacy in receptive field. To address these issues, here we proposed an efficient embedded feature pyramid network (EFPNet), which enables bidirectional information flow to create and utilize feature pyramids with richer multi-scale information. To establish the bidirectional parallel information flow, small three-level feature pyramids were firstly inserted into the main feature pyramid of the backbone model, leading to an enlarged receptive field and the increased utilization of multi-scale information, especially the deep semantic information. Moreover, we further designed a pathway enhancement module to form another cascading bidirectional information flow, enhancing the utilization of the shallow detailed information. Finally, we presented a global selection module, based on global pooling, to further expand the receptive field and select important features from the rich multi-scale information extracted in previous steps. Experiments on the MS COCO, PASCAL VOC and HRSID datasets validated the superiority of our proposed EFPNet approach over state-of-the-art baseline methods by achieving an improvement of 1.2 %-4.8 % in average precision (AP) for both object detection and instance segmentation tasks.}, archive = {J_NEUCOM}, author = {Chunning Meng and Zequn Sun and Lianzhi Huo and Tao Li and Shengjiang Chang and Zhiqing Zhang}, doi = {10.1016/j.neucom.2024.129196}, journal = {Neurocomputing}, month = {2}, pages = {129196}, shortjournal = {Neurocomputing}, title = {An embedded feature pyramid network enables bidirectional information flow for object detection and instance segmentation}, volume = {619}, year = {2025}, }
• (2025). FlowST-net: Tackling non-uniform spatial and temporal distributions for scene flow estimation in point clouds. NEUCOM, 619, 129183. (www)

  @article{ , abstract = {Accurate scene flow estimation from 3D point clouds is critical for applications such as robotics and autonomous driving. In this paper, we present FlowST-Net, a novel approach designed to address the challenges posed by non-uniform spatial and temporal distributions in point cloud sequences. Firstly, we propose a single-frame voxelization interpolation module to capture locally consistent and discriminative features within each frame, mitigating non-uniform spatial distribution. Secondly, a cross-frame feature-aware attention module is introduced to handle non-uniform temporal distributions by leveraging a cross-attention mechanism to model inter-frame point dependencies. Thirdly, we introduce a cross-frame geometry-aware augmentation module to enhance the consistency of local correspondences across frames. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the FlyingThings3D dataset and exhibits strong generalization to the KITTI dataset, highlighting its effectiveness in scene flow estimation.}, archive = {J_NEUCOM}, author = {Xiaohu Yan and Mian Zhang and Xuefeng Tan and Yiqi Wu and Dejun Zhang}, doi = {10.1016/j.neucom.2024.129183}, journal = {Neurocomputing}, month = {2}, pages = {129183}, shortjournal = {Neurocomputing}, title = {FlowST-net: Tackling non-uniform spatial and temporal distributions for scene flow estimation in point clouds}, volume = {619}, year = {2025}, }
• (2025). Optimized deep learning model for comprehensive medical image analysis across multiple modalities. NEUCOM, 619, 129182. (www)

  @article{ , abstract = {This study presents a novel amalgamated model for the diagnosis of multiple medical conditions using various imaging modalities, including Chest X-ray, MRI, and endoscopic images. Each imaging modality has unique protocols and feature characteristics, presenting significant complexity and challenges in diagnosis. To address these issues, we propose a truncated lightweight model that effectively fuses diverse image features while reducing computational requirements. The model utilizes deep learning (DL) techniques and multi-scale feature learning to enhance diagnostic capabilities across different medical images. Specifically, it employs an efficient MobileNet architecture to simultaneously diagnose multiple diseases. Key innovations include model truncation, a modified naive inception block for multi-scale feature extraction, and metaheuristic optimization methods. By optimizing the architecture with techniques such as Chain Foraging and Cyclone Aging, the model achieves robustness and scalability, improving generalization across varying image resolutions. Additionally, an integrated ConvLSTM unit before the Softmax layer enhances feature extraction across spatial and temporal dimensions, addressing challenges related to differing feature sizes and scales in multi-disease diagnosis. We conducted comprehensive testing on publicly available multi-class medical image datasets, including brain MRI, chest X-ray, and gastro endoscopic images, which demonstrates that the proposed model outperforms existing methods, achieving an overall accuracy of 97.37 %. To further support clinical decision-making, we utilized visualization techniques such as GradCAM, and Feature Map analysis to enhance the interpretability of the model predictions. Overall, the proposed model not only showcases exceptional performance in classifying various medical images but also presents an optimal balance between accuracy and computational efficiency, establishing its potential as a practical solution for accurate multi-modal medical image analysis.}, archive = {J_NEUCOM}, author = {Saif Ur Rehman Khan and Sohaib Asif and Ming Zhao and Wei Zou and Yangfan Li and Xiangmin Li}, doi = {10.1016/j.neucom.2024.129182}, journal = {Neurocomputing}, month = {2}, pages = {129182}, shortjournal = {Neurocomputing}, title = {Optimized deep learning model for comprehensive medical image analysis across multiple modalities}, volume = {619}, year = {2025}, }
• (2025). FinalGNN: A dual feature graph enhanced model for CTR prediction. NEUCOM, 619, 129181. (www)

  @article{ , abstract = {Click-through rate (CTR) prediction is crucial in computational advertising and recommendation systems, with superior performance relying on effectively modeling feature interactions across various fields. Recent studies construct graphs with features as nodes and feature interactions as edges, and stack graph neural networks to capture interactions of arbitrary orders. However, these methods have limitations, such as introducing high-order noisy interactions and failing to comprehensively capture beneficial high-order feature interactions. We introduce FinalGNN, a novel approach, building upon a meticulous analysis of generating noisy interactions and omitting beneficial interactions. FinalGNN eschews direct reliance on beneficial low-order interactions for learning beneficial high-order interactions. Instead, it constructs two types of graphs at each layer: one identifies currently beneficial feature interactions and enhances prediction accuracy at that layer; the other captures potential feature interactions that are deemed essential for fully capturing beneficial high-order interactions. Additionally, FinalGNN incorporates feature-edges and field-edges into its interaction aggregation process, and utilizes a graph-level prediction layer and loss function to fully exploit multi-layer feature interactions. Experiments conducted on six publicly available datasets demonstrate the effectiveness of our method in comparison to state-of-the-art approaches.}, archive = {J_NEUCOM}, author = {Yue Wang and Bilian Chen}, doi = {10.1016/j.neucom.2024.129181}, journal = {Neurocomputing}, month = {2}, pages = {129181}, shortjournal = {Neurocomputing}, title = {FinalGNN: A dual feature graph enhanced model for CTR prediction}, volume = {619}, year = {2025}, }
• (2025). A hybrid graph transformer network for fingerprint and palmprint minutiae matching of young children. NEUCOM, 619, 129180. (www)

  @article{ , abstract = {Deep neural networks have successfully applied for fingerprints and palmprints recognition. However, the performances of most fingerprint and palmprint recognition methods are often degraded in young children, due to the small physical size and poor image quality. Thus, it is still challenging for reliable fingerprint and palmprint recognition of young children. In this study, we propose a hybrid graph transformer network to gradually learn both local and global features of minutiae points, which are integrated to enhance the matching accuracy of fingerprints and palmprints in young children. First, a multi-scale transformer is designed to extract the features of minutiae points detected in fingerprints and palmprints. Then, a graph attention network (GAN) is proposed to capture the topological structure of minutiae set for more powerful representation and precise minutiae matching. Furthermore, to address the class imbalance problem between the positive and negative matched minutiae pairs, a filtering-and-refining strategy is developed for coarse-to-fine minutiae matching of fingerprints and palmprints. The extensive experiments and comparisons are performed on two fingerprint datasets and one palmprint dataset of young children, showing the promising performance of the proposed method.}, archive = {J_NEUCOM}, author = {Aitong Liu and Yelin Shi and Ajay Kumar and Xin Chen and Manhua Liu}, doi = {10.1016/j.neucom.2024.129180}, journal = {Neurocomputing}, month = {2}, pages = {129180}, shortjournal = {Neurocomputing}, title = {A hybrid graph transformer network for fingerprint and palmprint minutiae matching of young children}, volume = {619}, year = {2025}, }
• (2025). Nash equilibrium seeking in non-cooperative heterogeneous multi-robot systems via output regulation. NEUCOM, 619, 129179. (www)

  @article{ , abstract = {This paper presents a study on Nash equilibrium seeking in noncooperative games within multi-robot systems, a topic of increasing importance in diverse sectors including civil, security, and military. Unlike conventional approaches where players can directly observe the actions of others, our method assumes limited visibility, where players can only communicate through an undirected and connected communication graph. We introduce a novel distributed control approach, integrating gradient play with a consensus protocol. This method facilitates effective Nash equilibrium seeking by leveraging information shared among neighboring robots in heterogeneous linear dynamic systems. The proposed solution employs a high-level distributed Nash equilibrium-seeking algorithm, serving as an optimal reference generator for each robot to track the Nash equilibrium, and an advanced output regulation technique, aiming to regulate the output (e.g., position) of the robots with respect to the obtained references. Theoretical analysis confirms the convergence of our algorithm through Lyapunov stability analysis. The effectiveness and practical applicability of our approach are validated through numerical simulations and empirical testing with physical robots, highlighting its efficacy and utility in real-world scenarios.}, archive = {J_NEUCOM}, author = {Yi Dong and Zhongguo Li and Shoaib Ehsan and Zhengtao Ding and Xiaowei Huang}, doi = {10.1016/j.neucom.2024.129179}, journal = {Neurocomputing}, month = {2}, pages = {129179}, shortjournal = {Neurocomputing}, title = {Nash equilibrium seeking in non-cooperative heterogeneous multi-robot systems via output regulation}, volume = {619}, year = {2025}, }
• (2025). Is mamba effective for time series forecasting?. NEUCOM, 619, 129178. (www)

  @article{ , abstract = {In the realm of time series forecasting (TSF), it is imperative for models to adeptly discern and distill hidden patterns within historical time series data to forecast future states. Transformer-based models exhibit formidable efficacy in TSF, primarily attributed to their advantage in apprehending these patterns. However, the quadratic complexity of the Transformer leads to low computational efficiency and high costs, which somewhat hinders the deployment of the TSF model in real-world scenarios. Recently, Mamba, a selective state space model, has gained traction due to its ability to process dependencies in sequences while maintaining near-linear complexity. For TSF tasks, these characteristics enable Mamba to comprehend hidden patterns as the Transformer and reduce computational overhead compared to the Transformer. Therefore, we propose a Mamba-based model named Simple-Mamba (S-Mamba) for TSF. Specifically, we tokenize the time points of each variate autonomously via a linear layer. A bidirectional Mamba layer is utilized to extract inter-variate correlations and a Feed-Forward Network is set to learn temporal dependencies. Finally, the generation of forecast outcomes through a linear mapping layer. Experiments on thirteen public datasets prove that S-Mamba maintains low computational overhead and achieves leading performance. Furthermore, we conduct extensive experiments to explore Mamba’s potential in TSF tasks. Our code is available at https://github.com/wzhwzhwzh0921/S-D-Mamba .}, archive = {J_NEUCOM}, author = {Zihan Wang and Fanheng Kong and Shi Feng and Ming Wang and Xiaocui Yang and Han Zhao and Daling Wang and Yifei Zhang}, doi = {10.1016/j.neucom.2024.129178}, journal = {Neurocomputing}, month = {2}, pages = {129178}, shortjournal = {Neurocomputing}, title = {Is mamba effective for time series forecasting?}, volume = {619}, year = {2025}, }
• (2025). A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark. NEUCOM, 619, 129177. (www)

  @article{ , abstract = {Despite the recent emergence of video captioning models, how to generate the text description with specific entity names and fine-grained actions is far from being solved, which however has great applications such as basketball live text broadcast. In this paper, a new basketball benchmark for entity-aware video captioning is proposed. Specifically, we construct a multimodal basketball game knowledge graph (KG_NBA_2022) storing basketball game records as well as detailed information on teams and players. Then, a multimodal basketball game video captioning (VC_NBA_2022) dataset that contains 9 types of fine-grained shooting events and 286 players’ knowledge (i.e., images and names) is constructed based on KG_NBA_2022 in an automatic approach. We also develop a simple yet effective knowledge guided entity-aware video captioning network (KEANet) based on a candidate player list in an encoder–decoder form for basketball live text broadcast. The temporal contextual information in video is encoded by introducing the Bi-directional Gated Recurrent Unit (Bi-GRU) module. And the entity-aware module is designed to model relationships among players and emphasize key players. Extensive experiments on multiple sports benchmarks demonstrate that KEANet effectively leverages additional knowledge and outperforms advanced video captioning models.}, archive = {J_NEUCOM}, author = {Zeyu Xi and Ge Shi and Xuefen Li and Junchi Yan and Zun Li and Lifang Wu and Zilin Liu and Liang Wang}, doi = {10.1016/j.neucom.2024.129177}, journal = {Neurocomputing}, month = {2}, pages = {129177}, shortjournal = {Neurocomputing}, title = {A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark}, volume = {619}, year = {2025}, }
• (2025). Feature-weighted fuzzy clustering methods: An experimental review. NEUCOM, 619, 129176. (www)

  @article{ , abstract = {Soft clustering, a widely utilized method in data analysis, offers a versatile and flexible strategy for grouping data points. Most soft clustering algorithms assume that all the features present in the feature space of a dataset are of equal importance and neglect their degree of informativeness or irrelevance. Distinguishing between the relative importance of features in providing an optimal clustering structure has become a very challenging task. Many feature weighting methods have been proposed to deal with this problem in the field of soft clustering, which can broadly categorized into six major types: feature reduction-based, entropy-based, variance-based, membership-based, optimization-based, and meta-heuristic-based. This paper comprehensively reviews the most significant fuzzy clustering algorithms that employ feature weighting techniques. A taxonomy of the feature weighting-based fuzzy clustering algorithms is presented. Furthermore, all state-of-the-art approaches are implemented in Python and compared in terms of clustering performance by conducting various experimental evaluation schemes. In this comprehensive experimental analysis, 26 state-of-the-art clustering algorithms are evaluated on two synthetic and 18 benchmark UCI datasets based on Accuracy (ACC) , Normalized Mutual Information (NMI), Precision (PR), Recall (RE), F1, Silhouette (SI) and Davies-Bouldin (DB) evaluation criteria. Moreover, the significance of the experimental comparisons is examined using Friedman and Holm’s post-hoc statistical tests. The experimental analysis demonstrates the superior performance of variance-based feature weighting algorithms in most datasets. All the tested algorithms are implemented in Python, and the related source codes are shared publicly at https://github.com/Amin-Golzari-Oskouei/FWSCA .}, archive = {J_NEUCOM}, author = {Amin Golzari Oskouei and Negin Samadi and Shirin Khezri and Arezou Najafi Moghaddam and Hamidreza Babaei and Kiavash Hamini and Saghar Fath Nojavan and Asgarali Bouyer and Bahman Arasteh}, doi = {10.1016/j.neucom.2024.129176}, journal = {Neurocomputing}, month = {2}, pages = {129176}, shortjournal = {Neurocomputing}, title = {Feature-weighted fuzzy clustering methods: An experimental review}, volume = {619}, year = {2025}, }
• (2025). Predefined time sliding mode control based on interval type-2 fuzzy emotional neural network. NEUCOM, 619, 129173. (www)

  @article{ , abstract = {This article investigates the predefined time convergence problem based on interval 2 type fuzzy emotional neural networks (IT2FENN). Firstly, a new sliding surface is proposed which could enable the system to achieve predefined time convergence during both the arrival and sliding stages. Secondly, a new interval type-2 fuzzy emotional neural network (IT2FENN) is designed, which can be seen as integrating fuzzy neural networks (FNNs) into a regular emotional neural networks (ENNs) to improve the nonlinear approximation ability of the controller, meanwhile, the predefined time convergence is still guaranteed. Finally, the stability analysis for the whole system is presented by using the Lyapunov stability theorem. Three sets of simulation experiments were designed in this study. The experiments demonstrate that combining the designed controller with IT2FENN not only ensures the convergence of the system state within a predefined time but also shows that IT2FENN has stronger approximation capabilities and faster training speeds compared to traditional radial basis function neural network (RBFNN) and ENN.}, archive = {J_NEUCOM}, author = {Chao Jia and Zhiwen Yang and Zhiliang Shao and Yashuai Li}, doi = {10.1016/j.neucom.2024.129173}, journal = {Neurocomputing}, month = {2}, pages = {129173}, shortjournal = {Neurocomputing}, title = {Predefined time sliding mode control based on interval type-2 fuzzy emotional neural network}, volume = {619}, year = {2025}, }
• (2025). Scribble-supervised active learning for microscopy instance segmentation. NEUCOM, 619, 129169. (www)

  @article{ , abstract = {Accurate manual annotation for large-scale microscopy instance segmentation is time-consuming and expensive. Recently, weakly supervised learning has been proposed to reduce the annotation cost by exploiting weak annotation, such as scribble. However, existing weakly supervised learning methods work on randomly selected annotation and cannot fully exploit hard samples, resulting in unsatisfying performance. In this paper, we propose a two-stage, scribble-supervised active learning method for microscopy instance segmentation to automatically exploit hard samples and reduce the annotation cost. First, we propose a cross pseudo-label strategy to warm up the segmentation network leveraging a small number of scribbles, which generates the initial instances. Second, we propose an instance-based region selection strategy via active learning to exploit hard samples and further improve the segmentation performance. Specifically, we iteratively select hard regions and correct segmentation errors in these regions with a few scribbles to update the network parameters. Extensive experiments on 2D and 3D microscopy datasets demonstrate the superiority of our proposed method over existing weakly supervised learning and active learning methods.}, archive = {J_NEUCOM}, author = {Miaomiao Cai and Xiaoyu Liu and Wei Huang and Shenglong Zhou and Yueyi Zhang and Xuejin Chen and Zhiwei Xiong}, doi = {10.1016/j.neucom.2024.129169}, journal = {Neurocomputing}, month = {2}, pages = {129169}, shortjournal = {Neurocomputing}, title = {Scribble-supervised active learning for microscopy instance segmentation}, volume = {619}, year = {2025}, }
• (2025). Enhanced graph diffusion learning with dynamic transformer for anomaly detection in multivariate time series. NEUCOM, 619, 129168. (www)

  @article{ , abstract = {In recent years, deep learning-based multivariate time series anomaly detection methods have significantly improved detection performance by accurately modeling the spatiotemporal correlations in sensor data. However, they still face the following challenges: the complexity of signals acquired from sensors makes it difficult to capture the true hidden spatial relationships between sensors. Local features and global correlations of time series data are ignored in capturing accurately the temporal correlations of multivariate time series. To address these problems, we propose a G raph D iffusion T ransformed S patiotemporal model( GDTS ) for multivariate time series anomaly detection. Specifically, we first extend a newly developed diffusion probability model to spatial correlation modeling of sensors, which results in a new graph diffusion network. During the diffusion process, we design a novel L2 regularized weighted dot product diffusion function with energy constraints to guide information propagation for accurately capturing the hidden spatial dependencies between sensors. Moreover, we develop a variant of the transformer model with a hybrid sampling strategy to capture the local features and global correlations of time series, which comprehensively describes the temporal correlation of time series data. Subsequently, the multivariate time series features are jointly optimized at the spatio-temporal level for prediction. Finally, the threshold method is used to measure the deviation between the actual observed value and the predicted value to accomplish anomaly detection. Extensive experiments on real-world publicly available datasets demonstrate that GDTS significantly outperforms several state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Rong Gao and Jiming Wang and Yonghong Yu and Jia Wu and Li Zhang}, doi = {10.1016/j.neucom.2024.129168}, journal = {Neurocomputing}, month = {2}, pages = {129168}, shortjournal = {Neurocomputing}, title = {Enhanced graph diffusion learning with dynamic transformer for anomaly detection in multivariate time series}, volume = {619}, year = {2025}, }
• (2025). RobustMixGen: Data augmentation for enhancing robustness of visual–language models in the presence of distribution shift. NEUCOM, 619, 129167. (www)

  @article{ , abstract = {With the increasing deployment of Vision–Language Models (VLMs) in real-world applications, there is growing interest in enhancing their robustness to noise. Data augmentation has emerged as a prominent approach for improving robustness, and in the context of VLMs, MixGen has been widely adopted. Despite its success in improving performance, our experiments indicate that MixGen significantly degrades performance under distribution shift conditions, primarily due to the model’s reliance on spurious correlations induced by MixGen-augmented data. To address this limitation, we propose a novel augmentation method that enhances both model performance and robustness by mitigating the learning of spurious correlations. Our approach involves the pre-classification of object and background categories. For image synthesis, we introduce the CutMixup technique, while for text synthesis, we employ a conjunction concatenation strategy, both aimed at reducing the impact of spurious correlations. We evaluated the efficacy of our method using the COCO dataset, a large-scale benchmark comprising images and text. The effectiveness of our approach was assessed in a retrieval task under simulated distribution shift conditions. Our experimental results demonstrate the superiority of the proposed method, with a 17.11% improvement in the robustness metric (MMI) under distribution shift scenarios, establishing it as a more effective data augmentation technique. We would like to broaden the applicability of the augmentation method to various vision–language tasks beyond retrieval.}, archive = {J_NEUCOM}, author = {Sunwoo Kim and Hun Im and Woojun Lee and Seonggye Lee and Pilsung Kang}, doi = {10.1016/j.neucom.2024.129167}, journal = {Neurocomputing}, month = {2}, pages = {129167}, shortjournal = {Neurocomputing}, title = {RobustMixGen: Data augmentation for enhancing robustness of visual–language models in the presence of distribution shift}, volume = {619}, year = {2025}, }
• (2025). Combining long and short spatiotemporal reasoning for deep reinforcement learning. NEUCOM, 619, 129165. (www)

  @article{ , abstract = {Improving sample efficiency in deep reinforcement learning is a crucial challenge in sequential decision-making. Finding a rich representation learning method for sequential samples is an urgent need for visual reinforcement learning. Previous research shows that computational complexity or over-parameterization prevents agents from learning long-term spatiotemporal properties, which leads to a substantial decrease in sample usage. Considering these challenges, we provide an online reinforcement learning framework that integrates both short- and long-term dependencies to effectively describe spatiotemporal properties.We propose a method, which we name Spatio-Temporal Reasoning and Memory (STRM), that aims at reconciling these relationships. Specifically, the short-term spatiotemporal feature module extracts local spatiotemporal features and their relationships using a 3D convolutional neural network combined with a self-attention mechanism. In contrast, the long-term spatiotemporal feature module employs a Transformer-based external memory network, using the short-term spatiotemporal features as input. Ultimately, a comprehensive state representation is produced by combining these various aspects. The efficacy of our framework has been empirically validated, exhibiting markedly better performance than previous approaches.}, archive = {J_NEUCOM}, author = {Huiling Liu and Peng Liu and Chenjia Bai}, doi = {10.1016/j.neucom.2024.129165}, journal = {Neurocomputing}, month = {2}, pages = {129165}, shortjournal = {Neurocomputing}, title = {Combining long and short spatiotemporal reasoning for deep reinforcement learning}, volume = {619}, year = {2025}, }
• (2025). Multi-channel GCN network based on position-aware gating fusion for aspect sentiment triplet extraction. NEUCOM, 619, 129164. (www)

  @article{ , abstract = {Aspect sentiment triplet extraction (ASTE) is a crucial subtask of aspect-based sentiment analysis (ABSA). Each triplet consists of an aspect term, an opinion term, and the sentiment polarity of the associated aspect term. Existing approaches typically address ASTE tasks end-to-end through table-filling methods. However, these methods often fail to capture the interactions between words in the sentence and the extraction of implicit words. To overcome this limitation, we propose a multi-channel GCN network based on position-aware gating fusion (MPGF-GCN) to extract aspect sentiment triplets. We design a multi-channel GCN module integrated with aspect term extraction (ATE) and opinion term extraction (OTE) tasks. Each extraction module consists of two GCN components: a feature graph convolutional network (FGCN) and an original graph convolutional network (OGCN). The FGCN structures, combined with the refining strategy, merge adjacent nodes within each extraction module and facilitate information exchange between nodes. In addition, the model effectively integrates the aspect term and opinion term information captured by ATE and OTE through the position-aware gating fusion module (PagFM). We conducted extensive experiments on the benchmark dataset and the results demonstrate that our proposed method performed well.}, archive = {J_NEUCOM}, author = {Kun Yang and Xiao Wang and Bin Gao and Linlin Li and Shutian Liu and Zhengjun Liu}, doi = {10.1016/j.neucom.2024.129164}, journal = {Neurocomputing}, month = {2}, pages = {129164}, shortjournal = {Neurocomputing}, title = {Multi-channel GCN network based on position-aware gating fusion for aspect sentiment triplet extraction}, volume = {619}, year = {2025}, }
• (2025). Model compression through distillation with cross-layer integrated guidance at word level. NEUCOM, 619, 129162. (www)

  @article{ , abstract = {In various academic and applied domains including software engineering, lightweight software applications can be facilitated by knowledge distillation, which involves the transfer of insights from a teacher model to a student model, enabling the latter to gradually replicate the behavior of the former, thus achieving model compression and acceleration. Nonetheless, due to differences in the capacity between teacher and student models, the mixed types of knowledge in hidden representations, and variations in the proportion of knowledge in the hidden representations of each layer, there is still room for improvement in knowledge distillation. This paper proposes a Cross-layer Integration of Word-level Association (CI-WA) knowledge distillation. Firstly, CI-WA introduces an extractor constructed by dynamic sparse attention, which extracts task-related word-level associations from the hidden vectors at each layer, mitigating the influence of task-irrelevant information. Secondly, this paper introduces cross-layer connections into the knowledge distillation process, enhancing the performance of student model by jointly leveraging high-level and low-level features. Finally, the proposed method is validated on two tasks: natural language understanding and modeling. Experimental results illustrate that the proposed method surpasses state-of-the-art techniques, and language modeling provides a new perspective for knowledge distillation.}, archive = {J_NEUCOM}, author = {Guiyu Li and Shang Zheng and Haitao Zou and Hualong Yu and Shang Gao}, doi = {10.1016/j.neucom.2024.129162}, journal = {Neurocomputing}, month = {2}, pages = {129162}, shortjournal = {Neurocomputing}, title = {Model compression through distillation with cross-layer integrated guidance at word level}, volume = {619}, year = {2025}, }
• (2025). Multivariate hawkes spatio-temporal point process with attention for point of interest recommendation. NEUCOM, 619, 129161. (www)

  @article{ , abstract = {Point-of-interest (POI) recommendation is one of the most essential services in modern Location-Based Social Networks (LBSNs) to address the problem of information overload presented in various online applications. However, existing methods may ignore the spatio-temporal context, especially the interplay between time and space in user behavior patterns, and treat them as independent features. To fully exploit the spatio-temporal information for accurate POI recommendations, we propose a M ultivariate H awkes S patio- T emporal P oint P rocess with a ttention (MHSTPP-a), which predicts the user’s preference for the next POI by jointly modeling the correlations between users and POIs. In particular, a novel and lightweight representation learning approach is designed to model user interaction sequences from the perspective of spatio-temporal point processes. Specifically, MHSTPP-a first learn the informative embeddings of users and POIs based on historical check-in records. Then, a multivariate Hawkes process with the attention mechanism is designed to model the spatio-temporal dependencies and correlations between check-in behaviors in continuous time and space and learn users’ general and contextual preferences. Finally, MHSTPP-a can perform accurate next POI recommendations according to personal preferences. Extensive experiments carried out on three real-world datasets demonstrate that the proposed approach MHSTPP-a outperforms baselines, including several state-of-the-art methods in various evaluation metrics.}, archive = {J_NEUCOM}, author = {Xin Zhang and He Weng and Yuxin Wei and Dongjing Wang and Jia Chen and Tingting Liang and Yuyu Yin}, doi = {10.1016/j.neucom.2024.129161}, journal = {Neurocomputing}, month = {2}, pages = {129161}, shortjournal = {Neurocomputing}, title = {Multivariate hawkes spatio-temporal point process with attention for point of interest recommendation}, volume = {619}, year = {2025}, }
• (2025). TtfDiffusion: Training-free and text-free image editing in diffusion models with structural and semantic disentanglement. NEUCOM, 619, 129159. (www)

  @article{ , abstract = {Diffusion models achieve outstanding performance in various domains, particularly in the field of controlled image editing. Despite their great success, these models typically require both fine-tuning and the use of edited text prompts as conditions. However, fine-tuning diffusion models is highly time-consuming, and suitable edited text prompts may not be readily available or easy to devise in specific domains such as art, fashion, or specialized fields like medicine. To address this issue, we present our gradient-based method called TtfDiffusion, which is both t raining-free and t ext- f ree. This method enables efficient and effective semantic control using frozen pre-trained DDMs, eliminating the need for additional training and edited text prompts. Our method efficiently uncovers disentangled directions to separate the structure and semantics of images within the h-space of frozen pre-trained diffusion models during the denoising process, significantly enhancing image editing performance. Moreover, our method can improve images with poor original generation quality while editing semantics. It is applicable to various architectures, including DDPM, Stable Diffusion-v1.5, and Stable Diffusion-v2.1, and to datasets such as CelebA-HQ, LSUN-church, and LSUN-bedroom.}, archive = {J_NEUCOM}, author = {Zhenbo Yu and Jian Jin and Jinhan Zhao and Zhenyong Fu and Jian Yang}, doi = {10.1016/j.neucom.2024.129159}, journal = {Neurocomputing}, month = {2}, pages = {129159}, shortjournal = {Neurocomputing}, title = {TtfDiffusion: Training-free and text-free image editing in diffusion models with structural and semantic disentanglement}, volume = {619}, year = {2025}, }
• (2025). Skeleton-OOD: An end-to-end skeleton-based model for robust out-of-distribution human action detection. NEUCOM, 619, 129158. (www)

  @article{ , abstract = {Human action recognition is crucial in computer vision systems. However, in real-world scenarios, human actions often fall outside the distribution of training data, requiring a model to both recognize in-distribution (ID) actions and reject out-of-distribution (OOD) ones. Despite its importance, there has been limited research on OOD detection in human actions. Existing works on OOD detection mainly focus on image data with RGB structure, and many methods are post-hoc in nature. While these methods are convenient and computationally efficient, they often lack sufficient accuracy, fail to consider the exposure of OOD samples, and ignore the application in skeleton structure data. To address these challenges, we propose a novel end-to-end skeleton-based model called Skeleton-OOD, which is committed to improving the effectiveness of OOD tasks while ensuring the accuracy of ID recognition. Through extensive experiments conducted on NTU-RGB+D 60, NTU-RGB+D 120, and Kinetics-400 datasets, Skeleton-OOD demonstrates the superior performance of our proposed approach compared to state-of-the-art methods. Our findings underscore the effectiveness of classic OOD detection techniques in the context of skeleton-based action recognition tasks, offering promising avenues for future research in this field. Code is available at https://github.com/YilliaJing/Skeleton-OOD.git .}, archive = {J_NEUCOM}, author = {Jing Xu and Anqi Zhu and Jingyu Lin and Qiuhong Ke and Cunjian Chen}, doi = {10.1016/j.neucom.2024.129158}, journal = {Neurocomputing}, month = {2}, pages = {129158}, shortjournal = {Neurocomputing}, title = {Skeleton-OOD: An end-to-end skeleton-based model for robust out-of-distribution human action detection}, volume = {619}, year = {2025}, }
• (2025). Neuroadaptive control achieving zero-error tracking and designated performance—A novel vanishing damping approach. NEUCOM, 619, 129157. (www)

  @article{ , abstract = {In this work, we present a novel control approach for uncertain nonlinear systems that is capable of steering the tracking errors towards a desired residual region within pre-specified time, and thereafter further regulating the errors asymptotically to zero. This is achieved by using prescribed time-varying function (PtvF) based transformation, together with neural adaptive law featuring a vanishing damping term. The role of the PtvF transformation is to drive the system errors (from any initial condition) into the desirable bounded region within user-assignable time (rather than infinite time as in existing prescribed performance control (PPC) methods). Whereas, the role of the neural adaptive law with a damping term that vanishes within prescribed time, combined with the “softsign” mechanism in the control, is to further asymptotically regulate the errors to zero, rather than to some stipulated or unknown ultimately uniformly bounded region (which again differs from most existing PPC methods). By using Lyapunov function incorporated with the lower bounds of control gains, a rigorous analysis of the closed-loop system’s stability is established. The results are further verified and validated through simulations.}, archive = {J_NEUCOM}, author = {Kaili Xiang and Yongduan Song}, doi = {10.1016/j.neucom.2024.129157}, journal = {Neurocomputing}, month = {2}, pages = {129157}, shortjournal = {Neurocomputing}, title = {Neuroadaptive control achieving zero-error tracking and designated performance—A novel vanishing damping approach}, volume = {619}, year = {2025}, }
• (2025). Double center swarm exploring varying parameter neurodynamic network for non-convex nonlinear programming. NEUCOM, 619, 129156. (www)

  @article{ , abstract = {To solve non-convex nonlinear programming problems, a double center swarm exploring varying parameter neurodynamic network (DCSE-VPNN) is proposed and analyzed. Firstly, a varying parameter neurodynamic network is proposed as a solver for nonlinear programming to seek local optimal solutions. Secondly, a double center particle swarm optimization algorithm is exploited, wherein each neural network serves as a particle. Each particle independently explores a local optimal solution. Through information exchange among particles, the subsequent positions to be explored are updated. As a result, DCSE-VPNN acquires the capability of global search. Computer simulation experiments verify the efficacy of the proposed approach in solving non-convex nonlinear programming problems. In comparison with two existing methods, the results show that the proposed DCSE-VPNN approach has fewer iterations and higher search accuracy.}, archive = {J_NEUCOM}, author = {Zhijun Zhang and Minghao Zhu and Xiaohui Ren}, doi = {10.1016/j.neucom.2024.129156}, journal = {Neurocomputing}, month = {2}, pages = {129156}, shortjournal = {Neurocomputing}, title = {Double center swarm exploring varying parameter neurodynamic network for non-convex nonlinear programming}, volume = {619}, year = {2025}, }
• (2025). A center-anchored adaptive hierarchical graph neural network with application in structure-aware recognition of enzyme catalytic specificity. NEUCOM, 619, 129155. (www)

  @article{ , abstract = {Non-Euclidean modeling of biomolecules such as enzymes has been advanced by utilizing graph neural networks (GNNs). While existing whole-molecule modeling approaches offer significant advantages, they present two primary challenges: optimizing long-range interactions within the enzyme molecule and ensuring model interpretability. Furthermore, current GNN-based enzyme modeling methods pay insufficient attention to the hierarchical structure of biological functions. To address these challenges, we propose a novel Center-Anchored Adaptive Hierarchical GNN (CAAH-GNN) model that integrates a center-anchored method with an adaptive hierarchical strategy for structure-aware recognition of enzyme catalytic specificity. Our model features three key characteristics: (1) It autonomously identifies and integrates local key residues within enzyme molecules while embedding global graph information into each node's features, significantly enhancing prediction accuracy. (2) The spherical spatial radius is adaptively learned using a two-step hierarchically sampling method. Compared to full-molecule models, this approach reduces spatial complexity more than 80 %. (3) Our model exhibits strong generalization performance and interpretability, as it is not tailored to specific enzyme families and does not remove any nodes during training. Extensive experiments based on two enzyme datasets and various graph convolution kernels demonstrate that CAAH-GNN outperforms baseline models and other hierarchical strategy models in prediction accuracy. In summary, our research not only introduces a novel GNN-based enzyme modeling approach but also provides a new perspective and a tool for enzyme engineering.}, archive = {J_NEUCOM}, author = {Yi Li and Yahui Liu and Jiangchun Hou and Xiaohui Liu}, doi = {10.1016/j.neucom.2024.129155}, journal = {Neurocomputing}, month = {2}, pages = {129155}, shortjournal = {Neurocomputing}, title = {A center-anchored adaptive hierarchical graph neural network with application in structure-aware recognition of enzyme catalytic specificity}, volume = {619}, year = {2025}, }
• (2025). HRPVT: High-resolution pyramid vision transformer for medium and small-scale human pose estimation. NEUCOM, 619, 129154. (www)

  @article{ , abstract = {Human pose estimation on medium and small scales has long been a significant challenge in this field. Most existing methods focus on restoring high-resolution feature maps by stacking multiple costly deconvolutional layers or by continuously aggregating semantic information from low-resolution feature maps while maintaining high-resolution ones, which can lead to information redundancy. Additionally, due to quantization errors, heatmap-based methods have certain disadvantages in accurately locating keypoints of medium and small-scale human figures. This study proposes HRPVT, which utilizes PVT v2 as the backbone to model long-range dependencies. Building on this, the authors introduce the High-Resolution Pyramid Module (HRPM), designed to generate higher quality high-resolution representations by incorporating the intrinsic inductive biases of Convolutional Neural Networks (CNNs) into the high-resolution feature maps. The integration of HRPM enhances the performance of pure transformer-based models for human pose estimation at medium and small scales. Furthermore, the heatmap-based method is replaced with SimCC approach, which eliminates the need for costly upsampling layers. To accommodate models with varying parameter scales, the authors have developed two insertion strategies of HRPM, each designed to enhance the model’s ability to perceive medium and small-scale human poses from two distinct perspectives. The proposed method achieved scores of 76.3 AP and 75.5 AP on the MS COCO Keypoint validation and test-dev datasets, respectively, while reducing the number of parameters by 60% and the GFLOPs by 62% compared to HRNet-W48. Furthermore, the proposed method achieved the higher score in the AP M metric among most state-of-the-art methods, validating its superiority in medium and small-scale human pose estimation.}, archive = {J_NEUCOM}, author = {Zhoujie Xu and Meng Dai and Qing Zhang and Xiaodi Jiang}, doi = {10.1016/j.neucom.2024.129154}, journal = {Neurocomputing}, month = {2}, pages = {129154}, shortjournal = {Neurocomputing}, title = {HRPVT: High-resolution pyramid vision transformer for medium and small-scale human pose estimation}, volume = {619}, year = {2025}, }
• (2025). Cycle pixel difference network for crisp edge detection. NEUCOM, 619, 129153. (www)

  @article{ , abstract = {Edge detection, as a fundamental task in computer vision, has garnered increasing attention. The advent of deep learning has significantly advanced this field. However, recent deep learning-based methods generally face two significant issues: (1) reliance on large-scale pre-trained weights, and (2) generation of thick edges. We construct a U-shape encoder–decoder model named CPD-Net that successfully addresses these two issues simultaneously. In response to issue (1), we propose a novel cycle pixel difference convolution (CPDC), which effectively integrates edge prior knowledge with modern convolution operations, consequently successfully eliminating the dependence on large-scale pre-trained weights. As for issue (2), we construct a multi-scale information enhancement module (MSEM) and a dual residual connection-based (DRC) decoder to enhance the edge location ability of the model, thereby generating crisp and clean contour maps. Comprehensive experiments conducted on four standard benchmarks demonstrate that our method achieves competitive performance on the BSDS500 dataset (ODS = 0.813 and AC = 0.352), NYUD-V2 (ODS = 0.760 and AC = 0.223), BIPED dataset (ODS = 0.898 and AC = 0.426), and CID (ODS = 0.59). Our approach provides a novel perspective for addressing these challenges in edge detection.}, archive = {J_NEUCOM}, author = {Changsong Liu and Wei Zhang and Yanyan Liu and Mingyang Li and Wenlin Li and Yimeng Fan and Xiangnan Bai and Liang Zhang}, doi = {10.1016/j.neucom.2024.129153}, journal = {Neurocomputing}, month = {2}, pages = {129153}, shortjournal = {Neurocomputing}, title = {Cycle pixel difference network for crisp edge detection}, volume = {619}, year = {2025}, }
• (2025). Flocking with nonlinear systems based on distributed observers. NEUCOM, 619, 129152. (www)

  @article{ , abstract = {This paper investigates the output feedback-based flocking control problem for nonlinear dynamical systems (NDSs). To address this, distributed observers and consensus-based leader-following control algorithms are incorporated into each sensor in the network. By combining state estimation with flocking control algorithms, the observed state information of the NDS serves as the virtual leader for each sensor. To handle scenarios involving obstacles, we propose enclosing obstacles of various shapes within circles and developing a flocking control algorithm using artificial potential functions. Simulation results validate the effectiveness of the proposed algorithm.}, archive = {J_NEUCOM}, author = {Jun Chi and Qing An and Xiaoling Wang and Housheng Su}, doi = {10.1016/j.neucom.2024.129152}, journal = {Neurocomputing}, month = {2}, pages = {129152}, shortjournal = {Neurocomputing}, title = {Flocking with nonlinear systems based on distributed observers}, volume = {619}, year = {2025}, }
• (2025). MA-net: Resource-efficient multi-attentional network for end-to-end speech enhancement. NEUCOM, 619, 129150. (www)

  @article{ , abstract = {Deep Neural Networks (DNNs) have transformed speech enhancement (SE) by solving the complex relationships within speech signals through their multi-layered hierarchical representations. However, their computational demands remain a challenging problem. Self-attention has emerged as a key technique for capturing long-range dependencies in speech signals by measuring attention between vectors through scaled-dot products. Despite its widespread utility across various domains, self-attention encounters limitations when applied to SE. Specifically, its efficiency reduces in low signal-to-noise ratio (SNR) conditions due to its sensitivity to the scale of input vectors influenced by factors such as low SNRs. To address these challenges, we propose a resource-efficient Multi-Attention Network (MA-Net) speech enhancement model to effectively capture local and long-range dependencies in speech signals, while maintaining a low computational footprint. MA-Net integrates two fundamental modules: Spectral Temporal Hybrid Attention (STHA) and Dynamic Feedback Shuffle Attention (DFSA). The STHA module is designed to model long-range dependencies in spectral and temporal features by using hybrid self-attention (HSA). This mechanism computes attention weights between query ( Q ) and key ( K ) vectors using dot-product and cosine similarity scores to mitigate the impact of scale variations in input vectors, enabling more consistent and reliable attention mechanisms. The DFSA module iteratively applies channel and spatial attention to dynamically refine feature representations by adjusting the weight of each iteration’s output based on input spectral features. Evaluations performed on two benchmark datasets (WSJ0-SI84 and VCTK+DEMAND) show that the MA-Net outperforms recent models in terms of SE performance at a considerably reduced computational complexity, with 0.92M parameters, 0.09 RTF, and 1.32G/s MACs. On the WSJ0-SI84 dataset, MA-Net improves PESQ, STOI, and SI-SDR by 1.26, 20.3%, and 9.76 dB over noisy mixtures, highlighting the usefulness of MA-Net in real-world SE conditions.}, archive = {J_NEUCOM}, author = {Fazal E Wahab and Zhongfu Ye and Nasir Saleem and Rizwan Ullah and Amir Hussain}, doi = {10.1016/j.neucom.2024.129150}, journal = {Neurocomputing}, month = {2}, pages = {129150}, shortjournal = {Neurocomputing}, title = {MA-net: Resource-efficient multi-attentional network for end-to-end speech enhancement}, volume = {619}, year = {2025}, }
• (2025). SDVS-net: A spatial dilated convolution and variable self-attention network for multivariate long-term time series forecasting. NEUCOM, 619, 129148. (www)

  @article{ , abstract = {Multivariate long-term time series data exhibit complex time and spatial (variable) characteristics. Transformer-based methods have achieved some success in multivariate long-term time series forecasting (MTSF) tasks. However, the embedding strategy of transformers tends to amalgamate variables, leading to the loss of variable dependency features. Additionally, these methods overly emphasize the extraction of time features, neglecting to explore the complex dependencies among variables. Therefore, MTSF poses an important and challenging problem in time series analysis. This paper proposes a spatial dilated convolution and variable self-attention network for MTSF, called SDVS-Net. SDVS-Net thoroughly considers multivariate long-term time series’ time and spatial (variable) features. Spatial dilated convolution (SDC) is employed to extract local features among variables to fully exploit the local dependencies among variables. To further capture the global dependencies among variables, Multivariate multi-head attention (MMA) is introduced, utilizing multi-head mechanisms to supplement the dependencies of each variable at different times. Time multilayer perceptron(TMP) with enhanced features from hidden layers is employed to extract features from the time dimension further. Finally, a single-layer feedforward neural network (FNN) outputs the prediction series. The experimental results on six benchmark datasets from different real-world domains demonstrate that, compared to state-of-the-art methods, SDVS-Net achieves higher prediction accuracy in MTSF problems.}, archive = {J_NEUCOM}, author = {Xiaoyi Tian and Siyuan Huang and Jinjie Xiao and Han Wang and Yepeng Liu}, doi = {10.1016/j.neucom.2024.129148}, journal = {Neurocomputing}, month = {2}, pages = {129148}, shortjournal = {Neurocomputing}, title = {SDVS-net: A spatial dilated convolution and variable self-attention network for multivariate long-term time series forecasting}, volume = {619}, year = {2025}, }
• (2025). Quasi-synchronization of discrete-time tempered fractional-order memristive neural networks with time delays. NEUCOM, 619, 129146. (www)

  @article{ , abstract = {Due to the finite lifespan of particles or constraints imposed by physical space, research on tempered fractional calculus has become increasingly popular. In this paper, to enhance both the practicality and precision of our model, we combine tempered fractional operators with discrete-time memristive neural networks (DTFMNNs) to investigate synchronization issues. With the aid of discrete convolution, we derive the discrete Laplace transform formulas for discrete tempered fractional calculus. Next, by combining the inequality technique with the Lyapunov function method and Laplace transform formulas, some easily verifiable sufficient conditions are established to ensure synchronization of DTFMNNs under the proposed controller. Finally, numerical simulations are conducted to validate the effectiveness of the theoretical results.}, archive = {J_NEUCOM}, author = {Xiao-Li Zhang and Yongguang Yu and Hu Wang and Di Nie}, doi = {10.1016/j.neucom.2024.129146}, journal = {Neurocomputing}, month = {2}, pages = {129146}, shortjournal = {Neurocomputing}, title = {Quasi-synchronization of discrete-time tempered fractional-order memristive neural networks with time delays}, volume = {619}, year = {2025}, }
• (2025). A novel span-based knowledge-enhanced framework for aspect sentiment triplet extraction. NEUCOM, 619, 129145. (www)

  @article{ , abstract = {Aspect Sentiment Triplet Extraction (ASTE) is a subtask of Aspect-based Sentiment Analysis, which aims to identify all aspect sentiment triplets in given sentences. ASTE is an important research task to discover sentimental opinions in online social media. Existing ASTE models are usually based on pipeline or joint strategy. The pipeline-based methods may suffer from error propagation. The joint-based methods could avoid error propagation in an end-to-end manner, which have become the more popular choice. Among these joint-based ASTE models, the syntactic information from the dependency tree is widely used. However, the dependency tree may fail to connect the aspect and opinion terms in some cases, which causes the lack of interaction in triplets. Inspired by the work of Liang et al. on the Aspect Sentiment Classification (ASC) task, we integrate common-sense knowledge into the ASTE task, which can bring more important information to the modeling of context representation. To address these limitations, this paper first involves common-sense sentiment knowledge along with syntactic dependency knowledge to get better representations of contexts. The syntactic and common-sense knowledge can contribute to the interpretability when extracting the sentiment elements. We also introduce a self-attention mechanism based on the orthogonal loss function to better capture the interactions between words. We evaluated our model on four public datasets with recently proposed baselines in F1-score metric, especially 3.62%, 4.14%, 3.58% and 2.22% higher on 14Res, 14Lap, 15Res, and 16Res, compared to the SSJE approach, respectively. Experimental results show that our method outperforms SOTA models on three datasets.}, archive = {J_NEUCOM}, author = {Heng-yang Lu and Rui Cong and Wei Nie and Tian-ci Liu and Wei Fang}, doi = {10.1016/j.neucom.2024.129145}, journal = {Neurocomputing}, month = {2}, pages = {129145}, shortjournal = {Neurocomputing}, title = {A novel span-based knowledge-enhanced framework for aspect sentiment triplet extraction}, volume = {619}, year = {2025}, }
• (2025). Anatomical structures detection using topological constraint knowledge in fetal ultrasound. NEUCOM, 619, 129143. (www)

  @article{ , abstract = {The accurate recognition of anatomical structures in fetal ultrasound images is crucial for prenatal diagnosis and determining ultrasound standard planes. However, this task can be arduous and time-consuming for sonographers. Existing methods usually treat anatomical structures as isolated individuals, ignore their interconnections, and lead to false and missed detections due to the similarities between these structures. In this paper, we propose a key anatomical structure-aware module, named KASM, which consists of class-aware representation learning and topological information interaction and is designed to integrate seamlessly into any proposal feature-based detection framework. Given the potential issue of information interactions between proposal features leading to feature over-smoothing, our approach adopts class-aware representation learning as an initial step. This process aims to enhance the intra-class compactness of proposal features while simultaneously dispersing inter-class features. The objective is to derive highly discriminative proposal features, thus facilitating subsequent accurate information interactions. Subsequently, proposal features are modeled using topological constraint relationships between anatomical structures and informative interactions are performed for structure recognition. We conduct extensive experiments on collected fetal ultrasound datasets in femoral, abdominal, and thalamic planes. The results consistently demonstrate that KASM effectively mitigates issues such as false detection and missed detection of anatomical structures across different baseline models, e.g., integrating KASM in Faster R-CNN reduces the false detection rate from 49.7% to 38.4% (−11.3%) and the missed detection rate from 60.8% to 57.0% (−3.8%) under the confidence threshold of 0.2. This underscores the broad applicability and efficacy of KASM in addressing these challenges.}, archive = {J_NEUCOM}, author = {Juncheng Guo and Guanghua Tan and Jianxin Lin and Bin Pu and Xin Wen and Chunlian Wang and Shengli Li and Kenli Li}, doi = {10.1016/j.neucom.2024.129143}, journal = {Neurocomputing}, month = {2}, pages = {129143}, shortjournal = {Neurocomputing}, title = {Anatomical structures detection using topological constraint knowledge in fetal ultrasound}, volume = {619}, year = {2025}, }
• (2025). Micro-expression recognition using dual-view self-supervised contrastive learning with intensity perception. NEUCOM, 619, 129142. (www)

  @article{ , abstract = {Micro-expressions, as indicators of true emotions, have significant applications in medical care and public safety. These expressions are characterized by their short duration, low intensity, and localized occurrence. These characteristics lead to the small sample problem in micro-expressions, making feature learning challenging and limiting the improvement of recognition performance. To address this issue, we propose a multimodal contrastive learning pre-training model based on Action Unit (AU) intensity perception. We conducted an experiment to determine the minimum threshold for recognizing facial expressions. Using this threshold, we filtered a large volume of unsupervised samples. The first stage involves unsupervised multimodal contrastive learning, where the model learns from differences in facial actions across various modalities. Subsequently, the model is trained on the micro-expression recognition task using a small amount of labeled data, overcoming the limitations of small sample sizes. Comparative experiments using the MEGC2019-CD and the multimodal dataset CAS(ME) 3 datasets demonstrate the superiority of our method. Our method is available at https://github.com/MELABIPCAS/DVSCL.git .}, archive = {J_NEUCOM}, author = {Jingting Li and Haoliang Zhou and Yu Qian and Zizhao Dong and Su-Jing Wang}, doi = {10.1016/j.neucom.2024.129142}, journal = {Neurocomputing}, month = {2}, pages = {129142}, shortjournal = {Neurocomputing}, title = {Micro-expression recognition using dual-view self-supervised contrastive learning with intensity perception}, volume = {619}, year = {2025}, }
• (2025). Towards unsupervised domain-specific open-world recognition. NEUCOM, 619, 129141. (www)

  @article{ , abstract = {Open-World Recognition (OWR) is an emerging study that constructs machine-learning models to recognize unknown classes and learn them continually. The classical formalization of OWR relies on three main components: classifier, unknown identification, and continual learning. However, for a model that operates on domain-specific tasks, training rejected unknown classes directly will harm the models in terms of effectivity and efficiency (i.e., a waste collector robot will learn unnecessary classes and will collect novel non-waste objects). Filtering these novel objects manually requires human-in-the-loop which is costly and unable to learn on the job. Therefore, in this study, we introduce and formalize Unsupervised Domain-specific Open-world Recognition (UDOR) that has the potential framework to achieve a fully automated agent in an open-world environment. In addition, we formalize the specific component in UDOR called novelty manager to assist the model to learn on the job. Furthermore, we propose a unified model using Continual Multi-Channel Contrastive Prototype Networks (CMCCPN), Automated Machine learning (AutoML), and class discovery with Hierarchical DBSCAN (HDBSCAN) or First Integer Neighbor Clustering Hierarchy (FINCH) as a step towards UDOR. Our experimentation results suggest that CMCCPN produced the highest performance, AutoML provides almost exemplary capability in differentiating novel classes, and Vision Transformer with HDBSCAN or FINCH shows a good technique to be investigated in discovering classes with a small number of classes. Our source code is available at https://github.com/gusti-alfarisy/udor .}, archive = {J_NEUCOM}, author = {Gusti Ahmad Fanshuri Alfarisy and Owais Ahmed Malik and Ong Wee Hong}, doi = {10.1016/j.neucom.2024.129141}, journal = {Neurocomputing}, month = {2}, pages = {129141}, shortjournal = {Neurocomputing}, title = {Towards unsupervised domain-specific open-world recognition}, volume = {619}, year = {2025}, }
• (2025). Learning the beneficial, forgetting the harmful: High generalization reinforcement learning with in evolving representations. NEUCOM, 619, 129139. (www)

  @article{ , abstract = {In visual Reinforcement Learning (RL), one of the key problems is how to learn policies, which can be generalized to unseen environments. Recently, saliency guidance and representation learning have shown excellent performance in improving agent policy generalization performance. However, due to bottlenecks such as the difficulty in utilizing dynamic features, primacy bias, and error accumulation caused by hyperparameter selection in visual RL, further improvements in generalization performance are affected. To alleviate these problems, we propose a novel method: L earning the B eneficial, F orgetting the H armful (LBFH). LBFH is an effective combination of Dynamic model-based Dual Feature Alignment (DDFA) and Reinforcement Learning with Periodic Resets (RLPR). DDFA uses dynamical models and self-supervised learning to learn representations with dynamical features. RLPR uses the saliency-guided Q-network and periodic reset, allowing the agent to focus on task-relevant pixels while periodically releasing accumulated errors. LBFH combines the advantages of both, allowing the agent to forget accumulated errors in the evolving representation, thereby performing high generalization learning. In order to verify the effectiveness of LBFH, we conducted full experiments on multiple tasks in DMControl Generalization Benchmark (DMControl-GB) and Robotic Manipulation. The results show that LBFH outperforms previous state-of-the-art methods when tested on 2 different visual control benchmarks.}, archive = {J_NEUCOM}, author = {Jiawei Zheng and Yonghong Song and Guojiao Lin and Jiayi Duan and Hao Lin and Shuaitao Li}, doi = {10.1016/j.neucom.2024.129139}, journal = {Neurocomputing}, month = {2}, pages = {129139}, shortjournal = {Neurocomputing}, title = {Learning the beneficial, forgetting the harmful: High generalization reinforcement learning with in evolving representations}, volume = {619}, year = {2025}, }
• (2025). Multimodal large model pretraining, adaptation and efficiency optimization. NEUCOM, 619, 129138. (www)

  @article{ , abstract = {Multimodal large models, leveraging extensive datasets and parameters, have provided superior solutions for multimodal tasks and have been widely applied across various domains in everyday life. However, the development of multimodal large models also faces numerous challenges. Among the most critical challenges are how to scientifically and rationally pre-train multimodal large models and how to apply these pre-trained models to downstream tasks more effectively and at a lower cost. In response to these challenges, this paper initially analyzes and summarizes various multimodal pre-training methods employed by state-of-the-art multimodal large models. By analyzing the characteristics these methods, the paper elucidates the different capabilities imparted to multimodal large models by various multimodal pre-training methods. This provides a basis for researchers to select multimodal pre-training methods according to specific needs. Subsequently, the paper categorizes and outlines general adaptation methods for downstream tasks using multimodal pre-trained models. By comparing their advantages and disadvantages, the distinct characteristic of each method are highlighted. Furthermore, we categorize multimodal downstream tasks into five types. Then, for different types of tasks based on their characteristics and application scenarios, we outline suitable adaptation methods, offering researchers a more definitive direction for adaptation studies related to various types of tasks. Additionally, from the perspectives of parameter efficiency, memory efficiency, and efficiency of data utilization, a comparative analysis of research on the efficiency optimization of multimodal large models is conducted. In conclusion, the paper summarizes the shortcomings of existing research and looks forward to the issues and development directions that need to be addressed for multimodal large models. It points out that integrating more modalities into multimodal large models, reducing resource consumption costs, enhancing model explainability and transparency, and addressing privacy and security issues will be hotspots for future research.}, archive = {J_NEUCOM}, author = {Lixia Ji and Shijie Xiao and Jingmei Feng and Wenzhao Gao and Han Zhang}, doi = {10.1016/j.neucom.2024.129138}, journal = {Neurocomputing}, month = {2}, pages = {129138}, shortjournal = {Neurocomputing}, title = {Multimodal large model pretraining, adaptation and efficiency optimization}, volume = {619}, year = {2025}, }
• (2025). A triple-phase boost transformer for industrial equipment fault prediction. NEUCOM, 619, 129137. (www)

  @article{ , abstract = {To address the critical demand for predictive maintenance (PdM) in the industrial scene, our proposed solution is the triple-phase boost transformer (THOR) model. This model achieves accurate time-series multi-label classification and provides an end-to-end solution that includes unsupervised anomaly detection for fault prediction. To further enhance the model’s ability to detect anomalies, we have implemented a min–max loss training approach. This approach amplifies the discrepancy between normal and abnormal patterns, enabling the identification of more obscure anomalies. Additionally, THOR incorporates the use of Discrete Wavelet Transformation (DWT) and designs a DWT attention mechanism for developing frequency-enhanced transformers. This advancement lends greater power to both anomaly detection and capturing temporal signatures for fault prediction. In our real-world case analysis, the effectiveness of the proposed method is verified. With our approach, we achieved precision rates of at least 97.44%, rates of 92.77%, and F1-scores of 0.95 or above, surpassing the performance of other existing fault prediction models in multi-class fault prediction. Therefore, our THOR model offers a valuable solution to the limitations faced by current fault prediction models.}, archive = {J_NEUCOM}, author = {Ruiyun Yu and Mingda Chen and Bin Liu}, doi = {10.1016/j.neucom.2024.129137}, journal = {Neurocomputing}, month = {2}, pages = {129137}, shortjournal = {Neurocomputing}, title = {A triple-phase boost transformer for industrial equipment fault prediction}, volume = {619}, year = {2025}, }
• (2025). Multi-task convolutional deep neural network for recommendation based on knowledge graphs. NEUCOM, 619, 129136. (www)

  @article{ , abstract = {Collaborative filtering suffers from data sparseness and cold start heavily in recommendation systems. Although several methods have tried to use auxiliary information in knowledge graphs to discover key representations of items and users to mitigate the issue mentioned above, there is still a certain space for improvement in their performances. This work proposes a multi-task learning scheme that exploits knowledge graphs to enhance recommendations. This scheme mainly includes recommendation tasks and knowledge graph embedding tasks, both of which are associated via convolutional neural networks to obtain richer potential information. Additionally, to better understand users’ personalized needs and interests, we have incorporated user preference information into item feature representations to enhance recommendation performance. Furthermore, we have integrated relation importance information into entity feature representations, which enables the head entity to aggregate more crucial neighbor information. Experimental results show the proposed algorithm significantly improves the recommendations of MovieLens-1M, Book-Crossing, and Last.FM, compared with ten competing baselines. Meanwhile, we conduct ablation studies to prove the effectiveness of each module in the proposed scheme on performance improvement. We have made the source code of our proposed model available at https://github.com/MYJiang1102/recommendation-code .}, archive = {J_NEUCOM}, author = {Mingyang Jiang and Man Li and Wenming Cao and Mingming Yang and Luxin Zhou}, doi = {10.1016/j.neucom.2024.129136}, journal = {Neurocomputing}, month = {2}, pages = {129136}, shortjournal = {Neurocomputing}, title = {Multi-task convolutional deep neural network for recommendation based on knowledge graphs}, volume = {619}, year = {2025}, }
• (2025). Micro-expression recognition based on direct learning of graph structure. NEUCOM, 619, 129135. (www)

  @article{ , abstract = {Micro-expressions are brief and subtle facial movements that can reveal an individual’s true emotions. In this paper, we introduce the idea of direct learning of graph structures for feature analysis in micro-expression recognition and propose an Optical Flow-based VisionGNN Network (OFVIG-Net). Specifically, The optical flow map leverages the OnsetFrame and ApexFrame of the micro-expression to extract feature information. VisionGNN treats the optical flow map as a graph, where the image is divided into patches treated as nodes, and a graph is constructed by linking neighboring nodes. Feature analysis is then performed based on this graph structure. Together, feature extraction and feature analysis constitute the micro-expression recognition task. We conducted a series of rigorous experiments, and the results show that OFVIG-Net achieves an accuracy of 66.32% on the composite micro-expression databases (SMIC, SAMM, CASMEII). Additionally, it attains the current state-of-the-art (SOTA) accuracy of 88.75% on the CASMÊ3 micro-expression database.}, archive = {J_NEUCOM}, author = {Lijun Zhang and Yifan Zhang and Xinzhi Sun and Weicheng Tang and Xiaomeng Wang and Zhanshan Li}, doi = {10.1016/j.neucom.2024.129135}, journal = {Neurocomputing}, month = {2}, pages = {129135}, shortjournal = {Neurocomputing}, title = {Micro-expression recognition based on direct learning of graph structure}, volume = {619}, year = {2025}, }
• (2025). DFD-NAS: General deepfake detection via efficient neural architecture search. NEUCOM, 619, 129129. (www)

  @article{ , abstract = {As the saying goes, “ seeing is believing ”. However, with the development of digital face editing tools, we can no longer trust what we can see. Though face forgery detection has made promising progress, most current methods are manually designed by human experts, which is labor-intensive. In this paper, we develop an end-to-end framework based on neural architecture search (NAS) for deepfake detection, which can automatically design network architectures without human intervention. First, a forgery-oriented search space is created to choose appropriate operations for this task, which facilitates the search process in finding the most valuable gradient information for face forgery detection. Second, inspired by the fact that the gap between training error and test error is a good indicator of generalization ability for a classification task, we propose a novel performance estimation metric. This metric encourages the error in the performance estimation phase to be close to the error in the search phase, which guides the search process to select more general models. The cross-dataset search is also considered to develop more general architectures. Eventually, we design a cell cascaded pyramid network (C2PN) for final detection, which aggregates multiscale features for performance improvements. Compared with state-of-the-art networks artificially designed, our method achieves competitive performance in both in-dataset and cross-dataset scenarios.}, archive = {J_NEUCOM}, author = {Xiao Jin and Wen Yu and Dai-Wei Chen and Wei Shi}, doi = {10.1016/j.neucom.2024.129129}, journal = {Neurocomputing}, month = {2}, pages = {129129}, shortjournal = {Neurocomputing}, title = {DFD-NAS: General deepfake detection via efficient neural architecture search}, volume = {619}, year = {2025}, }
• (2025). UPL-net: Uncertainty-aware prompt learning network for semi-supervised action recognition. NEUCOM, 619, 129126. (www)

  @article{ , abstract = {This paper focuses on understanding human behavior in videos by reframing the traditional video classification task as a transfer learning problem centered on visual concepts. Unlike existing action recognition approaches that rely solely on single-modal representations and video classifiers, our method leverages an uncertainty-aware prompt learning network (UPL-Net). This network is designed to extract spatiotemporal features that are pertinent to action-related concepts in videos while ensuring that the visual concepts derived from images are preserved. Furthermore, we introduce an uncertainty-guided semi-supervised learning strategy that harnesses unlabeled videos to enhance the model’s generalizability. Extensive experiments conducted on benchmark datasets, namely UCF and HMDB, demonstrate the superiority of our approach over state-of-the-art semi-supervised action recognition methods. Notably, under a 1% labeling rate on the UCF dataset, our method achieves a significant improvement of 12.8%, underscoring its effectiveness in leveraging limited labeled data and abundant unlabeled videos for improved performance.}, archive = {J_NEUCOM}, author = {Shu Yang and Ya-Li Li and Shengjin Wang}, doi = {10.1016/j.neucom.2024.129126}, journal = {Neurocomputing}, month = {2}, pages = {129126}, shortjournal = {Neurocomputing}, title = {UPL-net: Uncertainty-aware prompt learning network for semi-supervised action recognition}, volume = {619}, year = {2025}, }
• (2025). Adversarial training via multi-guidance and historical memory enhancement. NEUCOM, 619, 129124. (www)

  @article{ , abstract = {Deep neural networks (DNNs) are often susceptible to the influence of adversarial examples, potentially leading to severe security issues. Adversarial training stands out as one of the most effective defenses. In this paper, we empirically demonstrate that weakly robust models maintain resilience to adversarial examples generated on a fully trained robust model. Building on this observation, we introduce a novel weight ensemble training method named Multi-Guided Adversarial Training (MGAT). MGAT improves both clean and adversarial accuracy through two guiding strategies: self-guidance and expert-guidance. Self-guidance promotes diversity in decision boundaries among member models, while expert-guidance aligns them more closely with natural decision boundaries. Additionally, MGAT enhances the historical memory of model weights using an exponential moving average (EMA) with a memory factor, thereby better memorizing the weights of weakly robust models. Our experiments demonstrate that MGAT is highly effective in defending against a range of adversarial attacks. Notably, MGAT achieves an accuracy of 55.48% against AutoAttack, marking a 10.92% improvement over standard adversarial training on CIFAR-10.}, archive = {J_NEUCOM}, author = {Chenyu Zhao and Yaguan Qian and Bin Wang and Zhaoquan Gu and Shouling Ji and Wei Wang and Yanchun Zhang}, doi = {10.1016/j.neucom.2024.129124}, journal = {Neurocomputing}, month = {2}, pages = {129124}, shortjournal = {Neurocomputing}, title = {Adversarial training via multi-guidance and historical memory enhancement}, volume = {619}, year = {2025}, }
• (2025). Applications of scientific machine learning for the analysis of functionally graded porous beams. NEUCOM, 619, 129119. (www)

  @article{ , abstract = {This study help investigates different Scientific Machine Learning (SciML) approaches for the analysis of functionally graded (FG) porous beams and compares them under a new framework. The beam material properties are assumed to vary as an arbitrary continuous function. The methods consider the output of a neural network/operator as an approximation to the displacement fields and derive the equations governing beam behavior based on the continuum formulation. They are implemented in the framework and formulated by three approaches: (a) the vector approach leads to a Physics-Informed Neural Network (PINN), (b) the energy approach brings about the Deep Energy Method (DEM), and (c) the data-driven approach, which results in a class of Neural Operator methods. Finally, a neural operator has been trained to predict the response of the porous beam with functionally graded material under any porosity distribution pattern and any arbitrary traction condition. The results are validated with analytical and numerical reference solutions. The data and code accompanying this manuscript will be publicly available at https://github.com/eshaghi-ms/DeepNetBeam .}, archive = {J_NEUCOM}, author = {Mohammad Sadegh Eshaghi and Mostafa Bamdad and Cosmin Anitescu and Yizheng Wang and Xiaoying Zhuang and Timon Rabczuk}, doi = {10.1016/j.neucom.2024.129119}, journal = {Neurocomputing}, month = {2}, pages = {129119}, shortjournal = {Neurocomputing}, title = {Applications of scientific machine learning for the analysis of functionally graded porous beams}, volume = {619}, year = {2025}, }
• (2025). Exploring the landscape of compressed DeepFakes: Generation, dataset and detection. NEUCOM, 619, 129116. (www)

  @article{ , abstract = {In today’s era of social media, where information spreads rapidly through platforms like YouTube, Facebook, and Twitter, the development of generative models have given rise to a phenomenon called DeepFakes. This survey aims to provide a comprehensive overview of compressed DeepFakes research, covering various detection and generation techniques and datasets. It presents the details of detection methods, including experimental settings such as datasets, algorithms, feature selection, and results. The survey also highlights the existing challenges and future directions.}, archive = {J_NEUCOM}, author = {Muhammad Zubair and Saqib Hakak}, doi = {10.1016/j.neucom.2024.129116}, journal = {Neurocomputing}, month = {2}, pages = {129116}, shortjournal = {Neurocomputing}, title = {Exploring the landscape of compressed DeepFakes: Generation, dataset and detection}, volume = {619}, year = {2025}, }
• (2025). Domain knowledge boosted adaptation: Leveraging vision-language models for multi-source domain adaptation. NEUCOM, 619, 129114. (www)

  @article{ , abstract = {Multi-source domain adaptation (MSDA) aims to adapt a model trained on multiple labeled source domains to an unlabeled target domain. Existing MSDA methods primarily focus on reducing domain gaps by aligning the source domains with the target domain, either jointly or separately. However, these methods often distort semantic-related features and overlook the valuable domain-related information present in diverse domains. In this paper, we propose a novel MSDA method called Domain Knowledge Boosted Adaptation (DKBA) that leverages domain-related information to enhance model performance. Firstly, we employ prompt learning to embed domain-related information learned from a pretrained vision-language model into prompt embeddings. These embeddings serve as conditional priors, allowing the classification model to adaptively embed semantic-related features and obtain domain-invariant semantic features without excessively aligning domains. Our proposed DKBA approach achieves state-of-the-art results on four MSDA datasets, highlighting its effectiveness in leveraging domain knowledge for improved adaptation performance.}, archive = {J_NEUCOM}, author = {Yuwei He and Juexiao Feng and Guiguang Ding and Yuchen Guo and Tao He}, doi = {10.1016/j.neucom.2024.129114}, journal = {Neurocomputing}, month = {2}, pages = {129114}, shortjournal = {Neurocomputing}, title = {Domain knowledge boosted adaptation: Leveraging vision-language models for multi-source domain adaptation}, volume = {619}, year = {2025}, }
• (2025). UDNet: Unified deep network based on transformer and multi-stage fusion for brain tumor classification from undersampled MRI. NEUCOM, 619, 129109. (www)

  @article{ , abstract = {Magnetic resonance imaging (MRI) exhibits no ionizing radiation and offers high-resolution imaging. The development of MRI technology is currently limited primarily by two factors. On the one hand, prolonged k-space acquisition time often leads to patient discomfort and motion artifacts. On the other hand, in areas such as brain tumor classification, research often relies on fully sampled brain MRI images, necessitating time-consuming manual annotation and frequently overlooking the impact of the reconstruction process on downstream tasks. To address these limitations, we introduce a novel unified deep network (UDNet) based on Transformer and multi-stage fusion for brain tumor classification from undersampled MRI by aggregating reconstruction and classification tasks through a joint learning strategy with shared parameters, aiming to enhance brain tumor classification performance in undersampled MRI scenarios. Furthermore, UDNet employs a hierarchical fusion design combining CNN with Transformer architectures for multiple information fusion. Improvements include enhancing the skip connections in the original U-shaped structure with a designed residual structure and depth-separable convolution, along with utilizing the output of different decoder stages for interactive feature fusion. Experimental results show that the proposed UDNet significantly enhances the brain tumor classification effect in undersampled MRI scenarios.}, archive = {J_NEUCOM}, author = {Zhenyu Huang and Jizhong Duan and Yunshuang Xie and Yu Liu}, doi = {10.1016/j.neucom.2024.129109}, journal = {Neurocomputing}, month = {2}, pages = {129109}, shortjournal = {Neurocomputing}, title = {UDNet: Unified deep network based on transformer and multi-stage fusion for brain tumor classification from undersampled MRI}, volume = {619}, year = {2025}, }
• (2025). Universal embedding for pre-trained models and data bench. NEUCOM, 619, 129107. (www)

  @article{ , abstract = {The transformer architecture has shown significant improvements in the performance of various natural language processing (NLP) tasks. One of the great advantages of transformer-based model is that they allow for the addition of an extra layer to a pre-trained model (PTM) and fine-tuning, rather than requiring the development of a separate architecture for each task. This approach has provided great promising performance in NLP tasks. Therefore, selecting an appropriate PTM from the model zoo, such as Hugging Face , becomes a crucial task. Despite the importance of PTM selection, it still requires further investigation. The main challenge in PTM selection for NLP tasks is the lack of a publicly available benchmark to evaluate model performance for each task and dataset. To address this challenge, we introduce the first public data benchmark to evaluate the performance of popular transformer-based models on diverse ranges of NLP tasks. Furthermore, we propose graph representations of transformer-based models with node features that represent the matrix weight on each layer. Empirical results demonstrate that our proposed graph neural network (GNN) model outperforms existing PTM selection methods.}, archive = {J_NEUCOM}, author = {Namkyeong Cho and Taewon Cho and Jaesun Shin and Eunjoo Jeon and Taehee Lee}, doi = {10.1016/j.neucom.2024.129107}, journal = {Neurocomputing}, month = {2}, pages = {129107}, shortjournal = {Neurocomputing}, title = {Universal embedding for pre-trained models and data bench}, volume = {619}, year = {2025}, }
• (2025). Facial action units guided graph representation learning for multimodal depression detection. NEUCOM, 619, 129106. (www)

  @article{ , abstract = {Depression has become highly prevalent worldwide. Applying technological solutions to compensate for limited resources is crucial for enabling improved early detection and intervention of depressive emotions. Facial expressions and speech are often utilized to assess depression levels since they contain rich emotional cues. However, the inherent inter-individual variability presented in facial expressions and speech introduces noise and redundancy into these cues. Since Facial Action Units (FAUs) represent identity-free facial muscle movements, we proposed a novel FAU-Guided Fusion (FAU-GF) strategy that extracts less noisy and redundant emotional features from human facial and speech behaviors. Moreover, depressive symptoms manifest in a recurrent, episodic, and intermittent fashion, with disjointed temporal connections. To enhance temporal dependency extraction, we innovatively represent temporal events as a graph structure and designed a Feature-Aware Edge Updater (FAEU) to better model contextual dependencies by learning the edges. Subsequently, we adopted Graph Attention Networks (GAT) to update the nodes and edges. After that, a Temporal Attention (TA) module is constructed to selectively focus on informative affect dynamics for each node. Finally, a Depression Predictor is used to assess the depression severity level. The proposed model was evaluated on the AVEC 2014 and AVEC 2019 datasets. Experimental results demonstrate that our model outperforms existing methods.}, archive = {J_NEUCOM}, author = {Changzeng Fu and Fengkui Qian and Yikai Su and Kaifeng Su and Siyang Song and Mingyue Niu and Jiaqi Shi and Zhigang Liu and Chaoran Liu and Carlos Toshinori Ishi and Hiroshi Ishiguro}, doi = {10.1016/j.neucom.2024.129106}, journal = {Neurocomputing}, month = {2}, pages = {129106}, shortjournal = {Neurocomputing}, title = {Facial action units guided graph representation learning for multimodal depression detection}, volume = {619}, year = {2025}, }
• (2025). Graph regularized least squares regression for automated breast ultrasound imaging. NEUCOM, 619, 129065. (www)

  @article{ , abstract = {Breast cancer, recognized as one of the most pervasive malignancies affecting females, manifests a perpetual escalation in its worldwide morbidity. Timely screening offers patients interventions that are not only more expeditious but also more efficacious. ABUS, as an automated breast ultrasound imaging technology, has the advantages of high resolution, reproducible results, and strong objectivity. However, manual interpretation of ABUS images can be highly tedious and laborious when handling huge amounts of data from images, may leading to the risk of missed diagnosis. In practical clinical diagnosis, the incorporation of Computer-Aided Diagnosis (CAD) systems emerges as a more favorable option. In consideration of the holistic factors encompassing model construction efficiency, training cost and overall performance, we provide a unique subspace clustering approach known as Graph regularized Least Squares Regression (GLSR) in this paper. Firstly, we encode the data's intrinsic geometry into the framework of least squares regression to respect local and global correlation structures at the same time. Secondly, we construct a nearest neighbor graph to model the smooth structure of the representation, which may lead to a more accurate representation. Finally, a nonnegative constraint is introduced into our framework so that the representation obtained by GLSR has clearer physical meaning. We provide an optimization approach for solving such a framework based on iterative updates of four blocks. Extensive experimental results indicate the effectiveness of the proposed method compared with state-of-the-art approaches on real-world breast cancer images and other biological data.}, archive = {J_NEUCOM}, author = {Yi Zhou and Menghui Zhang and Ying Pan and Shibin Cai and Aifen Wu and Xi Shu and Mingwang Xu and Xuesong Yin and Guodao Zhang and Huiling Chen and Shuzheng Chen}, doi = {10.1016/j.neucom.2024.129065}, journal = {Neurocomputing}, month = {2}, pages = {129065}, shortjournal = {Neurocomputing}, title = {Graph regularized least squares regression for automated breast ultrasound imaging}, volume = {619}, year = {2025}, }
• (2025). Bridging the gap: An integrated approach to motif discovery and discord detection in time-series data. NEUCOM, 619, 129056. (www)

  @article{ , abstract = {Anomaly detection in time-series data is a critical task with implications for healthcare, law enforcement, and smart policing, yet it presents considerable challenges. Traditional methods often require extensive computational resources and exhibit limited success in discerning various anomaly types, hindering their practical deployment. To address these shortcomings, we propose the Motif Discovery with Discord Removal (MDR) algorithm, which markedly enhances computational efficiency and the accuracy of anomaly detection. The MDR algorithm adopts an integrated approach, combining motif analysis with discord removal, to furnish both a computationally frugal and robust anomaly identification system. To overcome scalability constraints, we introduce a variant of MDR utilizing Particle Swarm Optimization for parallel processing, thereby leveraging distributed computing to efficiently handle voluminous datasets. Our empirical assessments show that the MDR and its parallel adaptation surpass existing methods, supporting agile and precise real-time anomaly detection. Of particular interest to law enforcement, the enhancements in temporal crime data analysis offered by our methods can significantly advance predictive policing capabilities and the discernment of criminal patterns, leading to more proactive and intelligent law enforcement practices.}, archive = {J_NEUCOM}, author = {Wentao Hu}, doi = {10.1016/j.neucom.2024.129056}, journal = {Neurocomputing}, month = {2}, pages = {129056}, shortjournal = {Neurocomputing}, title = {Bridging the gap: An integrated approach to motif discovery and discord detection in time-series data}, volume = {619}, year = {2025}, }
• (2025). Artificial intelligence without restriction surpassing human intelligence with probability one: Theoretical insight into secrets of the brain with AI twins of the brain. NEUCOM, 619, 129053. (www)

  @article{ , abstract = {Artificial Intelligence (AI) has apparently become one of the most important techniques discovered by humans in history while the human brain is widely recognized as one of the most complex systems in the universe. One fundamental critical question which would affect human sustainability remains open: Will artificial intelligence (AI) evolve to surpass human intelligence in the future? This paper shows that in theory new AI twins with fresh cellular level of AI techniques for neuroscience could approximate the brain and its functioning systems (e.g. perception and cognition functions) with any expected small error and AI without restrictions could surpass human intelligence with probability one in the end. This paper indirectly proves the validity of the conjecture made by Frank Rosenblatt 70 years ago about the potential capabilities of AI, especially in the realm of artificial neural networks. This paper also gives the answer to the two widely discussed fundamental questions: 1) whether AI could have potentials of discovering new principles in nature; 2) whether error backpropagation (BP) algorithm commonly and efficiently used in tuning parameters in AI applications is also adopted in the brain. Intelligence is just one of fortuitous but sophisticated creations of the nature which has not been fully discovered. Like mathematics and physics, with no restrictions artificial intelligence would lead to a new subject with its self-contained systems and principles. We anticipate that this paper opens new doors for 1) AI twins and other AI techniques to be used in cellular level of efficient neuroscience dynamic analysis, functioning analysis of the brain and brain illness solutions; 2) new worldwide collaborative scheme for interdisciplinary teams concurrently working on and modelling different types of neurons and synapses and different level of functioning subsystems of the brain with AI techniques; 3) development of low energy of AI techniques with the aid of fundamental neuroscience properties; and 4) new controllable, explainable and safe AI techniques with reasoning capabilities of discovering principles in nature.}, archive = {J_NEUCOM}, author = {Guang-Bin Huang and M. Brandon Westover and Eng-King Tan and Haibo Wang and Dongshun Cui and Wei-Ying Ma and Tiantong Wang and Qi He and Haikun Wei and Ning Wang and Qiyuan Tian and Kwok-Yan Lam and Xin Yao and Tien Yin Wong}, doi = {10.1016/j.neucom.2024.129053}, journal = {Neurocomputing}, month = {2}, pages = {129053}, shortjournal = {Neurocomputing}, title = {Artificial intelligence without restriction surpassing human intelligence with probability one: Theoretical insight into secrets of the brain with AI twins of the brain}, volume = {619}, year = {2025}, }
• (2025). Double negative sampled graph adversarial representation learning with motif-based structural attention network. NEUCOM, 619, 129030. (www)

  @article{ , abstract = {In recent years, graph neural networks have achieved remarkable performance in various downstream tasks by aggregating node neighborhoods hierarchically. However, prior methods usually treat neighboring nodes equally based solely on graph structure or rely on feature similarity for neighborhood aggregation, thereby underutilizing the integrated structural and attribute information. Inspired by the adversarial training mechanism, we propose a Double Negative sampled generative Adversarial network with Motif-based Structural Attention Network ( DNA-MSAN ), a novel framework for graph representation learning. Our method enhances the discriminator’s capability to distinguish between positive and negative samples by preserving the original graph’s motif-based higher-order structure and effectively learning node attribute information. Specifically, we first introduce an attention network based on the higher-order connectivity of motifs, which preserves the higher-order connectivity by perceiving motifs in the graph and using a normalized motif adjacency matrix for neighborhood aggregation. Next, we design a double negative sampling strategy, which cleverly introduces a negative sampling approach based on the current parameters of the discriminator to obtain additional negative samples, serving as useful supplementary samples to those generated by the generator. For the generator, we implement a local graph softmax that significantly optimizes the time complexity of traditional graph softmax by restricting the depth of breadth-first search traversal, leading to marked improvements in model performance. We evaluate DNA-MSAN on link prediction, node classification, and visualization tasks across five datasets, demonstrating substantial advancements in performance across multiple downstream tasks.}, archive = {J_NEUCOM}, author = {Yinglong Zhang and Shangying Yang and Mingyue Kong and Xuewen Xia and Xing Xu}, doi = {10.1016/j.neucom.2024.129030}, journal = {Neurocomputing}, month = {2}, pages = {129030}, shortjournal = {Neurocomputing}, title = {Double negative sampled graph adversarial representation learning with motif-based structural attention network}, volume = {619}, year = {2025}, }
• (2025). Knowledge based attribute completion for heterogeneous graph node classification. NEUCOM, 619, 129023. (www)

  @article{ , abstract = {Heterogeneous graphs, with diverse node and edge types, are prevalent in real-world scenarios. Graph Neural Networks have gained significant attention for processing such complex data structures, delivering impressive results across various tasks like node classification. However, current GNN approaches overlook quality concerns within heterogeneous graphs, like noise or missing attributes, which directly affect the performance of downstream tasks. Previous research has attempted to address these challenges by employing simple one-hot embedding, which is disconnected from the original graph and prone to introducing noise. However, the improvements brought by these methods are limited when the quality of the original graph itself is poor. To this end, we propose a novel attribute completion method based on external knowledge bases, which incorporates a knowledge based completion framework with a relation-aware attention mechanism. The model first employs an attention mechanism to embed the external knowledge subgraph related to the topological structure of the original graph. Then, for nodes requiring attribute completion, the model leverages the association between the external knowledge base and the original graph to complete the missing attributes. Our proposed method can be combined with existing heterogeneous graph neural networks to enhance their performance. Extensive node classification experiments on three real-world datasets underscore the superior performance of our proposed method.}, archive = {J_NEUCOM}, author = {Haibo Yu and Zhangkai Zheng and Yun Xue and Yiping Song and Zhuoming Liang}, doi = {10.1016/j.neucom.2024.129023}, journal = {Neurocomputing}, month = {2}, pages = {129023}, shortjournal = {Neurocomputing}, title = {Knowledge based attribute completion for heterogeneous graph node classification}, volume = {619}, year = {2025}, }
• (2025). ExpDrug: An explainable drug recommendation model based on space feature mapping. NEUCOM, 619, 129021. (www)

  @article{ , abstract = {Drug recommendation uses AI technology to combine a patient’s electronic health records with medical knowledge to help doctors recommend safer and more accurate drug combinations. Time-dependence of patients’ historical records plays a crucial role in existing methods, but some of them ignore or disregard the explainability of the recommendation, and some recommendation results are accompanied by high drug-drug interactions (DDIs). Therefore, an explainable drug recommendation model(ExpDrug) with low DDIs is proposed in this article. It maps the ”black-box” features of patient diagnoses, procedures and medications to the explainable aspect features, and minimizes both rating prediction loss and explainable loss to improve recommendation performance. In addition, a controllable threshold strategy is proposed to reduce the DDI rate. Extensive experiments are conducted on the MIMIC-III dataset, and the proposed ExpDrug achieves new state-of-the-art performance across safety, explainability and four accuracy metrics. The source code is publicly available at https://github.com/hyh0606/ExpDrug .}, archive = {J_NEUCOM}, author = {Xuan Lu and Yanhong Hao and Furong Peng and Zheqing Zhu and Zhanwen Cheng}, doi = {10.1016/j.neucom.2024.129021}, journal = {Neurocomputing}, month = {2}, pages = {129021}, shortjournal = {Neurocomputing}, title = {ExpDrug: An explainable drug recommendation model based on space feature mapping}, volume = {619}, year = {2025}, }
• (2025). Deep neuro-fuzzy system for violence detection. NEUCOM, 619, 129007. (www)

  @article{ , abstract = {With the increasing demand for violence detection in video content, this paper introduces an innovative deep neuro-fuzzy system. This system, which merges Bidirectional Long Short-Term Memory (BiLSTM) networks, MobileNet architecture, and fuzzy logic, offers a robust and swift violence detection solution. Combining BiLSTM’s temporal sequence analysis with MobileNet’s efficient spatial feature extraction, our system accurately identifies violent behavior while minimizing computational resources. Trained on a diverse dataset, the model achieved an impressive accuracy of 96.37% in violence detection before integrating neuro-fuzzy classification techniques. Subsequently, with the incorporation of these techniques, a notable enhancement was observed, with the accuracy further increasing to 97.87%. This improvement underscores the efficacy and potential of our proposed approach in achieving robust and accurate violence detection in real-world scenarios. Moreover, this research explores transfer learning for domain-specific fine-tuning, offering promising adaptability across various contexts such as surveillance and social media. The practical applications of our system are evident, as it can be used to enhance security measures and monitor online content. The model was also applied to surveillance videos and achieved an accuracy of 95.47%. Evaluation metrics validate the model’s effectiveness, underscoring its superiority over existing methods.}, archive = {J_NEUCOM}, author = {Sidharth Mishra and Vishal Jain and Yash Ajay Saraf and Ilanthenral Kandasamy and Vasantha W.B.}, doi = {10.1016/j.neucom.2024.129007}, journal = {Neurocomputing}, month = {2}, pages = {129007}, shortjournal = {Neurocomputing}, title = {Deep neuro-fuzzy system for violence detection}, volume = {619}, year = {2025}, }
• (2025). A multi-view GNN-based network representation learning framework for recommendation systems. NEUCOM, 619, 129001. (www)

  @article{ , abstract = {Graph neural network-based methods have been widely adopted by recommendation systems to provide robust backbones for learning embeddings from multi-view graphs. They have the potential to enhance the performance of recommendation systems by transferring abundant knowledge across different views. However, current works still have deficiencies in capturing the various types of relationships present in multi-view data. They overlook the semantic consistency between fused and view representations and have difficulty in modeling the complementary information between different views. To address these limitations, we propose a novel solution named CrossViRec , a Cross - Vi ew graph neural network-based model to learn low-dimensional node representations for link Rec ommendation. The model leverages two key aspects of multi-view networks: diversity and collaboration. Diversity preserves the diverse semantics of different views, while collaboration facilitates the cooperation between these views. Therefore, the model defines two types of embeddings: a high-dimensional shared embedding and a low-dimensional edge embedding for each view. In particular, a function is learned to generate edge embeddings by aggregating features from node’s neighborhood inside various views through a BiLSTM aggregator followed by a self-attention mechanism. To get the overall vector representation on each view, the self-attention mechanism combines different edge embeddings of a given node and then merges them with the shared embedding vector that encodes the cross-view structural features. Experimental evaluations for the proposed model were conducted on four different challenging datasets: YouTube, Amazon, Twitter, and a private cross-social network dataset. The extensive experiments demonstrate that the proposed model can significantly outperform the existing baseline methods for the link recommendation task, and also achieve better performance for cross-social networks link recommendation task.}, archive = {J_NEUCOM}, author = {Amina Amara and Mohamed Ali Hadj Taieb and Mohamed Ben Aouicha}, doi = {10.1016/j.neucom.2024.129001}, journal = {Neurocomputing}, month = {2}, pages = {129001}, shortjournal = {Neurocomputing}, title = {A multi-view GNN-based network representation learning framework for recommendation systems}, volume = {619}, year = {2025}, }
• (2025). Hyperspectral image restoration via RPCA model based on spectral-spatial correlated total variation regularizer. NEUCOM, 619, 128885. (www)

  @article{ , abstract = {In the realm of hyperspectral image (HSI) restoration, numerous challenges revolve around the separation of low-rank and sparse components. Typical tasks include HSI denoising and anomaly detection. Effectively capturing the inherent priors of an HSI is pivotal for achieving pleasing outcomes, where the low-rankness among spectral dimension and the local smoothness/continuity among the spectral/spatial dimension are the most two significant ones. However, prevailing methods often struggle to accurately capture the above two priors and often lack theoretical guarantees. To tackle these challenges, we propose a novel regularizer called spectral–spatial correlated total variation (SSCTV), which can embed both low-rankness and joint local smoothness into one single regularizer. This integration enables the new regularizer to offer improved prior characterization. Furthermore, by integrating SSCTV regularizer with the Robust Principal Component Analysis (RPCA) framework, we create the new variant RPCA model named SSCTV-RPCA model. Leveraging the incoherence condition and golf scheme techniques, we establish the exact recoverable guarantee of SSCTV-RPCA model. Besides, relying on the Fast Fourier Transform (FFT), we design a computationally efficient solution method that guarantees convergence. Thorough experiments substantiate the efficacy of both the model and its theoretical foundation. The code is available at https://github.com/self-controler/SSCTV-RPCA.git .}, archive = {J_NEUCOM}, author = {Junheng Gao and Hailin Wang and Jiangjun Peng}, doi = {10.1016/j.neucom.2024.128885}, journal = {Neurocomputing}, month = {2}, pages = {128885}, shortjournal = {Neurocomputing}, title = {Hyperspectral image restoration via RPCA model based on spectral-spatial correlated total variation regularizer}, volume = {619}, year = {2025}, }
• (2025). Corrigendum to “Emergent population activity in metric-free and metric networks of neurons with stochastic spontaneous spikes and dynamic synapses” [Neurocomputing 461 (2021) 727–742]. NEUCOM, 618, 129184. (www)

  @article{ , archive = {J_NEUCOM}, author = {Dmitrii Zendrikov and Alexander Paraskevov}, doi = {10.1016/j.neucom.2024.129184}, journal = {Neurocomputing}, month = {2}, pages = {129184}, shortjournal = {Neurocomputing}, title = {Corrigendum to “Emergent population activity in metric-free and metric networks of neurons with stochastic spontaneous spikes and dynamic synapses” [Neurocomputing 461 (2021) 727–742]}, volume = {618}, year = {2025}, }
• (2025). PNC-based trend feature extraction method for time series data. NEUCOM, 618, 129174. (www)

  @article{ , abstract = {With the rapid development of manufacturing industry, air pollution has become a global hot issue, and the accurate prediction of PM2.5 has been an important topic, which can provide valuable information for governmental decision-making in environmental management affairs, among which the trend feature information is quite important. In this paper, in order to solve the problem of insufficient trend feature extraction for time series data, a trend feature extraction method based on PNC (Positive-Negative Correlation) for time series data is proposed. The method focuses on the trend feature extraction of time series data, and skillfully adopts the positive-negative constraints and positive-inverse ratio constraints of PNC to extract trend features of time series data. Firstly, the missing values of the features are filled in the time series data by using the backward and forward sequence pattern, and secondly, the regression algorithm is applied to optimally fill in the missing values of the target columns; then the EMD algorithm is used to extract more features of the data with different frequencies as the new features; then the original unsupervised learning data is transformed into the supervised learning data; and finally, the PNC method is used to extract the trend features of the time series data. In this paper, four deep learning models are used in the experiments to train the prediction of PM2.5 experimental data in four cities. The experimental results show that the RMSE, MSE, MAE, and R2-Score metrics of the model set (Model-Ours) using the PNC-based trend feature extraction method for time-series data are better than those of the model set without trend feature extraction (Model-Base) in the four datasets; among which the MSE metrics are the most obvious, and Model-Ours are better than those of the model set with no trend feature extraction in each of the four datasets, with prediction steps of 48,96,192,384,768 are lower than Model-Base by 67.882, 29.152, 19.559, 9.57 on average.This proves that the trend feature extraction method of time series data based on PNC can improve the prediction accuracy of the model and reduce the model error, and it also proves the effectiveness and superiority of the method.}, archive = {J_NEUCOM}, author = {Bo He and Longbing Li and Qingqing Zhang}, doi = {10.1016/j.neucom.2024.129174}, journal = {Neurocomputing}, month = {2}, pages = {129174}, shortjournal = {Neurocomputing}, title = {PNC-based trend feature extraction method for time series data}, volume = {618}, year = {2025}, }
• (2025). Named entity recognition and relationship extraction for biomedical text: A comprehensive survey, recent advancements, and future research directions. NEUCOM, 618, 129171. (www)

  @article{ , abstract = {The rapid growth of biomedical literature has necessitated the development of advanced information extraction techniques to unlock valuable insights from unstructured text. Named Entity Recognition (NER) and Relationship Extraction (RE) are critical tools in extracting entities and their relationships from biomedical literature. This survey paper aims to provide a comprehensive overview of recent NER and RE techniques in the biomedical domain. We present an analysis of the existing literature, categorizing the techniques based on their underlying methodologies, such as machine learning and deep learning approaches. Additionally, we critically evaluate the strengths, limitations, and performance of each technique, highlighting their contributions and shortcomings. The survey also explores recent trends, challenges, and future research directions in NER and RE in biomedical literature. This survey serves as a valuable resource for researchers, practitioners, and industry professionals seeking a comprehensive understanding of the state-of-the-art techniques and advancements in NER and RE for biomedical literature.}, archive = {J_NEUCOM}, author = {Nandita Goyal and Navdeep Singh}, doi = {10.1016/j.neucom.2024.129171}, journal = {Neurocomputing}, month = {2}, pages = {129171}, shortjournal = {Neurocomputing}, title = {Named entity recognition and relationship extraction for biomedical text: A comprehensive survey, recent advancements, and future research directions}, volume = {618}, year = {2025}, }
• (2025). Modulated spike-time dependent plasticity (STDP)-based learning for spiking neural network (SNN): A review. NEUCOM, 618, 129170. (www)

  @article{ , abstract = {Computational models known as Spiking Neural Networks (SNNs) are inspired by the intricate information processing observed in the brain. A key learning principle, spike-timing-dependent plasticity (STDP), governs how the temporal relationship between pre- and post-synaptic spikes influences synaptic weight changes. STDP, a Hebbian learning rule, is widely employed in training algorithms for SNNs, which encode information in the precise timing of spikes. Recently, modified variants of STDP have emerged to enhance the learning and adaptability of SNNs. These variants incorporate additional elements, such as neuromodulators and dendritic processing. This review explores the fundamental principles, experimental findings, and computational models, providing a comprehensive overview of advancements in modulated STDP-based learning for SNNs. It also addresses challenges associated with modified STDP, including computational complexity, parameter optimisation, scalability, and the pursuit of biological plausibility. This review is a valuable resource for researchers and practitioners interested in developing effective and biologically plausible learning algorithms for modulated STDP.}, archive = {J_NEUCOM}, author = {Nazeerah Abdul Rahman and Nooraini Yusoff}, doi = {10.1016/j.neucom.2024.129170}, journal = {Neurocomputing}, month = {2}, pages = {129170}, shortjournal = {Neurocomputing}, title = {Modulated spike-time dependent plasticity (STDP)-based learning for spiking neural network (SNN): A review}, volume = {618}, year = {2025}, }
• (2025). Transformer-based correlation mining network with self-supervised label generation for multimodal sentiment analysis. NEUCOM, 618, 129163. (www)

  @article{ , abstract = {Multimodal Sentiment Analysis (MSA) aims to recognize and understand a speaker's sentiment state by integrating information from natural language, facial expressions, and voice, has gained much attention in recent years. However, modeling multimodal data poses two main challenges: 1) There are potential sentiment correlations between modalities and within contextual contexts, making it difficult to perform deep-level sentiment correlation mining and information fusion; 2) Sentiment information tends to be unevenly distributed across different modalities, posing challenges in fully leveraging information from each modality for collaborative learning. To address the above challenges, we propose CMLG based on correlation mining and label generation. This approach utilizes a Squeeze and Excitation Network (SEN) to recalibrate modality features and employs Transformer-based intra-modal and inter-modal feature extractors to mine the intrinsic connections between different modalities. In addition, we designed a Self-Supervised Label Generation Module (SLGM) that relies on the positive correlation between feature distances and label offsets to generate single-peak labels, and jointly train multi-peak and single-peak tasks to detect sentiment differences. Extensive experiments on three benchmark dataset (MOSI, MOSEI and SIMS) have shown that the above proposed method CMLG achieves excellent results.}, archive = {J_NEUCOM}, author = {Ruiqing Wang and Qimeng Yang and Shengwei Tian and Long Yu and Xiaoyu He and Bo Wang}, doi = {10.1016/j.neucom.2024.129163}, journal = {Neurocomputing}, month = {2}, pages = {129163}, shortjournal = {Neurocomputing}, title = {Transformer-based correlation mining network with self-supervised label generation for multimodal sentiment analysis}, volume = {618}, year = {2025}, }
• (2025). Unsupervised model-guided online transfer learning framework for multiple fault detection of satellite control system. NEUCOM, 618, 129149. (www)

  @article{ , abstract = {Satellite in-orbit fault detection is critical for satellite reliability and safety. Although fault detection has been extensively studied from signal processing to data-driven methods, there has been less discussion on the problem of fault detection when multiple faults coexist. To address this challenge, this paper proposes a novel unsupervised model-guided online transfer learning method. The method utilizes model-guided to reduce the influence of the control algorithm on the control system. Then, a complexity analyzer is designed to select an appropriate prediction model for each satellite component. Finally, an adaptive threshold setting method based on Wasserstein distance optimization and online transfer learning is combined to achieve accurate detection of each fault when multiple faults coexist. Simulation experiments on multiple fault datasets of six components of a satellite control system, including fault detection accuracy experiments, ablation experiments, and comparison experiments, show that the method is more effective when multiple faults coexist compared to other four advanced fault detection methods.}, archive = {J_NEUCOM}, author = {Huaitao Xia and Tao Meng}, doi = {10.1016/j.neucom.2024.129149}, journal = {Neurocomputing}, month = {2}, pages = {129149}, shortjournal = {Neurocomputing}, title = {Unsupervised model-guided online transfer learning framework for multiple fault detection of satellite control system}, volume = {618}, year = {2025}, }
• (2025). Cluster fusion based cross teaching for semi-supervised medical image segmentation. NEUCOM, 618, 129147. (www)

  @article{ , abstract = {Automated medical image segmentation plays a crucial role in Computer-Aided Diagnosis (CAD). However, the reliance on extensive annotated data in fully supervised methods poses practical challenges. We address this issue by introducing a novel consistency regularization method in semi-supervised learning. This approach integrates a cluster assumption-based mask transformer network with a cross-training strategy, aiming to improve segmentation accuracy and generalization. Diverging from traditional pixel-based segmentation, our proposed Cluster Fusion Net formulates semantic segmentation as a soft clustering assignment, effectively allocating easily confused regions in medical images to corresponding cluster centers for more accurate segmentation. Additionally, we enhance the semi-supervised training process by introducing cluster swapping, providing additional constraints and guidance to improve performance. Extensive experiments on three publicly available medical image datasets validate the effectiveness of our method, demonstrating superior performance compared to state-of-the-art methods in medical image segmentation, even with a limited number of labeled samples.}, archive = {J_NEUCOM}, author = {Huaikun Zhang and Xiangyu Lu and Pei Ma and Jizhao Liu and Jing Lian and Yide Ma}, doi = {10.1016/j.neucom.2024.129147}, journal = {Neurocomputing}, month = {2}, pages = {129147}, shortjournal = {Neurocomputing}, title = {Cluster fusion based cross teaching for semi-supervised medical image segmentation}, volume = {618}, year = {2025}, }
• (2025). Causal disentanglement for regulating social influence bias in social recommendation. NEUCOM, 618, 129133. (www)

  @article{ , abstract = {Social recommendation systems face the problem of social influence bias, which can lead to an overemphasis on recommending items that friends have interacted with. Addressing this problem is crucial, and existing methods often rely on techniques such as weight adjustment or leveraging unbiased data to eliminate this bias. However, we argue that not all biases are detrimental, i.e., some items recommended by friends may align with the user’s interests. Blindly eliminating such biases could undermine these positive effects, potentially diminishing recommendation accuracy. In this paper, we propose a C ausal D isentanglement-based framework for R egulating S ocial influence B ias in social recommendation, named CDRSB, to improve recommendation performance. From the perspective of causal inference, we find that the user social network could be regarded as a confounder between the user and item embeddings (treatment) and ratings (outcome). Due to the presence of this social network confounder, two paths exist from user and item embeddings to ratings: a non-causal social influence path and a causal interest path. Building upon this insight, we propose a disentangled encoder that focuses on disentangling user and item embeddings into interest and social influence embeddings. Mutual information-based objectives are designed to enhance the distinctiveness of these disentangled embeddings, eliminating redundant information. Additionally, a regulatory decoder that employs a weight calculation module to dynamically learn the weights of social influence embeddings for effectively regulating social influence bias has been designed. Experimental results on four large-scale real-world datasets Ciao, Epinions, Dianping, and Douban book demonstrate the effectiveness of CDRSB compared to state-of-the-art baselines. We release our code at https://github.com/Lili1013/CDRSB .}, archive = {J_NEUCOM}, author = {Li Wang and Min Xu and Quangui Zhang and Yunxiao Shi and Qiang Wu}, doi = {10.1016/j.neucom.2024.129133}, journal = {Neurocomputing}, month = {2}, pages = {129133}, shortjournal = {Neurocomputing}, title = {Causal disentanglement for regulating social influence bias in social recommendation}, volume = {618}, year = {2025}, }
• (2025). Integrating advanced vision-language models for context recognition in risks assessment. NEUCOM, 618, 129131. (www)

  @article{ , abstract = {This study proposes an open-environment, multi-label human risk classification framework, capable of identifying possible risks to which individuals appearing on input video data are exposed. The framework consists of an ensemble of models covering object detection, action recognition, context understanding and text classification tasks. Each model is evaluated separately in the context of home environments, with the overall framework performing well in each evaluation after fine tuning. The models were evaluated using a combination of several datasets, including Charades, ETRI-Activity3D, and custom video question answering and risk datasets. This study exploits the ability of large language models to interpret semantic visual features combined with textual input in order to understand the context in which the person is placed. The framework’s ability to output multiple risks and its cross-domain capabilities make it a powerful tool that can enhance current risk management systems in a variety of scenarios, such as homes, construction sites and industry.}, archive = {J_NEUCOM}, author = {Javier Rodriguez-Juan and David Ortiz-Perez and Jose Garcia-Rodriguez and David Tomás and Grzegorz J.Nalepa}, doi = {10.1016/j.neucom.2024.129131}, journal = {Neurocomputing}, month = {2}, pages = {129131}, shortjournal = {Neurocomputing}, title = {Integrating advanced vision-language models for context recognition in risks assessment}, volume = {618}, year = {2025}, }
• (2025). Advances in artificial neural networks, machine learning and computational intelligence. NEUCOM, 618, 129130. (www)

  @article{ , archive = {J_NEUCOM}, author = {Luca Oneto ( Guest editors ) and Frank-Michael Schleif and Alessandro Sperduti}, doi = {10.1016/j.neucom.2024.129130}, journal = {Neurocomputing}, month = {2}, pages = {129130}, shortjournal = {Neurocomputing}, title = {Advances in artificial neural networks, machine learning and computational intelligence}, volume = {618}, year = {2025}, }
• (2025). Topology-aware node injection attacks against graph neural networks. NEUCOM, 618, 129128. (www)

  @article{ , abstract = {Graph neural networks (GNNs) are widely applied in real-life scenarios due to their excellent performance in processing graph data. Meanwhile, GNNs are vulnerable to the node injection attack (NIA). The attacker can significantly reduce the effectiveness of GNNs by injecting only a few malicious nodes into the graph. Existing topology construction approaches for NIA are often either random or computationally expensive, which limits the effectiveness of attack strategies and hinders their applicability to large-scale graphs. In this work, we propose a novel NIA method, named Topology-Aware Node Injection Attack (TANIA), which achieves both high attack effectiveness and scalability. TANIA comprises a two-stage topology construction strategy and an adaptive feature optimization module. Specifically, TANIA first selects nodes with weak information-aware ability as the candidate neighbor set to scale down the topology construction search space of the injected nodes. Then, it establishes connections for the injected nodes following the best wrong class consistency strategy to refine the topology for an effective attack with a limited budget. Based on the reconstructed topology, TANIA adaptively optimizes the features of the injected nodes to enhance the attack ability. The experimental results show that TANIA exhibits outstanding attack performance against 14 GNNs compared with eight representative NIA methods while maintaining scalability on large graph scenarios.}, archive = {J_NEUCOM}, author = {Linlin Su and Jinyan Wang and Zeming Gan and De Li}, doi = {10.1016/j.neucom.2024.129128}, journal = {Neurocomputing}, month = {2}, pages = {129128}, shortjournal = {Neurocomputing}, title = {Topology-aware node injection attacks against graph neural networks}, volume = {618}, year = {2025}, }
• (2025). BGFE-net: A boundary-guided feature enhancement network for segmentation of targets with fuzzy boundaries. NEUCOM, 618, 129127. (www)

  @article{ , abstract = {The accurate segmentation of targets with fuzzy boundaries (e.g., polyps and nodules) is critical for the diagnosis and prevention of diseases (e.g., malignancies). Quite a few recent studies have been performed on medical image segmentation with special emphasis set upon the targets’ fuzzy boundaries. However, insufficient fusion of features of the target regions and those of the boundaries in these studies significantly limits their effectiveness in dealing with fuzzy boundaries. To address the issue, we propose a novel Boundary-Guided Feature Enhancement (BGFE) Network (BGFE-Net). Following existing methods, BGFE-Net also formulates image segmentation task as a multi-task problem, with the target segmentation task facilitated by the boundary segmentation task. BGFE-Net is characterized by its BGFE module, an X-shaped structure designed for better fusion of the features of the target regions and those of the boundaries. For the target segmentation branch of the module, the features of the uncertain target pixels are enhanced based on their cross-attention with those of the highly confident target pixels, whose indices are determined by the boundary segmentation branch. For the boundary segmentation branch of the module, the features of the uncertain boundary pixels are enhanced based on their cross-attention with those of the highly confident boundary pixels, whose indices are determined by the target segmentation branch. Furthermore, to place more emphasis on boundary regions, a boundary-based data augmentation strategy is proposed, with boundary maps generated based on Segment Anything Model (SAM) used as one channel for the augmented samples. Extensive experiments on two colon polyp datasets and one thyroid nodule dataset demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Dongyang Xie and Yang Zhang and Xiaoxi Tian and Le Xu and Lianhong Duan and Lixia Tian}, doi = {10.1016/j.neucom.2024.129127}, journal = {Neurocomputing}, month = {2}, pages = {129127}, shortjournal = {Neurocomputing}, title = {BGFE-net: A boundary-guided feature enhancement network for segmentation of targets with fuzzy boundaries}, volume = {618}, year = {2025}, }
• (2025). A review on multi-focus image fusion using deep learning. NEUCOM, 618, 129125. (www)

  @article{ , abstract = {Multi-focus image fusion (MFIF) is an image enhancement technique that investigates how to obtain a fully focused image from multiple defocus images, providing fundamental services for computer vision fields, such as image recognition, 3D reconstruction, medical diagnosis, computational photography, etc. In recent years, the deep learning-based MFIF has broken through the limitations of traditional MFIF and achieved superior fusion results. Existing reviews on MFIF mainly classify and describe the techniques at the technical level, such as supervised/unsupervised learning methods and network types, but lack discussion and analysis on problem scenarios. Therefore, based on the problem scenarios of MFIF, this paper categorizes deep learning-based MFIF research into six types: MFIF with lightweight networks, MFIF for artifacts and defocus spread effects, MFIF for information preservation, MFIF with unified fusion networks, MFIF on addressing suboptimal initial decision map and MFIF in challenging environments. Furthermore, this paper summarizes commonly used synthesis datasets and real datasets for MFIF, and statistically describes the main evaluation metrics. Finally, this paper analyzes the shortcomings of existing algorithms, and identifies potential future research directions, aiming to provide objective reviews for MFIF researchers focusing on different problem scenarios.}, archive = {J_NEUCOM}, author = {Fei Luo and Baojun Zhao and Joel Fuentes and Xueqin Zhang and Weichao Ding and Chunhua Gu and Luis Rojas Pino}, doi = {10.1016/j.neucom.2024.129125}, journal = {Neurocomputing}, month = {2}, pages = {129125}, shortjournal = {Neurocomputing}, title = {A review on multi-focus image fusion using deep learning}, volume = {618}, year = {2025}, }
• (2025). Adaptive opinion dynamics over community networks when agents cannot express opinions freely. NEUCOM, 618, 129123. (www)

  @article{ , abstract = {In this paper, based on the Spiral of Silence (SOS) theory, we propose a novel dynamic weight expressed and private opinion dynamics model (called DWEPO for simplicity) for social networks with community structures. We introduce a dynamic weight to enable the quantification of the indicators of opinion climate, quasi-sensory statistics, and fear of isolation index in SOS theory. When updating private opinions, agents adjust their confidence levels based on the communities to which their neighbors belong. When updating expressed opinions, agents adjust the dynamic weights based on the opinion climate within their communities. We conducted simulations of our model on two real networks and one synthetic network, and compared it with the Inertial Hegselmann–Krause (IHK) model and the original EPO model. The results show that our model is more conducive to opinion convergence and portrays the process of agent “silencing” in a group well. The DWEPO model can realize the consensus of agents in the community through confidence level settings. When an agent becomes “silent,” its opinion climate, quasi-sensory statistics and fear of isolation index decrease accordingly. When an agent’s opinion dominates the community, the opinion climate and fear of isolation index change from positive to negative.}, archive = {J_NEUCOM}, author = {Yuan Peng and Yiyi Zhao and Jianglin Dong and Jiangping Hu}, doi = {10.1016/j.neucom.2024.129123}, journal = {Neurocomputing}, month = {2}, pages = {129123}, shortjournal = {Neurocomputing}, title = {Adaptive opinion dynamics over community networks when agents cannot express opinions freely}, volume = {618}, year = {2025}, }
• (2025). ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. NEUCOM, 618, 129122. (www)

  @article{ , abstract = {Zero-Shot Anomaly Segmentation (ZSAS) aims to segment anomalies without any training data related to the test samples. Recently, while foundational models like CLIP and SAM have shown potential for ZSAS, existing approaches leveraging either CLIP or SAM individually encounter critical limitations: (1) CLIP emphasizes global feature alignment across different inputs, leading to imprecise segmentation of local anomalous parts; and (2) SAM tends to generate numerous redundant masks without proper prompt constraints, resulting in complex post-processing requirements. In this paper, we introduce ClipSAM, a novel collaborative framework that integrates CLIP and SAM to address these issues in ZSAS. The insight behind ClipSAM is to employ CLIP’s semantic understanding capability for anomaly localization and rough segmentation, which is further used as the prompt constraints for SAM to refine the anomaly segmentation results. Specifically, we propose a Unified Multi-scale Cross-modal Interaction (UMCI) module that learns local and global semantics about anomalous parts by interacting language features with visual features at both row-column and multi-scale levels, effectively reasoning about anomaly positions. Additionally, we develop a Multi-level Mask Refinement (MMR) module that guides SAM’s output through multi-level spatial prompts derived from CLIP’s localization, progressively merging the masks to refine the segmentation results. Extensive experiments validate the effectiveness of our approach, achieving the optimal segmentation performance on the MVTec-AD and VisA datasets.}, archive = {J_NEUCOM}, author = {Shengze Li and Jianjian Cao and Peng Ye and Yuhan Ding and Chongjun Tu and Tao Chen}, doi = {10.1016/j.neucom.2024.129122}, journal = {Neurocomputing}, month = {2}, pages = {129122}, shortjournal = {Neurocomputing}, title = {ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation}, volume = {618}, year = {2025}, }
• (2025). Two grids are better than one: Hybrid indoor scene reconstruction framework with adaptive priors. NEUCOM, 618, 129118. (www)

  @article{ , abstract = {Indoor scene reconstruction from multi-view images is a pivotal technology within the field of robotics and augmented reality. Previous researches have predominantly focused on neural radiance fields aided by geometric monocular priors. However, due to the inductive smoothness bias introduced by deep Multi-Layer Perceptron (MLP) networks, these methods struggle to recover the scene surface with complex and fine geometry details. Additionally, when used as additional supervision signals during optimization, priors in different regions make different contributions. Simply incorporating them in all regions may lead to a decrease in the accuracy. To tackle these issues, we present a generic end-to-end framework named AdaptSurf, which combines Signed Distance Field (SDF) voxel grids and feature voxel grids to enhance the capability of reconstructing accurate geometry details, respectively. Furthermore, we design a policy network to adaptively enable the estimated depth or normal priors to supervise the learning process, which improves the reconstruction accuracy and accelerates neural surface reconstruction. Qualitative and quantitative experiments show that AdaptSurf yields high-quality surfaces, especially for fine-grained details and smooth regions. Furthermore, the policy network exhibits an interpretable behavior that depends on the voxel features, which helps to improve the quality of surface reconstruction.}, archive = {J_NEUCOM}, author = {Boyuan Bai and Xiuquan Qiao and Ping Lu and Hongru Zhao and Wenzhe Shi and Hengjia Zhang and Yakun Huang}, doi = {10.1016/j.neucom.2024.129118}, journal = {Neurocomputing}, month = {2}, pages = {129118}, shortjournal = {Neurocomputing}, title = {Two grids are better than one: Hybrid indoor scene reconstruction framework with adaptive priors}, volume = {618}, year = {2025}, }
• (2025). Online dynamic influence maximization based on deep reinforcement learning. NEUCOM, 618, 129117. (www)

  @article{ , abstract = {Online influence maximization (OIM) aims to maximize the scale of information diffusion in social networks with unknown edge activation probabilities, which has gained significant attention due to its potential to optimize and control the spread of information. However, the existing related studies only focused on static networks with fixed topologies, failing to account for the dynamic nature of real-world social networks. Real-world social networks are dynamic, with constantly evolving network topologies, the continuous appearance and disappearance of user nodes, and fluctuating relationships between nodes. To address these challenges, we propose a novel method for influence maximization in online dynamic social networks, leveraging deep reinforcement learning to enhance the OIM process. Our approach utilizes node feedback to learn the relationships between network structures and edge activation probabilities, enabling the dynamic selection of seed sets that maximize influence based on the current state of the network. We validate the effectiveness and generalizability of our proposed method through experiments on multiple real-world datasets.}, archive = {J_NEUCOM}, author = {Nuan Song and Wei Sheng and Yanhao Sun and Tianwei Lin and Zeyu Wang and Zhanxue Xu and Fei Yang and Yatao Zhang and Dong Li}, doi = {10.1016/j.neucom.2024.129117}, journal = {Neurocomputing}, month = {2}, pages = {129117}, shortjournal = {Neurocomputing}, title = {Online dynamic influence maximization based on deep reinforcement learning}, volume = {618}, year = {2025}, }
• (2025). Jensen–Shannon divergence based novel loss functions for bayesian neural networks. NEUCOM, 618, 129115. (www)

  @article{ , abstract = {Bayesian neural networks (BNNs) are state-of-the-art machine learning methods that can naturally regularize and systematically quantify uncertainties using their stochastic parameters. Kullback–Leibler (KL) divergence-based variational inference used in BNNs suffer from unstable optimization and challenges in approximating light-tailed posteriors due to the unbounded nature of the KL divergence. To resolve these issues, we formulate a novel loss function for BNNs based on a new modification to the generalized Jensen–Shannon (JS) divergence, which is bounded. In addition, we propose a Geometric JS divergence-based loss, which is computationally efficient since it can be evaluated analytically. We found that the JS divergence-based variational inference is intractable, and hence employed a constrained optimization framework to formulate these losses. Our theoretical analysis and empirical experiments on multiple regression and classification data sets suggest that the proposed losses perform better than the KL divergence-based loss, especially when the data sets are noisy or biased. Specifically, there are approximately 5% and 8% improvements in accuracy for a noise-added CIFAR-10 dataset and a regression dataset, respectively. There is about 13% reduction in false negative predictions of a biased histopathology dataset. In addition, we quantify and compare the uncertainty metrics for the regression and classification tasks.}, archive = {J_NEUCOM}, author = {Ponkrshnan Thiagarajan and Susanta Ghosh}, doi = {10.1016/j.neucom.2024.129115}, journal = {Neurocomputing}, month = {2}, pages = {129115}, shortjournal = {Neurocomputing}, title = {Jensen–Shannon divergence based novel loss functions for bayesian neural networks}, volume = {618}, year = {2025}, }
• (2025). Boosting few-shot intent detection via feature enrichment and regularization. NEUCOM, 618, 129113. (www)

  @article{ , abstract = {Few-shot intent detection aims to detect fast-emerging new intents from limited labeled utterances, which attracts increasing attention recently. Although current few-shot learning methods have achieved promising results, their performance is still limited by improper features and poor generalization, especially when the number of classes is large, such as when both seen and unseen classes are present simultaneously. The main issues are that a handful of training utterances fail to represent the whole diversely expressed class and the trained model tends to overfit seen samples. To circumvent these issues, we propose a novel method based on feature enrichment and regularization to improve the utterance features for classification. Specifically, we design a meta fusion module to enhance utterance features by fully exploiting relevant augmented data from large-scale language generation models via the natural language inference paradigm to obtain comprehensive class prototypes. And, taking into account the overfitting problem, we design a pairwise loss and a weighted contrastive loss to impose further relational constraints in the embedding space to increase the generalization ability of models. Extensive experiments on both standard and generalized few-shot intent detection tasks demonstrate that our method outperforms previous strong baselines on four public intent datasets.}, archive = {J_NEUCOM}, author = {Feng Zhang and Wei Chen and Pei Zhao and Tengjiao Wang}, doi = {10.1016/j.neucom.2024.129113}, journal = {Neurocomputing}, month = {2}, pages = {129113}, shortjournal = {Neurocomputing}, title = {Boosting few-shot intent detection via feature enrichment and regularization}, volume = {618}, year = {2025}, }
• (2025). IE-NeRF: Exploring transient mask inpainting to enhance neural radiance fields in the wild. NEUCOM, 618, 129112. (www)

  @article{ , abstract = {We present a novel approach for synthesizing realistic novel views using Neural Radiance Fields (NeRF) with uncontrolled photos. While NeRF has shown impressive results in controlled settings, it struggles with transient objects commonly found in dynamic and time-varying scenes. Our framework called Inpainting Enhanced NeRF , or IE-NeRF addresses this issue by learning to decompose the static and dynamic components of the scene through exploring transient mask inpainting. Specifically, to model varying appearances, we propose a CNN-based appearance encoder, and a view-consistent appearance loss to transfer consistent photometric appearance in different views. Moreover, our approach extends the Multi-Layer Perceptrons (MLP) network of NeRF, enabling it to simultaneously generate intrinsic properties (static color, density) and extrinsic transient masks. During optimization we take advantage of inpainting to restore static scene image while eliminating occlusions with the guidance of transient masks. Additionally, we propose a new training strategy using frequency regularization with transient mask factor in integrated positional encoding, which facilitates faster inference and early separation of transient components during training. We evaluate our approach on internet photo collections of landmarks, demonstrating its ability to generate high-quality novel views and outperform the existing methods.}, archive = {J_NEUCOM}, author = {Shuaixian Wang and Haoran Xu and Yaokun Li and Jiwei Chen and Guang Tan}, doi = {10.1016/j.neucom.2024.129112}, journal = {Neurocomputing}, month = {2}, pages = {129112}, shortjournal = {Neurocomputing}, title = {IE-NeRF: Exploring transient mask inpainting to enhance neural radiance fields in the wild}, volume = {618}, year = {2025}, }
• (2025). Hybrid classical–quantum neural networks enhanced by quantum architecture search for coronary artery stenosis detection. NEUCOM, 618, 129111. (www)

  @article{ , abstract = {The quantum circuit is a pivotal component of the hybrid classical–quantum neural network, and its efficacy influences the network’s overall performance. Given the vast number of potential combinations of quantum gates, designing simple yet effective quantum circuits has become a significant challenge. To improve the performance of hybrid networks through the design of suitable quantum circuits, we propose a training framework that utilizes quantum architecture search to enhance hybrid classical–quantum neural networks. This framework employs a Monte Carlo tree search algorithm to optimize quantum circuits and integrates multiple neural network architectures to improve the performance of hybrid networks. Furthermore, this work considers the quantum circuit architecture of two-qubit quantum circuits as the search target, constructing distributed quantum circuits that significantly reduce the search space. Experiments conducted on a coronary artery stenosis detection dataset validate the proposed scheme and analyze the impact of varying search spaces on circuit design. The results demonstrate that, compared to several prevalent quantum circuit architectures, the quantum circuit designed using this search strategy exhibits superior performance across multiple evaluation metrics.}, archive = {J_NEUCOM}, author = {Shaochun Li and Junzhi Cui and Jingli Ren}, doi = {10.1016/j.neucom.2024.129111}, journal = {Neurocomputing}, month = {2}, pages = {129111}, shortjournal = {Neurocomputing}, title = {Hybrid classical–quantum neural networks enhanced by quantum architecture search for coronary artery stenosis detection}, volume = {618}, year = {2025}, }
• (2025). Enhancing zero-shot multilingual semantic parsing: A framework leveraging large language models for data augmentation and advanced prompting techniques. NEUCOM, 618, 129108. (www)

  @article{ , abstract = {In recent years, significant progress has been made in semantic parsing tasks due to the introduction of pre-trained language models. However, there remains a notable gap between English and other languages because of the limited availability of annotated data. One promising approach to bridge this gap is to augment multilingual datasets using labeled English data and then train semantic parsers with this enhanced dataset (known as zero-shot multilingual semantic parsing). In this study, we propose a novel framework for the zero-shot multilingual semantic parsing task through LLM-driven data augmentation. Our approach leverages annotated English data – consisting of sentences and their corresponding semantic representations – to generate augmented data in target languages. This is achieved through the implementation of multilingual chain-of-thought (CoT) prompting techniques that incrementally construct semantic forms. By deconstructing complex semantic structures into sub-fragments, our framework guides LLMs to progressively construct accurate semantic forms in target languages without the need for target language demonstration examples, enabling zero-shot learning. We demonstrate the effectiveness of our framework on two benchmark datasets, MTOP and MASSIVE, achieving state-of-the-art (SOTA) performance in zero-shot multilingual semantic parsing. Additionally, we analyze the effectiveness of the proposed framework under stricter constraints, where only limited annotated data in English is available. Our experiments indicate that our framework, trained using 600 samples of English data ( 5% of the full dataset), can achieve 80% of the performance of a system trained on the full dataset. This demonstrates the effectiveness of our proposed framework under conditions of limited annotated data.}, archive = {J_NEUCOM}, author = {Dinh-Truong Do and Minh-Phuong Nguyen and Le-Minh Nguyen}, doi = {10.1016/j.neucom.2024.129108}, journal = {Neurocomputing}, month = {2}, pages = {129108}, shortjournal = {Neurocomputing}, title = {Enhancing zero-shot multilingual semantic parsing: A framework leveraging large language models for data augmentation and advanced prompting techniques}, volume = {618}, year = {2025}, }
• (2025). Predicting pedestrian crossing intention in autonomous vehicles: A review. NEUCOM, 618, 129105. (www)

  @article{ , abstract = {Road traffic accidents involving collisions between vehicles and pedestrians are a major cause of death and injury globally. With recent technological progress in the field of autonomous driving, and as more and more vehicles with autonomous capabilities are introduced on the roads, the prediction of pedestrian crossing intention has emerged as an active area of research. The pedestrian crossing intention prediction task can be summarized as follows: are detected pedestrians going to cross the road or not? Answering this question is crucial so that control systems integrated into autonomous vehicles can avoid collision with pedestrians by executing actions like braking or steering away. Moreover, determining crossing intention long in advance is important so that the autonomous vehicle can offer a comfortable driving experience without unnecessary braking. The number of reviews published in the pedestrian behavior prediction literature is limited, and those that present approaches for crossing intention prediction do not provide a comprehensive view of the field. In this paper, we propose to fill this gap by presenting an extensive review of the various approaches proposed in the literature for predicting pedestrian crossing intention. We begin by formulating the problem at hand, followed by a description of datasets and an overview of pedestrian behavior understanding. We then provide a taxonomy of models as well as an analysis and comparison of the approaches proposed in the literature. We finalize our review with a discussion of current challenges in the field and provide possible directions for future research.}, archive = {J_NEUCOM}, author = {François-Guillaume Landry and Moulay A. Akhloufi}, doi = {10.1016/j.neucom.2024.129105}, journal = {Neurocomputing}, month = {2}, pages = {129105}, shortjournal = {Neurocomputing}, title = {Predicting pedestrian crossing intention in autonomous vehicles: A review}, volume = {618}, year = {2025}, }
• (2025). Comparative analysis of unsupervised anomaly detection techniques for heat detection in dairy cattle. NEUCOM, 618, 129088. (www)

  @article{ , abstract = {Population growth has increased the demand for meat and dairy products, making livestock, especially cattle, key to meeting this demand. This has led to an increase in herd size, complicating efficient herd management. To meet this challenge, innovative technologies, such as monitoring collars, have been developed to improve individual animal management. This research work evaluates and compares three unsupervised anomaly detection methods to identify estrus in dairy cows from intensive farms, based on daily activity data recorded by a commercial monitoring collar. Data from two different dairy farms have been used and the results have been compared by evaluating the behavior both individually and at herd level. The results obtained show a good performance of the selected techniques in the individual animal models. Thus, this research demonstrates that these techniques can be very useful tools in farm management, providing valuable information, improving productivity and, consequently, increasing the economic performance of the farm.}, archive = {J_NEUCOM}, author = {Álvaro Michelena and Antonio Díaz-Longueira and Paulo Novais and Dragan Simić and Óscar Fontenla-Romero and José Luis Calvo-Rolle}, doi = {10.1016/j.neucom.2024.129088}, journal = {Neurocomputing}, month = {2}, pages = {129088}, shortjournal = {Neurocomputing}, title = {Comparative analysis of unsupervised anomaly detection techniques for heat detection in dairy cattle}, volume = {618}, year = {2025}, }
• (2025). Information-based stealthy attacks against distributed state estimation over sensor networks. NEUCOM, 618, 129087. (www)

  @article{ , abstract = {In this article, the problem of information-based attacks on distributed state estimation over sensor networks is investigated. We consider a scenario where each sensor node obtains the state estimation by utilizing the measurements transmitted from local and neighboring nodes. An information-based attack is proposed to degrade the estimation performance of the remote estimator, which can be designed in two steps: First, the attacker runs a Kalman filter to estimate the compromised prediction error of remote estimator, then tampers with the transmitted measurements based on the obtained estimation. The attacker aims to maximize the estimation error of the estimator while remaining stealthy to the detector, which can be formulated as an optimization problem. The analytical expression of the optimal strict stealthy attack strategy is derived using the singular value decomposition technique. Moreover, the optimal ϵ -stealthy attack strategy can be obtained by solving a convex optimization problem. When attackers have limited access to all measurements due to resource constraints, extended discussions are provided on designing the optimal attack strategy with strict stealthiness and offering a selection method to further enhance the attack performance. Simulation results are presented to demonstrate the effectiveness of the designed attacks.}, archive = {J_NEUCOM}, author = {Yichao Wang and Kaijing Jin and Dan Ye}, doi = {10.1016/j.neucom.2024.129087}, journal = {Neurocomputing}, month = {2}, pages = {129087}, shortjournal = {Neurocomputing}, title = {Information-based stealthy attacks against distributed state estimation over sensor networks}, volume = {618}, year = {2025}, }
• (2025). Lagrange stability of quaternion-valued neural networks with mixed delays on time scales. NEUCOM, 618, 129086. (www)

  @article{ , abstract = {This study delves into the Lagrange stability of quaternion-valued neural networks (QVNNs) on time scales, the considered model incorporates the leaky delay, discrete delay and distributed delay. By constructing appropriate Lyapunov–Krasovskii functionals on time scales and applying the free-weighted matrix approach along with inequality techniques, a sufficient condition formulated as linear matrix inequalities (LMIs) is obtained, which ensures the Lagrange stability and the attractor regions of networks is estimated. Finally, a numerical example demonstrates the correctness and effectiveness of the method.}, archive = {J_NEUCOM}, author = {Jiang Wu and Qiankun Song and Yurong Liu}, doi = {10.1016/j.neucom.2024.129086}, journal = {Neurocomputing}, month = {2}, pages = {129086}, shortjournal = {Neurocomputing}, title = {Lagrange stability of quaternion-valued neural networks with mixed delays on time scales}, volume = {618}, year = {2025}, }
• (2025). Joint contrastive learning with semantic enhanced label referents for few-shot NER. NEUCOM, 618, 129081. (www)

  @article{ , abstract = {Few-shot named entity recognition (NER) faces a challenge of obtaining more generalized class representations with limited labeled instances to recognize named entities. However, few support instances are insufficient to encompass the intra-class diverse entity semantics, and learning the similarity between query instances and support instances in isolation predisposes dispersed class distributions and fuzzy boundaries of certain classes, leading to misclassification. Meanwhile, there are differences in the meanings of the same class labels under different target domains, which leads to low accuracy of label referents generated based on label words, affecting the similarity metric between query instances and label referents. To overcome the problem, we propose a new method, J oint C ontrastive learning with semantic E nhanced L abel R eferents for few-shot NER ( JCELRNER ). Specifically, we propose a strategy for representing the class using multiple referents, including support instance tokens and label prototypes. A joint contrastive loss objective that simultaneously is proposed to optimize the distributional distance between query instances and all class referents in the same semantic space by jointly considering token-to-token and token-to-label correlations, promoting the intra-class compactness. Meanwhile, JCELRNER proposes a dual semantic enhancement module to generate reliable label referents that adapt to different target domains by fusing entity and context semantic information. Extensive experiments show that by promoting the intra-class relevance of referents of the same class and enhancing the label referents with dual semantic information, the model achieves a comprehensive and accurate understanding of the entire class, thereby enhancing the generalization ability to novel classes, outperforming the other competitive approach.}, archive = {J_NEUCOM}, author = {Xiaoya Liu and Senlin Luo and Zhouting Wu and Limin Pan and Xinshuai Li}, doi = {10.1016/j.neucom.2024.129081}, journal = {Neurocomputing}, month = {2}, pages = {129081}, shortjournal = {Neurocomputing}, title = {Joint contrastive learning with semantic enhanced label referents for few-shot NER}, volume = {618}, year = {2025}, }
• (2025). Integrating real-time pose estimation and PPE detection with cutting-edge deep learning for enhanced safety and rescue operations in the mining industry. NEUCOM, 618, 129080. (www)

  @article{ , abstract = {The dynamic and often hazardous environments of mines necessitate advanced safety protocols, particularly in PPE compliance and emergency response. This study leverages cutting-edge pose estimation and object detection technologies to enhance miner safety by focusing on two primary goals: developing a comprehensive dataset for body situation classification from miner pose data to improve rescue operations, and refining PPE detection using deep learning models localized by pose keypoints. A robust dataset was collected in diverse mining scenarios under low-light and noisy conditions using the adapted YOLO Pose v8. This dataset was instrumental in extracting detailed metrics such as joint angles and bounding box sizes for body situation classification. For PPE detection, the study employed various YOLO models, including YOLOv8, YOLOv9, YOLOv10, YOLOv11, YOLO World, and RTDetr, across a public dataset and a custom dataset from five Moroccan mines to ensure effective PPE identification in different contexts. The implemented models showed superior performance in PPE detection and pose estimation, confirming the efficacy of integrating advanced YOLO architectures in mining operations. By enhancing detection and analyzing miner postures and PPE compliance, this research significantly improves mine safety, potentially reducing accident rates and improving rescue operation effectiveness. The methodologies and insights extend to occupational safety in other high-risk industries, setting new standards for safety technology integration and positioning these solutions at the forefront of innovation.}, archive = {J_NEUCOM}, author = {Mohamed Imam and Karim Baïna and Youness Tabii and El Mostafa Ressami and Youssef Adlaoui and Soukaina Boufousse and Intissar Benzakour and El Hassan Abdelwahed}, doi = {10.1016/j.neucom.2024.129080}, journal = {Neurocomputing}, month = {2}, pages = {129080}, shortjournal = {Neurocomputing}, title = {Integrating real-time pose estimation and PPE detection with cutting-edge deep learning for enhanced safety and rescue operations in the mining industry}, volume = {618}, year = {2025}, }
• (2025). Using offline data to speed up reinforcement learning in procedurally generated environments. NEUCOM, 618, 129079. (www)

  @article{ , abstract = {One of the key challenges of Reinforcement Learning (RL) is the ability of an agent to generalize its learned policy to unseen settings. Moreover, training an RL agent requires large numbers of interactions with the environment. Motivated by the success of Imitation Learning (IL), we conduct a study to investigate whether an agent can leverage offline data in the form of trajectories to improve the sample-efficiency in procedurally generated environments. We consider two settings of using IL from offline data for RL: (1) pre-training a policy before online RL training and (2) concurrently training a policy with online RL and IL from offline data. We analyze the impact of the quality (optimality of trajectories), quantity and diversity of available offline trajectories on the effectiveness of both approaches. Across four well-known sparse reward tasks in the MiniGrid environment, we find that using IL for both pre-training and concurrently during online RL training, consistently improves sample-efficiency, and in some tasks achieves higher returns compared to using either IL or RL alone. Furthermore, we show that training a policy from as few as two trajectories can make the difference between learning an optimal policy at the end of online training and not learning at all. Evaluation in two tasks of the Procgen environment further highlights that the diversity of the training data is more important than its quality. Our findings motivate the widespread adoption of IL for pre-training and concurrent IL in procedurally generated environments whenever offline trajectories are available or can be generated.}, archive = {J_NEUCOM}, author = {Alain Andres and Lukas Schäfer and Stefano V. Albrecht and Javier Del Ser}, doi = {10.1016/j.neucom.2024.129079}, journal = {Neurocomputing}, month = {2}, pages = {129079}, shortjournal = {Neurocomputing}, title = {Using offline data to speed up reinforcement learning in procedurally generated environments}, volume = {618}, year = {2025}, }
• (2025). A novel OL-mapping operator-based edge detection approach. NEUCOM, 618, 129078. (www)

  @article{ , abstract = {Accurate edge detection is of great significance to image processing. However, the traditional methods, which use differential operations or prior experience to determine the value of the operator, cannot meet the high demand for accuracy in the field of image processing. Deep-learning based detection algorithms restrict its flexible usability due to the high demand of computational resources (storage space, processor) for its large number of parameters. To date, neither traditional nor emerging techniques cannot achieve SOTA performance on detection tasks. Therefore, to address the need for both precision and resources, this paper proposes an edge detection operator (HQ operator) based on the relationship between the original image and label mapping, which enhances the edge detection effect while drastically reducing the computational complexity. Unlike previous edge extraction operators, HQ operator utilizes a combination of labels and original images for edge detection, and does not use iterative training, but through subgraph generation classification (SGC) method and subgraph feature fusion operation (SFFO) to obtain the edge detection operator at once. Compared to the traditional edge detection operator, HQ operator can accurately extract feature information of the edge of target region. And compared to deep learning, the number of parameters of the HQ operator is greatly reduced and the calculation time of parameters is reduced. In addition, the HQ operator obtained on one dataset can be ported to other datasets with good edge extraction results. The edge images processed by the HQ operator can be spliced with the original image in the form of channels, and the image after splicing is fed into the image segmentation network to improve the performance of the segmentation network to improve the performance of the segmentation network. The experimental results show that by using the HQ operator, the image edge detection effect is significantly improved (PA:96.49%), and the performance of the segmentation network is also improved to a certain extent when this operator is combined with the segmentation network(Dice:89.97%, IoU:81.97%).}, archive = {J_NEUCOM}, author = {Qi Han and Sihan Li and Mingyang Hou and Hongyi Wang and Tengfei Weng and Yuan Tian and Jiajia Qin and Yecheng Xu and Yi Wei}, doi = {10.1016/j.neucom.2024.129078}, journal = {Neurocomputing}, month = {2}, pages = {129078}, shortjournal = {Neurocomputing}, title = {A novel OL-mapping operator-based edge detection approach}, volume = {618}, year = {2025}, }
• (2025). MLFINet: A multi-level feature interaction 3D medical image segmentation network. NEUCOM, 618, 129077. (www)

  @article{ , abstract = {Accurate identification of brain tumors using magnetic resonance imaging (MRI) is crucial for clinical decision-making and surgery planning. However, it is a challenging task due to the diverse and complex nature of tumors. Some existing U-Net variants are outperformed in the brain tumor segmentation tasks, but these models usually have large parameters and limited ability to extract image features. In addition, the semantic difference between encoder and decoder also affects the segmentation performance of U-Net based variant network models. To address the above problems, a new segmentation network model called MLFINet is proposed based on a 3D U-Net structure. In this model, a dual separable convolution squeeze-and-excitation (DSSE) module is first designed to extract image features more efficiently while reducing the number of model parameters, and then a multi-level feature interaction module (MFIM) is designed to enhance semantic interaction between different levels. In addition, structural similarity (SSIM) loss is considered to help the model better identify and segment complex boundaries and details. Extensive experiments are conducted on open-source datasets, and the results demonstrate that the proposed model outperforms selected baselines in terms of several metrics, achieving the dice score of 86.63%, 85.79%, 94.10% on the BraTS2018, BraTS2020 and LiTS17 dataset, respectively.}, archive = {J_NEUCOM}, author = {Chuanlin Liao and Xiaolin Gou and Kemal Polat and Jingchun Zhou and Yi Lin}, doi = {10.1016/j.neucom.2024.129077}, journal = {Neurocomputing}, month = {2}, pages = {129077}, shortjournal = {Neurocomputing}, title = {MLFINet: A multi-level feature interaction 3D medical image segmentation network}, volume = {618}, year = {2025}, }
• (2025). Leveraging central-surrounding receptive fields for single image dehazing. NEUCOM, 618, 129075. (www)

  @article{ , abstract = {Image dehazing is a representative low-level vision task that estimates latent haze-free images from hazy images. In recent years, Vision Transformers (ViT) have shown promising dehazing performance, leveraging their capacity for global perception through long-sequence dependencies in cost of high resource consumption. Therefore, our approach seeks to integrate the utilization of global information into the Convolutional Neural Network (CNN) framework in a more resource-efficient manner. In this paper, we introduce the Adaptive Center-Surround Receptive Field (ACSRF) network architecture inspired by the central-peripheral receptive field in biological vision for single-image haze removal. This leads to a unique receptive field mechanism that effectively combines both central and surrounding information. Our ACRSF addresses this by initially compressing global information and then merging it within the CNN, significantly boosting the capability to integrate local and global information, and effectively handling dominant color tones. Experimental results on four publicly available real-world image dehazing datasets show that our ACRSF outperforms current state-of-the-art methods in recovering global information, especially in dominant color tones. Importantly, this technology demonstrates its effectiveness in realistic scenarios, contributing significantly to improving traffic safety in adverse weather conditions. The code is available at https://github.com/JavanTang/ACSRF .}, archive = {J_NEUCOM}, author = {Zhifeng Tang and Yunyao Zhang and Peng Zhang and Wenkang Dong and Zhiyong Zhang and Xingguang Zhang and Ning Zhang}, doi = {10.1016/j.neucom.2024.129075}, journal = {Neurocomputing}, month = {2}, pages = {129075}, shortjournal = {Neurocomputing}, title = {Leveraging central-surrounding receptive fields for single image dehazing}, volume = {618}, year = {2025}, }
• (2025). Class-prior probability estimation using density ratio between unlabeled instances and positively labeled noisy instances. NEUCOM, 618, 129074. (www)

  @article{ , abstract = {In this paper, the authors propose a new approach for determining the mixture proportions of positive and negative instances in both a small set of positively labeled instances contaminated with an unknown proportion of negative instances and an unlabeled instance set with an unknown mixture of positive and negative instances. The unique advantage of this method lies in its ability to derive these proportions with a high level of accuracy by directly leveraging probability density ratio estimation. Experiments conducted on synthetic and UCI repository data demonstrate that the proposed technique can robustly estimate class-prior probabilities with greater speed and better accuracy compared to existing methods, particularly in the presence of instance number and imbalance challenges.}, archive = {J_NEUCOM}, author = {Takeshi Yoshida and Eitaro Shin’ya and Takashi Washio}, doi = {10.1016/j.neucom.2024.129074}, journal = {Neurocomputing}, month = {2}, pages = {129074}, shortjournal = {Neurocomputing}, title = {Class-prior probability estimation using density ratio between unlabeled instances and positively labeled noisy instances}, volume = {618}, year = {2025}, }
• (2025). Context-based emotion recognition: A survey. NEUCOM, 618, 129073. (www)

  @article{ , abstract = {Emotions play a crucial role in human communication, and accurately recognizing them is essential for developing intelligent systems capable of effective human-computer interaction. Contextual information, including body language, tone of voice, situational cues, and facial expressions, significantly influences the analysis of emotions. This survey paper investigates the critical role of context in emotion recognition and comprehensively analyzes the various techniques employed in Context-Based Emotion Recognition (CBER). It highlights the associated challenges and limitations, examines the datasets used in emotion recognition research, and discusses the performance evaluation metrics utilized to assess the accuracy and effectiveness of algorithms and models in understanding emotions within their respective contexts. Furthermore, the paper explores practical applications of context-based emotion recognition across diverse fields. It serves as an invaluable resource for researchers seeking insights into the latest developments in this field and identifies future research directions to advance context-based emotion recognition.}, archive = {J_NEUCOM}, author = {Rizwan Abbas and Bingnan Ni and Ruhui Ma and Teng Li and Yehao Lu and Xi Li}, doi = {10.1016/j.neucom.2024.129073}, journal = {Neurocomputing}, month = {2}, pages = {129073}, shortjournal = {Neurocomputing}, title = {Context-based emotion recognition: A survey}, volume = {618}, year = {2025}, }
• (2025). Consensus of second-order multi-agent systems based on PIDD-like control protocol with time delay. NEUCOM, 618, 129072. (www)

  @article{ , abstract = {This paper concentrates on investigating the consensus for general second-order linear multiagent systems subjected to time-varying delay. Initially, a novel PIDD-like consensus control protocol with time-varying delay is designed, which includes information on position and its integration, as well as velocity and its differentiation. Subsequently, based upon model transformation, the consensus problem is transformed into the asymptotic stability of the error system. Through constructing of a novel augmented vector Lyapunov–Krasovskii functional, including delay-product terms and multiple integral terms, and then by adopting auxiliary-function-based integral inequalities and the delay-dependent-matrix-based reciprocally convex inequality to address the derivative of the constructed functional, the less conservative consensus conditions for multi-agent systems are obtained. Additionally, numerical simulations are given to demonstrate that the presented consensus conditions not only ensure the consensus of position and velocity states but also exhibit better dynamic performance.}, archive = {J_NEUCOM}, author = {Jianfeng Guo and Wei Qian and Yanmin Wu}, doi = {10.1016/j.neucom.2024.129072}, journal = {Neurocomputing}, month = {2}, pages = {129072}, shortjournal = {Neurocomputing}, title = {Consensus of second-order multi-agent systems based on PIDD-like control protocol with time delay}, volume = {618}, year = {2025}, }
• (2025). DSU-net: A dynamic stage unfolding network for high-noise image compressive sensing denoising. NEUCOM, 618, 129071. (www)

  @article{ , abstract = {Deep unfolding networks (DUNs) have demonstrated considerable efficacy in the domain of compressive sensing (CS), attributed to their superior performance and interpretability. Nonetheless, many existing CS methodologies fail to account for the influence of noise prior to image sampling, resulting in blurring and distortion in the reconstructed images. To address these issues, this paper proposes a dynamic stage unfolding network (DSUNet). Firstly, a novel dynamic stage unfolding mechanism is proposed to achieve feature refinement by dynamically optimizing various noisy images in a sequential manner. Secondly, a step inertia fusion module (SIFM) is developed to execute multi-tiered information fusion from adjacent stages, thereby promoting feature reuse and minimizing the loss of detailed information. Finally, a step cross transformer denoiser (SCTD) is designed to capture correlations between distant pixels, effectively addressing the limitations associated with local attributes and significantly improving image reconstruction performance. Comprehensive experimental evaluations indicate that the proposed DSUNet achieves outstanding performance under both low CS ratios and high noise levels, successfully addressing the challenges of denoising and reconstruction in high-resolution and high-noise images. Codes are available at https://github.com/zzuli407/DSU-Net .}, archive = {J_NEUCOM}, author = {Jie Zhang and Miaoxin Lu and Wenxiao Huang and Xiaoping Shi and Yanfeng Wang}, doi = {10.1016/j.neucom.2024.129071}, journal = {Neurocomputing}, month = {2}, pages = {129071}, shortjournal = {Neurocomputing}, title = {DSU-net: A dynamic stage unfolding network for high-noise image compressive sensing denoising}, volume = {618}, year = {2025}, }
• (2025). Fast gradient-free activation maximization for neurons in spiking neural networks. NEUCOM, 618, 129070. (www)

  @article{ , abstract = {Elements of neural networks, both biological and artificial, can be described by their selectivity for specific cognitive features. Understanding these features is important for understanding the inner workings of neural networks. For a living system, such as a neuron, whose response to a stimulus is unknown and not differentiable, the only way to reveal these features is through a feedback loop that exposes it to a large set of different stimuli. The properties of these stimuli should be varied iteratively in order to maximize the neuronal response. To utilize this feedback loop for a biological neural network, it is important to run it quickly and efficiently in order to reach the stimuli that maximizes certain neurons’ activation with the least number of iterations possible. Here we present a framework with an efficient design for such a loop. We successfully tested it on an artificial spiking neural network (SNN), which is a model that simulates the asynchronous spiking activity of neurons in the brain. Our optimization method for activation maximization is based on the low-rank Tensor Train decomposition of the discrete activation function. The optimization space is the latent parameter space of images generated by SN-GAN or VQ-VAE generative models. To our knowledge, this is the first time that gradient-free activation maximization has been applied to SNNs. We track changes in the optimal stimuli for artificial neurons during training and show that highly selective neurons can form already in the early epochs of training and in the early layers of a convolutional spiking network. This formation of refined optimal stimuli is associated with an increase in classification accuracy. Some neurons, especially in the deeper layers, may gradually change the concepts they are selective for during learning, potentially explaining their importance for model performance. The source code of our framework, MANGO (for Maximization of neuronal Activation via Non-Gradient Optimization) is available on GitHub ( https://github.com/iabs-neuro/mango )}, archive = {J_NEUCOM}, author = {Nikita Pospelov and Andrei Chertkov and Maxim Beketov and Ivan Oseledets and Konstantin Anokhin}, doi = {10.1016/j.neucom.2024.129070}, journal = {Neurocomputing}, month = {2}, pages = {129070}, shortjournal = {Neurocomputing}, title = {Fast gradient-free activation maximization for neurons in spiking neural networks}, volume = {618}, year = {2025}, }
• (2025). FGA-YOLO: A one-stage and high-precision detector designed for fine-grained aircraft recognition. NEUCOM, 618, 129067. (www)

  @article{ , abstract = {Aircraft recognition in remote sensing images (RSIs) plays a pivotal role in various applications, such as disaster relief, airport monitoring. However, various aircraft types exhibit subtle inter-class variations, and the same model can display diverse visual characteristics under varying imaging conditions, presenting a significant challenge for the fine-grained aircraft recognition. This paper proposes an efficient detection model called FGA-YOLO. Firstly, we introduce a triple feature fusion module based on feature selection (STFF) to aggregate feature maps from three different levels, thereby constructing an efficient feature pyramid network that enables the network to capture salient information hidden across different levels of feature maps. Furthermore, we design a global multi-scale gather-and-distribute module (GMGD) to inject global information into the network, enhancing the extraction of the objects’ key discriminative features. Finally, we develop EMA Slide Loss by employing exponential moving average (EMA) method to mitigate the imbalance between easy and hard samples in aircraft recognition. Extensive experiments on two aircraft recognition datasets (i.e., MAR20 and FAIR1M) demonstrate that our proposed method can significantly improve the accuracy of fine-grained aircraft recognition. Moreover, experiments on embedded devices demonstrate that the proposed FGA-YOLO is also friendly for lightweight deployment.}, archive = {J_NEUCOM}, author = {Jian Wu and Fanyu Zhao and Guitao Yao and Zhonghe Jin}, doi = {10.1016/j.neucom.2024.129067}, journal = {Neurocomputing}, month = {2}, pages = {129067}, shortjournal = {Neurocomputing}, title = {FGA-YOLO: A one-stage and high-precision detector designed for fine-grained aircraft recognition}, volume = {618}, year = {2025}, }
• (2025). Learning to break: Knowledge-enhanced reasoning in multi-agent debate system. NEUCOM, 618, 129063. (www)

  @article{ , abstract = {Multi-agent debate system (MAD) imitating the process of human discussion in pursuit of truth, aims to align the cognition of different agents to reach the optimal solution. It is challenging for various agents to perform correctly and maintain highly consistent cognition due to their limited and different knowledge backgrounds (i.e., cognitive islands), which hinders the search for the optimal solution. To address the challenge, we propose a novel M ulti- A gent D ebate with K nowledge- E nhanced framework ( MADKE ) to promote the system to find the solution. First, we involve a shared retrieval knowledge pool in the debate process to solve the problem of limited and different knowledge backgrounds. Then, we propose an adaptive knowledge selection method to guarantee the accuracy and personalization of knowledge. This method allows agents to choose whether to use external knowledge in each conversation round according to their own needs. Our experimental results on six datasets show that our method achieves state-of-the-art results compared to existing single-agent and multi-agent methods. Further analysis reveals that the introduction of retrieval knowledge can help the agent to break cognitive islands in the debate process and effectively improve the consistency and correctness of the model. Moreover, MADKE using Qwen1.5-72B-Chat surpasses GPT-4 by +1.26% on average in six datasets, which validates that our method can help open-source LLMs achieve or even surpass the performance of GPT-4. Our code is available at https://github.com/FutureForMe/MADKE .}, archive = {J_NEUCOM}, author = {Haotian Wang and Xiyuan Du and Weijiang Yu and Qianglong Chen and Kun Zhu and Zheng Chu and Lian Yan and Yi Guan}, doi = {10.1016/j.neucom.2024.129063}, journal = {Neurocomputing}, month = {2}, pages = {129063}, shortjournal = {Neurocomputing}, title = {Learning to break: Knowledge-enhanced reasoning in multi-agent debate system}, volume = {618}, year = {2025}, }
• (2025). IntelELM: A python framework for intelligent metaheuristic-based extreme learning machine. NEUCOM, 618, 129062. (www)

  @article{ , abstract = {This study introduces IntelELM, an open-source Python library designed for hybrid neural networks that integrate Extreme Learning Machine (ELM) with Metaheuristic Algorithms (MHAs). Built on the foundations of two well-established libraries, Scikit-Learn and Mealpy, IntelELM offers four primary strategies for addressing regression and classification tasks. These strategies are implemented through the ElmRegressor and ElmClassifier classes for traditional ELM, as well as the MhaElmRegressor and MhaElmClassifier for hybrid metaheuristic-based ELM models. The library is easy to install and use, especially for individuals familiar with the Scikit-Learn ecosystem. IntelELM comprises at least 402 distinct models across these four primary classes, encompassing classical ELM regression and classification models, as well as over 200 metaheuristic-based ELM regression and classification models each. To demontrade the power of the proposed library, we evaluate several hybrid models from the IntelELM library alongside traditional machine learning models across three benchmark datasets. Experimental results demonstrate that the hybrid models within IntelELM exhibit competitive performance across various performance metrics compared to traditional machine learning approaches. These findings underscore the library's potential to offer effective solutions to real-world problems and contribute significantly to the computer science community. We have released the source code of the library as open-source, inviting the research community to conduct widespread evaluations of this comprehensive framework as a promising tool for research studies and real-world solutions. The source code can be found at https://github.com/thieu1995/IntelELM .}, archive = {J_NEUCOM}, author = {Nguyen Van Thieu and Essam H. Houssein and Diego Oliva and Nguyen Duy Hung}, doi = {10.1016/j.neucom.2024.129062}, journal = {Neurocomputing}, month = {2}, pages = {129062}, shortjournal = {Neurocomputing}, title = {IntelELM: A python framework for intelligent metaheuristic-based extreme learning machine}, volume = {618}, year = {2025}, }
• (2025). SO-TAD: A surveillance-oriented benchmark for traffic accident detection. NEUCOM, 618, 129061. (www)

  @article{ , abstract = {Automatic traffic accident detection is a crucial component of Vehicle-to-Infrastructure (V2I) systems and is gaining increasing attention from researchers due to its significant impact on road safety. However, existing datasets often suffer from issues such as data contamination, small scale, and lack of open access, which hinder the advancement of research in this field. To overcome these challenges, we introduce the high-quality, large-scale Surveillance-Oriented Traffic Accident Detection (SO-TAD) dataset, designed with long-tail distribution characteristics. The SO-TAD dataset comprises 2186 samples, including 282 traffic accident instances, each providing spatio-temporal localization information. Furthermore, within the proposed frame prediction-based traffic accident detection framework, we designed a three-channel scene feature fusion generative adversarial network (TCFF-GAN). This network integrates scene introduction features and scene difference features to generate predictive frame features. Subsequently, traffic accidents are detected based on the stability of the cosine similarity sequence between the predicted frames and the actual frames. We validated the effectiveness of the proposed TCFF-GAN not only through public datasets but also by conducting extensive experiments on the newly introduced SO-TAD dataset. This work also establishes a new benchmark for automatic traffic accident detection research. The code and dataset are available at https://github.com/cccxy-299/so-tad .}, archive = {J_NEUCOM}, author = {Xingyuan Chen and Huahu Xu and Mingyang Ruan and Minjie Bian and Qishen Chen and Yuzhe Huang}, doi = {10.1016/j.neucom.2024.129061}, journal = {Neurocomputing}, month = {2}, pages = {129061}, shortjournal = {Neurocomputing}, title = {SO-TAD: A surveillance-oriented benchmark for traffic accident detection}, volume = {618}, year = {2025}, }
• (2025). DPC-MFP: An adaptive density peaks clustering algorithm with multiple feature points. NEUCOM, 618, 129060. (www)

  @article{ , abstract = {Density peaks clustering (DPC) is a simple and efficient density-based clustering algorithm. However, the original DPC algorithm still has some shortcomings including manual selection of cluster centers, single cluster center representation of each cluster, and error cascade in the one-step assignment strategy of the remaining data points. To address these shortcomings, a novel density peaks clustering algorithm with multiple feature points adapting to the cluster structures (DPC-MFP) is developed. First, the DPC-MFP algorithm automatically selects multiple feature points for each cluster to eliminate the influence of manual cluster center selection. Second, multiple feature points belonging to the same cluster are identified by their similarity and are further used to represent the same cluster. DPC-MFP can adapt to the complex structures of different clusters and overcome the shortcoming of the inability of cluster centers used in DPC in representing complex clusters. Finally, in order to overcome the shortcoming of the one-step assignment strategy in DPC, dependent points are divided into ascertained and uncertain dependent points, and a hybrid strategy is adopted to assign the remaining points to clusters. Numerical experiments on a large number of real-word and synthetic datasets show that the DPC-MFP algorithm outperforms other existing algorithms, especially on some synthetic datasets with complex cluster structures.}, archive = {J_NEUCOM}, author = {Wenke Zang and Xincheng Liu and Linlin Ma and Minghe Sun and Jing Che and Yuzhen Zhao and Yuanhua Wang and Da Wang and Xiyu Liu}, doi = {10.1016/j.neucom.2024.129060}, journal = {Neurocomputing}, month = {2}, pages = {129060}, shortjournal = {Neurocomputing}, title = {DPC-MFP: An adaptive density peaks clustering algorithm with multiple feature points}, volume = {618}, year = {2025}, }
• (2025). Envelope rotation forest: A novel ensemble learning method for classification. NEUCOM, 618, 129059. (www)

  @article{ , abstract = {Ensemble learning is an important field and research hotpot of machine learning. However, the existing ensemble learning methods construct subsets by directly sampling the same original sample set. Therefore, the subsets suffer from the low quality of original sample set, such as low separability, low diversity, high sensitivity to noisy samples. As a result, the improvement effect of existing ensemble methods is limited. To address these problems, a new ensemble learning method - envelope rotation forest was proposed. Firstly, envelope sample transformation mechanism (ESTM) is designed to transform the original sample set to multiple layers of envelope sample sets, thereby improving the quality of the sample subsets. The ESTM consists of multilayer iterative mean clustering (MIMC) and interlayer consistency mechanism (ICM). Secondly, the two-dimensional base classifier matrix generation and fusion mechanism (2D_BCMG&FM) is designed to enhance classification performance. The final classification results are obtained by fusing the two-dimensional classification labels. For verification, the proposed method is tested on 26 datasets and is compared with representative methods. It is observed that the accuracy is improved by up to 18.77 %, which highlights the significance of considering envelope transformation on original samples for subsequent subset construction. Data and code can be found at: https://github.com/GRabase/ERF}, archive = {J_NEUCOM}, author = {Jie Ma and Huan Cheng and Hong Chen and Yi Zhang and Yongming Li and Yinghua Shen and Pin Wang and Wei Wang}, doi = {10.1016/j.neucom.2024.129059}, journal = {Neurocomputing}, month = {2}, pages = {129059}, shortjournal = {Neurocomputing}, title = {Envelope rotation forest: A novel ensemble learning method for classification}, volume = {618}, year = {2025}, }
• (2025). General decay anti-synchronization of coupled delayed memristive neural networks with constant and time-varying distributed-delay coupling. NEUCOM, 618, 129058. (www)

  @article{ , abstract = {In this study, we investigate the general decay anti-synchronization (GDAS) problem for coupled delayed memristive neural networks (CDMNNs) and delayed CDMNNs (DCDMNNs) with constant and time-varying distributed-delay coupling, respectively. First, we introduce the concept of GDAS to the considered network on the basis of the definitions of ψ -type stability and the ψ -type function. Furthermore, several criteria are inferred for the GDAS of CDMNNs with constant and time-varying distributed-delay coupling by selecting the correct Lyapunov functionals and devising an appropriate nonlinear controller. Additionally, several sufficient conditions for realizing GDAS for DCDMNNs with constant and time-varying distributed-delay coupling are obtained. Finally, simulation experiments demonstrate the feasibility of the obtained conclusions.}, archive = {J_NEUCOM}, author = {Yan-Li Huang and Ao-Bo Li}, doi = {10.1016/j.neucom.2024.129058}, journal = {Neurocomputing}, month = {2}, pages = {129058}, shortjournal = {Neurocomputing}, title = {General decay anti-synchronization of coupled delayed memristive neural networks with constant and time-varying distributed-delay coupling}, volume = {618}, year = {2025}, }
• (2025). Trajectory tracking of QUAV based on cascade DRL with feedforward control. NEUCOM, 618, 129057. (www)

  @article{ , abstract = {The adoption of advanced control strategies has become increasingly critical for the quadrotor unmanned aerial vehicle (QUAV). Deep reinforcement learning (DRL) stands at the forefront of these developments, showcasing significant utility, particularly in the realm of static target tracking for QUAVs. However, existing DRL methodologies encounter substantial challenges with latency issues when applied to trajectory tracking scenarios. This paper tackles this challenge by incorporating a feedforward technique, which empowers agents to analyze high order trajectory information. Initially, a cascade DRL controller for fixed point tracking is trained, where multiple agents are separately responsible for translation and rotation movement along each axis. Then, the controller is implemented for trajectory tracking by incorporating the feedforward information from the trajectory, which significantly solves the latency issue. Through extensive tracking demonstrations and quantitative analysis, the proposed integrated scheme demonstrates a significant enhancement in the trajectory tracking performance of QUAVs.}, archive = {J_NEUCOM}, author = {Shuliang He and Haoran Han and Jian Cheng}, doi = {10.1016/j.neucom.2024.129057}, journal = {Neurocomputing}, month = {2}, pages = {129057}, shortjournal = {Neurocomputing}, title = {Trajectory tracking of QUAV based on cascade DRL with feedforward control}, volume = {618}, year = {2025}, }
• (2025). Deep learning technology for face forgery detection: A survey. NEUCOM, 618, 129055. (www)

  @article{ , abstract = {Currently, the rapid development of computer vision and deep learning has enabled the creation or manipulation of high-fidelity facial images and videos via deep generative approaches. This technology, also known as deepfake, has achieved dramatic progress and become increasingly popular in social media. However, the technology can generate threats to personal privacy and national security by spreading misinformation. To diminish the risks of deepfake, it is desirable to develop powerful forgery detection methods to distinguish fake faces from real faces. This paper presents a comprehensive survey of recent deep learning-based approaches for facial forgery detection. We attempt to provide the reader with a deeper understanding of the current advances as well as the major challenges for deepfake detection based on deep learning. We present an overview of deepfake techniques and analyze the characteristics of various deepfake datasets. We then provide a systematic review of different categories of deepfake detection and state-of-the-art deepfake detection methods. The drawbacks of existing detection methods are analyzed, and future research directions are discussed to address the challenges in improving both the performance and generalization of deepfake detection.}, archive = {J_NEUCOM}, author = {Lixia Ma and Puning Yang and Yuting Xu and Ziming Yang and Peipei Li and Huaibo Huang}, doi = {10.1016/j.neucom.2024.129055}, journal = {Neurocomputing}, month = {2}, pages = {129055}, shortjournal = {Neurocomputing}, title = {Deep learning technology for face forgery detection: A survey}, volume = {618}, year = {2025}, }
• (2025). A fast sparse graph based clustering technique using dispersion of data points. NEUCOM, 618, 129054. (www)

  @article{ , abstract = {Minimum spanning tree (MST) have been employed in practice for various exploratory data analyses, e.g., to discover clusters of arbitrary shapes and sizes from diversified datasets. However, the computational complexity of these algorithms becomes a bottleneck when they are applied on very large datasets. The main overhead associated with these algorithms is the proximity search in the construction of a similarity graph which incurs O ( N 2 ) time on a set of N data points. To conquer this issue, several graph sparsification techniques have been proposed which take O ( N 3 / 2 ) time. This paper proposes a O ( N 4 / 3 ) time local neighborhood similarity graph construction technique using two levels of partitioning and merging, which is asymptotically O ( N 1 / 6 ) factor improvement over the existing methods. To the best of our knowledge, this is the asymptotically fastest known algorithm using two levels of partitioning and merging. Experimental analysis shows that the proposed sparse graph construction technique reduces 99.46% edges of the complete graph by preserving the relevant neighborhood information. Also, MST constructed from the proposed graph captures the local neighborhood of data points efficiently which is shown in terms of edge error and weight error rates. Finally, we demonstrate that the proposed approximate MST based clustering technique outperforms the best-known existing algorithms in terms of clustering accuracy and execution time on diversified datasets.}, archive = {J_NEUCOM}, author = {Mohammad Maksood Akhter and Abdul Atif Khan and Rashmi Maheshwari and R. Jothi and Sraban Kumar Mohanty}, doi = {10.1016/j.neucom.2024.129054}, journal = {Neurocomputing}, month = {2}, pages = {129054}, shortjournal = {Neurocomputing}, title = {A fast sparse graph based clustering technique using dispersion of data points}, volume = {618}, year = {2025}, }
• (2025). Multiple adverse weather image restoration: A review. NEUCOM, 618, 129044. (www)

  @article{ , abstract = {Adverse weather conditions significantly impact the machine vision perception of unmanned platforms such as drones and autonomous vehicles. Therefore, restoring and enhancing images affected by these conditions is crucial. However, there is a lack of a comprehensive systematic review on the latest research progress in unified weather image restoration under all-in-one weather conditions and other factors. In this paper, we survey recent progress in various network architectures for multitask image restoration. Specifically, we review and compare different methods including, U-Net, GAN, Transformer, and U-Net/Transformer hybrids. Additionally, we evaluate publicly available datasets for single-task scenarios and compare the performance of these methods comprehensively, and analyze corresponding evaluation metrics. Based on these findings, we believe that Transformer and U-Net models are particularly promising for multi-task image restoration. Nevertheless, further research is needed to fully explore this area. Researchers can improve the image restoration effect from aspects such as data dynamic flow, encoder–decoder internal structure, etc. These insights contribute to advancing image acquisition technologies and addressing challenges in military image information retrieval, including military reconnaissance, terrain analysis, target identification, and battlefield surveillance.}, archive = {J_NEUCOM}, author = {Hanguang Xiao and Shihong Liu and Kun Zuo and Haipeng Xu and Yuyang Cai and Tianqi Liu and Zhiying Yang}, doi = {10.1016/j.neucom.2024.129044}, journal = {Neurocomputing}, month = {2}, pages = {129044}, shortjournal = {Neurocomputing}, title = {Multiple adverse weather image restoration: A review}, volume = {618}, year = {2025}, }
• (2025). Enhancing performance of 3D point completion network using consistency loss. NEUCOM, 618, 129037. (www)

  @article{ , abstract = {Point cloud completion networks are conventionally trained to minimize the disparities between the completed point cloud and the ground-truth counterpart. However, an incomplete object-level point cloud can have multiple valid completion solutions when it is examined in isolation. This one-to-many mapping issue can cause contradictory supervision signals to the network because the loss function may produce different values for identical input–output pairs of the network. In many cases, this issue could adversely affect the network optimization process. In this work, we propose to enhance the conventional learning objective using a novel completion consistency loss to mitigate the one-to-many mapping problem. Specifically, the proposed consistency loss ensure that a point cloud completion network generates a coherent completion solution for incomplete objects originating from the same source point cloud. Experimental results across multiple well-established datasets and benchmarks demonstrated the proposed completion consistency loss have excellent capability to enhance the completion performance of various existing networks without any modification to the design of the networks. The proposed consistency loss enhances the performance of the point completion network without affecting the inference speed, thereby increasing the accuracy of point cloud completion. Notably, a state-of-the-art point completion network trained with the proposed consistency loss can achieve state-of-the-art accuracy on the challenging new MVP dataset. The code and result of experiment various point completion models using proposed consistency loss will be available at: https://github.com/kaist-avelab/ConsistencyLoss .}, archive = {J_NEUCOM}, author = {Kevin Tirta Wijaya and Christofel Rio Goenawan and Seung-Hyun Kong}, doi = {10.1016/j.neucom.2024.129037}, journal = {Neurocomputing}, month = {2}, pages = {129037}, shortjournal = {Neurocomputing}, title = {Enhancing performance of 3D point completion network using consistency loss}, volume = {618}, year = {2025}, }
• (2025). Few-shot learning with distribution calibration for event-level rumor detection. NEUCOM, 618, 129034. (www)

  @article{ , abstract = {With the rapid evolution of social media, rumors travel at unprecedented speeds. Automatic recognition of rumors is important for making users receive truthful information and maintaining social harmony. Recently, deep learning models have demonstrated strong rumor detection ability by capturing semantics of rumors, learning rumor propagation patterns and integrating users’ properties, etc. However, most existing rumor detection models perform poorly on unseen events because they are powerful at capturing event-specific features in seen data, which cannot be transferred to newly emergent events due to distribution differences between events. Therefore, in this study, we propose a novel model named E-Rumor in a few-shot setting to learn and transfer event-invariant features from historic events with sufficient samples to new events with only a few examples. The model first calculates the rumor class distributions of historic rumors, and then calibrates the class distributions of a new event with the old ones. Furthermore, an adequate number of samples are generated from the calibrated distributions to expand the training set for a new-event rumor classifier. The proposed model can be paired with any rumor detection classifier as a feature extractor without extra parameters. Empirical studies have shown that a simple MLP (Multilayer Perceptron) trained on the samples generated from the calibrated distributions can outperform the state-of-the-art baseline models on two benchmark rumor datasets PHEME5, PHEME9 and one topic-level dataset T-Twitter, which is generated by topic extraction and cross-data manner, with more than 15.15%, 4.7% and 7.38% accuracy improvements, respectively.}, archive = {J_NEUCOM}, author = {Hongyan Ran and Caiyan Jia and Xiaohong Li and Zhichang Zhang}, doi = {10.1016/j.neucom.2024.129034}, journal = {Neurocomputing}, month = {2}, pages = {129034}, shortjournal = {Neurocomputing}, title = {Few-shot learning with distribution calibration for event-level rumor detection}, volume = {618}, year = {2025}, }
• (2025). Fusion or not: Learning visual relocalization with matrix fisher distribution. NEUCOM, 618, 129033. (www)

  @article{ , abstract = {Visual relocalization is crucial for applications such as simultaneous localization and mapping (SLAM) and augmented reality. State-of-the-art methods based on multiple view geometry struggles with robust localization in challenging environments due to visual perturbations. Recently, deep learning-based absolute pose regression (APR) has gained significant attention owing to its robust feature extraction capabilities, yet it consistently falls short of matching the localization accuracy achieved by structure-based localization methods. In the paper, a novel visual localization framework is developed, which fuses geometric structure methods with neural network inference via the matrix Fisher distribution on the special orthogonal group SO ( 3 ) . To be specific, a deep neural network is trained to generate the exponential probability density model on SO ( 3 ) , i.e., the matrix Fisher distribution, which does not require the implementation of complex functions to constrain the rotational variables and can simultaneously model the uncertainty of the rotation estimates. A Bayesian fusion method is then introduced to fuse the orientation predictions recovered from the Fisher parameters with the rotation estimates computed by a structure-based approach. Moreover, the fusion results are concatenated with the feature encoding of the image to compensate for the correlation with the rotational variables, and the position prediction is output through a fully connected layer. The performance of the developed localization framework is evaluated via public indoor and outdoor datasets, demonstrating that the approach significantly improves localization accuracy while remaining robust in challenging environments.}, archive = {J_NEUCOM}, author = {Minglei Fu and Shengzhou Li and Yuqiang Jin and Wen-An Zhang and Uladzislau Sychou and Vadim Skobtsov and Vladislav Sobolevskii and Boris Sokolov}, doi = {10.1016/j.neucom.2024.129033}, journal = {Neurocomputing}, month = {2}, pages = {129033}, shortjournal = {Neurocomputing}, title = {Fusion or not: Learning visual relocalization with matrix fisher distribution}, volume = {618}, year = {2025}, }
• (2025). A systematic study of echo state networks topologies for chaotic time series prediction. NEUCOM, 618, 129032. (www)

  @article{ , abstract = {In the last twenty years, Echo State Networks have become a prominent method for the prediction of time series with a large variety of proposed topologies of connections. These topologies inside the ESN are typically shown to be advantageous for a certain prediction task in terms of reducing the computational complexity and increasing the explainability regarding the selection of weights. Still, a thorough comparison of these different topologies is missing in terms of their formal description and their performance as well as characteristics when applied to different prediction tasks. In this paper, we study 16 topologies for the task of time series prediction in the context of chaotic dynamics. We restrict our focus to those since they are considered among the most challenging to predict, while still being of interest for applications and being representative for other time series. We categorize the selected topologies into intralevel connections, concurrent and sequential reservoirs as well as complex topologies and implemented all of them. We parametrized all according to their original publications but also ran additional tests regarding suitable parametrization. All topologies are evaluated on the eight time series, in auto-regressive as well as single-step predictions. Our results emphasize on the benefit of choosing a suitable topology, reducing the error by up to 94.87% compared to the baseline, the conventional Echo State Network. We also compared the 16 topologies to the widely used Gated Recurrent Unit (GRU) network and observe the structured reservoir yielding a decreased error of up to 99.99% compared to the GRU.}, archive = {J_NEUCOM}, author = {Johannes Viehweg and Philipp Teutsch and Patrick Mäder}, doi = {10.1016/j.neucom.2024.129032}, journal = {Neurocomputing}, month = {2}, pages = {129032}, shortjournal = {Neurocomputing}, title = {A systematic study of echo state networks topologies for chaotic time series prediction}, volume = {618}, year = {2025}, }
• (2025). Modeling and optimization of trajectory deviation for compound directional drilling in coal mines. NEUCOM, 618, 129029. (www)

  @article{ , abstract = {During the sliding mode of compound directional drilling process in coal mines, the inclined angle and azimuth angle need to be changed to minimize the trajectory deviation between the actual trajectory and the designed trajectory. The inclined angle and azimuth angle can be changed by adjusting the tool face angle. In this article, a modeling and optimization method is proposed to obtain optimal tool face angle for sliding mode. Firstly, the gaussian process regression method is employed to build calculation model of trajectory deviation. The optimization model of trajectory deviation consists of the established calculation model of trajectory deviation and the set constraints. Then, the optimal inclination angle and azimuth angle are obtained by solving the constitutive optimization model using an improved particle swarm optimization algorithm. Next, the calculation model of tool face angle is built to represent the relationship between inclination angle, azimuth angle and tool face angle. The optimal tool face angle is calculated based on optimal inclined angle and azimuth angle. Finally, an industrial case study based on the actual drilling data verifies that the proposed method is beneficial to track designed trajectory during drilling process. The proposed method meets the requirements of on-site construction accuracy and ensures the safety and stability of drilling.}, archive = {J_NEUCOM}, author = {Wangnian Li and Xiao Yang and Chengda Lu and Quanxin Li and Peng Fang and Xiao Wu and Hengyu Huang and Haipeng Fan and Ningping Yao and Hongliang Tian and Min Wu}, doi = {10.1016/j.neucom.2024.129029}, journal = {Neurocomputing}, month = {2}, pages = {129029}, shortjournal = {Neurocomputing}, title = {Modeling and optimization of trajectory deviation for compound directional drilling in coal mines}, volume = {618}, year = {2025}, }
• (2025). CSPDAE: Colorization via SuperPixel downsampler denoising AutoEncoder towards enhanced color structural clarity. NEUCOM, 618, 129028. (www)

  @article{ , abstract = {Based on the various structural textures, sizes, and shapes of the gray image, the SuperPixel (SP)-based methods convert the rigid pixels into adaptable image patches that share common features. As a result, the SP-based colorization technique not only enhances the color appearance but also preserves the topological integrity of the images. Despite the effectiveness of neural network-based colorization methods, their integration with SP segmentation has traditionally been complex and cumbersome. To address this issue, we propose the C olorization SP Downsampler D enoising A uto E ncoder (CSPDAE), where the SP downsampler integrates SP segmentation directly into the network, eliminating the need for any prior input. The SP downsampler addresses the challenges of computational complexity and small region clustering, which are the main obstacles preventing Transformer architectures from being applied to pixel-level segmentation, through the use of SP cross-attention and aggregated positional embedding (APE). Furthermore, we have incorporated a Color Weight (CW) loss, based on the CIEDE2000 color difference formula, to ensure balanced pixel sampling and to improve the precision of detailed color representation. The experimental results confirm the effectiveness of our method, demonstrating its capacity to produce colors with greater structural accuracy and visually appealing details.}, archive = {J_NEUCOM}, author = {Wenqu Zhao and Lingxue Wang and Miao Dong and Yi Cai}, doi = {10.1016/j.neucom.2024.129028}, journal = {Neurocomputing}, month = {2}, pages = {129028}, shortjournal = {Neurocomputing}, title = {CSPDAE: Colorization via SuperPixel downsampler denoising AutoEncoder towards enhanced color structural clarity}, volume = {618}, year = {2025}, }
• (2025). Multi-view topology assisted dynamic graph learning for fMRI-based alzheimer’s disease identification. NEUCOM, 618, 129025. (www)

  @article{ , abstract = {Resting-state functional magnetic resonance imaging (rs-fMRI) provides a non-invasive technology that helps identify abnormal connections in the brain by tracking changes in blood flow components. Functional connectivity networks (FCNs) derived from fMRI data are commonly used for identifying brain disorders such as Alzheimer’s disease (AD). Unfortunately, most existing AD identification models rely solely on static single-view FCN (e.g., Pearson correlation, PC), neglecting the temporal information and complex topological structures present in fMRI data. Additionally, the imbalance of training data significantly hampers the generalization ability of the model. To this end, we propose a multi-view topology assisted dynamic graph learning (MTDGL) framework for fMRI analysis and AD automated identification, with a three-pipeline fMRI feature extraction structure composed of a spatiotemporal branch, a topological branch and a raw information branch. At the outset, we adopt a category balancing strategy for the training data to enhance the representation ability of the model. In the spatiotemporal branch, we employ a spatiotemporal attention graph isomorphism network (STAGIN) to capture latent dynamic representations in fMRI data. In the topological branch, we integrate topological information extracted from multiple views (i.e., PC, mutual information, MI). In the raw information branch, we extract the fine-grained information in the raw signal and adaptively fuse them with the dynamic representations for the final AD classification task. This can overcome the limitations of single-view methods in describing topological structures. In the feature fusion stage, the relevant features are first integrated, and then the original signal features are merged into these integrated relevant features through the Transformer model to achieve deep fusion between the two main information streams. The experimental results on the ADNI dataset demonstrate that the proposed MTDGL achieves an accuracy of 92.43% in AD recognition, outperforming several state-of-the-art approaches.}, archive = {J_NEUCOM}, author = {Xinyue Yan and Xianfu Zhang and Shengxiang Xia}, doi = {10.1016/j.neucom.2024.129025}, journal = {Neurocomputing}, month = {2}, pages = {129025}, shortjournal = {Neurocomputing}, title = {Multi-view topology assisted dynamic graph learning for fMRI-based alzheimer’s disease identification}, volume = {618}, year = {2025}, }
• (2025). Fine-grained adaptive contrastive learning for unsupervised feature extraction. NEUCOM, 618, 129014. (www)

  @article{ , abstract = {Self-supervised contrastive learning (SCL) has emerged as a crucial technique for unsupervised fine-grained feature extraction in the deep learning field. However, model collapse is a prevalent issue that hinders the potential of this approach, particularly in fine-grained tasks. To address the performance degradation in SCL, we propose a novel framework called Fine-grained Adaptive Contrastive Learning (FACL), which consists of three carefully designed modules. The first module is the adaptive center-point regression, which empowers the segment anything model to extract regions of interest adaptively. The second module, Diffusion Gaussian Noise, creates proxy distributions for the original samples, thereby augmenting the data with increased diversity. Finally, the lightweight multi-scale instant extraction module improves the efficiency and accuracy of feature extraction by integrating multi-scale sensory fields. Extensive experiments on real-world datasets demonstrate the effectiveness of the FACL framework in various tasks, including biometric authentication and anomaly detection. FACL outperforms existing SCL methods, reducing the equal error rate up to 54.8% and 64.1% compared to SimCLR and BYOL, respectively.}, archive = {J_NEUCOM}, author = {Tianyi Yin and Jingwei Wang and Yukai Zhao and Han Wang and Yunlong Ma and Min Liu}, doi = {10.1016/j.neucom.2024.129014}, journal = {Neurocomputing}, month = {2}, pages = {129014}, shortjournal = {Neurocomputing}, title = {Fine-grained adaptive contrastive learning for unsupervised feature extraction}, volume = {618}, year = {2025}, }
• (2025). RAMIS: Increasing robustness and accuracy in medical image segmentation with hybrid CNN-transformer synergy. NEUCOM, 618, 129009. (www)

  @article{ , abstract = {Hybrid architectures based on Convolutional Neural Network (CNN) and Vision Transformer (ViT) have become an important research direction in medical image segmentation in recent years. However, the currently popular hybrid architectures weaken the decision making process within the Transformer model, the way in which the output of the Transformer is post-processed by the upsampling of the convolution stack makes it difficult to restore the blurred boundaries of the target area. To improve the feature learning capability by addressing these issues, we propose RAMIS, a novel hybrid architecture for general medical image segmentation. RAMIS develops implicit neural representation and self-distillation to simultaneously obtain the super-resolution details and core features of the image as input to the Transformer encoder. Meanwhile, RAMIS explores an unsupervised learning CNN to obtain the initial input to the Transformer decoder, which not only explicitly considers the correlation within different samples, reduces the constraints on small datasets, but also fully leverages the potential of Transformer’s cross-attention for optimizing segmentation results. RAMIS designs a multi-resolution interaction network to post-process the Transformer output and solves the problem of blurred segmentation boundaries by combining super-resolution image. We extensively evaluate RAMIS on five datasets from three typical publicly available medical image segmentation datasets. Extensive experimental results demonstrate the general applicability and superior performance of the proposed method. The code and pre-trained models are available on our website https://ramis.netlify.app .}, archive = {J_NEUCOM}, author = {Jia Gu and Fangzheng Tian and Il-Seok Oh}, doi = {10.1016/j.neucom.2024.129009}, journal = {Neurocomputing}, month = {2}, pages = {129009}, shortjournal = {Neurocomputing}, title = {RAMIS: Increasing robustness and accuracy in medical image segmentation with hybrid CNN-transformer synergy}, volume = {618}, year = {2025}, }
• (2025). Zonotopic set-membership state estimation for nonlinear systems based on the deep koopman operator. NEUCOM, 618, 129004. (www)

  @article{ , abstract = {In this study, a novel nonlinear zonotopic set-membership estimation (SME) approach is proposed based on the Koopman operator and the deep neural networks (DNNs). The core concept involves transforming the original nonlinear system into a higher-dimensional linear system, enabling the development of linear SME algorithms. To address noise in real-world data, we propose a three-step offline training strategy for computing the linear approximation of nonlinear dynamics. Building on the deep-Koopman-based linearized system, we construct a reduced-order filter structure that directly estimates the original states and then computes the lifted states for improved observability. We derive two optimal gains based on different size criteria and present a data-driven nonlinear mapping approach that ensures a closed-form solution for the algorithm. Additionally, we introduce a local model update method that refines estimation accuracy using the set of state estimates obtained from SME. The effectiveness and superiority of our proposed methods is demonstrated through two illustrative examples.}, archive = {J_NEUCOM}, author = {Zhichao Pan and Siyu Liu and Biao Huang and Fei Liu}, doi = {10.1016/j.neucom.2024.129004}, journal = {Neurocomputing}, month = {2}, pages = {129004}, shortjournal = {Neurocomputing}, title = {Zonotopic set-membership state estimation for nonlinear systems based on the deep koopman operator}, volume = {618}, year = {2025}, }
• (2025). Select & re-rank: Effectively and efficiently matching multimodal data with dynamically evolving attention. NEUCOM, 618, 129003. (www)

  @article{ , abstract = {Extracting semantically consistent representations from multi-modal data helps computers understand the human world more comprehensively. Visual-semantic matching, as one of the fundamental tasks for multi-modal learning, attracts continuous attention. Recent research makes unflagging endeavors to enhance the matching performance, but sometimes at the expense of overlooking the delicate balance between efficiency and effectiveness. In this paper, we aim to address this dilemma through a newly proposed attention-mechanism-based architecture. To ensure optimal effectiveness , we adopt a more advanced Transformer Encoder (TE) as our basic model and introduce two significant ameliorations to tailor it for the visual-semantic matching task. Initially, we incorporate fine-grained supervision into the classic TE, allowing our model to capture sophisticated correspondences between different modalities. Subsequently, we employ a dynamic attention-evolving strategy to selectively pass useful information and strengthen the attention pattern consistency between adjacent TE blocks. To maintain efficiency , we propose a novel Select & Re-rank strategy that enables the model to ignore redundant information. This approach significantly reduces the computational cost and increases the matching speed with relatively minimal performance degradation. The proposed architecture can gradually capture and reorganize useful information from inter-modality as well as intra-modality under the supervision of both fine-grained and global similarity, which leads to more comprehensive and discriminative embeddings. Experiments on two benchmark datasets show that the proposed method achieves competitive results in terms of both efficiency and effectiveness.}, archive = {J_NEUCOM}, author = {Weikuo Guo and Xiangwei Kong and Huaibo Huang}, doi = {10.1016/j.neucom.2024.129003}, journal = {Neurocomputing}, month = {2}, pages = {129003}, shortjournal = {Neurocomputing}, title = {Select & re-rank: Effectively and efficiently matching multimodal data with dynamically evolving attention}, volume = {618}, year = {2025}, }
• (2025). DCVSMNet: Double cost volume stereo matching network. NEUCOM, 618, 129002. (www)

  @article{ , abstract = {We introduce the Double Cost Volume Stereo Matching Network (DCVSMNet 1 ), a novel architecture characterized by two upper (group-wise correlation) and lower (norm correlation) small cost volumes. Each cost volume is processed separately, and a coupling module is proposed to fuse the geometry information extracted from the upper and lower cost volumes. DCVSMNet is a fast stereo matching network with a 67 ms inference time and strong generalization ability which can produce competitive results compared to state-of-the-art methods. The results on several benchmark datasets show that DCVSMNet achieves better accuracy than methods such as CGI-Stereo and BGNet at the cost of greater inference time.}, archive = {J_NEUCOM}, author = {Mahmoud Tahmasebi and Saif Huq and Kevin Meehan and Marion McAfee}, doi = {10.1016/j.neucom.2024.129002}, journal = {Neurocomputing}, month = {2}, pages = {129002}, shortjournal = {Neurocomputing}, title = {DCVSMNet: Double cost volume stereo matching network}, volume = {618}, year = {2025}, }
• (2025). MixEI: Mixing explicit and implicit commonsense knowledge in open-domain dialogue response generation. NEUCOM, 618, 128999. (www)

  @article{ , abstract = {The inadequate awareness of real-world knowledge often causes machines to produce generic responses, such as ‘I think so.’, which may bring the degression of user interests. Consequently, enriching knowledge awareness and fabricating informative responses are long-standing challenges in open-domain dialogue systems. Previous studies have shown incorporating everyday commonsense knowledge can significantly enhance the open-domain dialogue response models. Nonetheless, previous works can only use explicit or implicit knowledge. Unlike them, this work presents a novel MixEI to leverage both explicit and implicit commonsense knowledge in dialogue generation. MixEI uses Dual-Way Knowledge Alignment and MixEI-Ranker to retrieve a set of contextually relevant commonsense facts as the explicit background knowledge and identify implicit knowledge labels by selecting clue facts that can tightly connect the dialogue context. MixEI uses BART as the backbone. After jointly encoding the background knowledge and dialogue history, MixEI first tries to externalize the implicit clue knowledge; then, the response decoding can seek information from both explicit and implicit knowledge. Extensive experiments on Chinese Weibo and English Reddit have verified the superior performance of the proposed MixEI-Ranker and MixEI .}, archive = {J_NEUCOM}, author = {Sixing Wu and Jiong Yu and Wei Zhou}, doi = {10.1016/j.neucom.2024.128999}, journal = {Neurocomputing}, month = {2}, pages = {128999}, shortjournal = {Neurocomputing}, title = {MixEI: Mixing explicit and implicit commonsense knowledge in open-domain dialogue response generation}, volume = {618}, year = {2025}, }
• (2025). Audio–visual correspondences based joint learning for instrumental playing source separation. NEUCOM, 618, 128997. (www)

  @article{ , abstract = {Sound source separation is usually challenged by the identification of intermittently static sounding objects in realistic scenarios, especially when the sound generation is intrinsically related to variational motion as instrumental playing. Current solutions based on audio–visual learning have demonstrated a capability to enhance the sound source by modality interaction, but the training sample deficiency caused by unreliable pair-wise annotation still limits the separation performance to be further improved. Inspired by the sound-motion connection in realistic symphony scenario, in this work, a two-stage audio–visual network is proposed for instrumental audio separation. With a coarse-grained to fine-grained separation, a Patch-level Channel Audio–Visual Interaction (PCAVI) module is also introduced for elegant association between features of audio-motion modalities. By designing a regularization loss to compensate the imbalance of audio–visual feature distribution, a novel circulant learning is proposed to achieve more accurate and reliable performance. Substantial experiments have demonstrated that the proposed two-stage audio–visual instrumental audio separation network outperforms the motion-free methods on the MUSIC dataset, and achieves strong audio separation results in challenging real-world symphony scenarios.}, archive = {J_NEUCOM}, author = {Tianyu Liu and Peng Zhang and Siliang Wang and Wei Huang and Yufei Zha and Yanning Zhang}, doi = {10.1016/j.neucom.2024.128997}, journal = {Neurocomputing}, month = {2}, pages = {128997}, shortjournal = {Neurocomputing}, title = {Audio–visual correspondences based joint learning for instrumental playing source separation}, volume = {618}, year = {2025}, }
• (2025). Mobile crowdsourcing based on 5G and 6G: A survey. NEUCOM, 618, 128993. (www)

  @article{ , abstract = {The rapid development of the Internet of Things (IoT), 5G and 6G networks, and advancements in hardware and software devices have enabled mobile devices to be equipped with rich sensors (smartphones, smart bracelets, etc.). This led to devices with powerful sensing capabilities that can easily connect and communicate with other devices. As a result, ordinary users could efficiently complete various sensing data tasks and get paid accordingly, promoting the rapid development of Mobile Crowdsourcing (MCS) technology and effectively facilitating people’s daily lives. However, the current MCS is still a long way from being complete and still faces several challenges. Therefore, many researchers have conducted extensive research on MCS in recent years. In this survey, we aim to provide a comprehensive summary of the recent research progress on MCS. Concerning the main research directions and real-world applications of MCS, we review the research in three main areas (task allocation, privacy protection, and incentive mechanism) and real-world application scenarios of MCS. Although there is a lot of existing research on MCS, there are still some crucial problems to be solved. Therefore, we conclude with some lessons learned and suggest future research directions by analyzing the studies covered in this paper.}, archive = {J_NEUCOM}, author = {Yingjie Wang and Yingxin Li and Weilong Wang and Peiyong Duan and Akshita Maradapu Vera Venkata Sai and Zhipeng Cai}, doi = {10.1016/j.neucom.2024.128993}, journal = {Neurocomputing}, month = {2}, pages = {128993}, shortjournal = {Neurocomputing}, title = {Mobile crowdsourcing based on 5G and 6G: A survey}, volume = {618}, year = {2025}, }
• (2025). Rethinking sketch-based 3D shape retrieval: A simple baseline and benchmark reconstruction. NEUCOM, 618, 128990. (www)

  @article{ , abstract = {Sketch-based 3D shape retrieval has been receiving more and more attention. However, most previous works have two important problems: complex pipelines of their methods and unreasonable data split of the public benchmarks. The growing complexity of previous works leads to the difficulties in analyzing and following them. The unreasonable benchmarks cause the information leak of the evaluation of algorithms. In this paper, our work aims to addresses the aforementioned two problems for sketch-based 3D shape retrieval. First, contrary to the previous complex approaches, we provide a simple and effective baseline method based on prototype learning. It outperforms all the previous methods by a clear margin, indicating that the complexity in existing literature might be excessive. Second, leveraging our method for analysis, we demonstrate that the current benchmark is problematic in its data split, which impairs its validity for evaluation. Accordingly, we reconstruct the benchmark via more rigorous data splitting schemes based on zero-shot learning task. It shows that our prototype learning-based method shows more significant advantages over previous methods when facing unseen classes. Comprehensive experimental results validate our two contributions. We hope this work could benefit future works in this field.}, archive = {J_NEUCOM}, author = {Shuang Liang and Weidong Dai and Changmao Cheng and Yichen Wei and Yiyang Cai}, doi = {10.1016/j.neucom.2024.128990}, journal = {Neurocomputing}, month = {2}, pages = {128990}, shortjournal = {Neurocomputing}, title = {Rethinking sketch-based 3D shape retrieval: A simple baseline and benchmark reconstruction}, volume = {618}, year = {2025}, }
• (2025). Deep feature blend attention: A new frontier in super resolution image generation. NEUCOM, 618, 128989. (www)

  @article{ , abstract = {Super Resolution (SR) images contain more useful information than Low Resolution (LR) images. Generally, SR images are highly preferred over LR images in the medical field due to their high quality. Generally, SR images are affected by many factors, like blur, noise, and decimation. Therefore, the present study proposes Deep Feature Blend Attention Mechanism for generating SR images with optimal outcomes. In this study, a Deep Learning (DL) based encoder-decoder is used to extract the detailed information from the LR image, assist in noise removal, and improve image quality. Using Gaussian blur, an LR image is produced, which is given as input to the encoder-decoder. The Attention mechanism with feature blending is performed to produce the reconstructed SR image. The low and high frequency component issues from feature map are improved using the attention mechanism. Also, feature blend attention mechanism is implied to reconstruct SR images from LR images. The present study utilizes the feature blend technique, which helps select the optimal features and blending to attain the optimal feature set. The study is performed with and without the feature blend technique in attention mechanism to expose the impact of the feature blend technique, and it is used to avoid overfitting issues and produce optimal results. The performance of the proposed system is assessed by using PSNR and SSIM to show the system's efficiency. The proposed system is compared with other state-of-the-art studies to expose the efficacy of the present system.}, archive = {J_NEUCOM}, author = {P.B. Dhanusha and A. Muthukumar and A. Lakshmi}, doi = {10.1016/j.neucom.2024.128989}, journal = {Neurocomputing}, month = {2}, pages = {128989}, shortjournal = {Neurocomputing}, title = {Deep feature blend attention: A new frontier in super resolution image generation}, volume = {618}, year = {2025}, }
• (2025). Dual-level constraint based distributed graph convolution network for multimodal emotion recognition in conversation. NEUCOM, 618, 128937. (www)

  @article{ , abstract = {Multimodal emotion recognition in conversation (MERC) has attracted increasing attention for its wide applications. While previous methods have focused on extracting and merging features from various modalities, they often do not sufficiently consider interactions within and across modalities, as well as the semantic and modal correlations of utterances within a dialogue. To address this gap, we propose a dual-level constraint-based distributed GCN model. This model constructs a graph convolution network by simultaneously incorporating multiple modal-central graphs to capture the multimodal features of utterances. Additionally, we adopt a dual-level constraint strategy to identify the semantic and modal correlation of each speaker’s utterances in a dialogue. By comprehensively capturing the utterance features, our model significantly enhances emotion classification performances. The comparative experiment results show that our model outperforms the state-of-the-art models on the benchmark datasets IEMOCAP, MELD and M3ED. In addition, we verify the effectiveness of each component of our model by ablation experiments.}, archive = {J_NEUCOM}, author = {Yan Xiang and Lu Wang and Xiaocong Tan and Junjun Guo}, doi = {10.1016/j.neucom.2024.128937}, journal = {Neurocomputing}, month = {2}, pages = {128937}, shortjournal = {Neurocomputing}, title = {Dual-level constraint based distributed graph convolution network for multimodal emotion recognition in conversation}, volume = {618}, year = {2025}, }
• (2025). ConvSeq-MF: Convo-sequential matrix factorization for recommender system. NEUCOM, 618, 128932. (www)

  @article{ , abstract = {In the domain of recommendation, the fusion of user historical preferences and item features has emerged as a promising field to enhance recommendation accuracy and mitigate the sparsity problem. This research presents a novel approach to recommendation systems by integrating both user and item content auxiliary information. By harnessing the rich contextual data associated with users and items, our hybrid approach ConvSeq-MF endeavors to provide more personalized and accurate recommendations. As the name indicates, it is based on a matrix factorization approach driven by joint venture of semantic and contextual information of both user’s and item’s. Series of experimentation on three real-world datasets shows the suggested methodology’s practicality and efficacy over other baselines. To the best of our knowledge, the data demonstrate that the suggested model works better than other state of the art frameworks.}, archive = {J_NEUCOM}, author = {Zeeshan Khan and Zafran Khan and Naima Iltaf}, doi = {10.1016/j.neucom.2024.128932}, journal = {Neurocomputing}, month = {2}, pages = {128932}, shortjournal = {Neurocomputing}, title = {ConvSeq-MF: Convo-sequential matrix factorization for recommender system}, volume = {618}, year = {2025}, }
• (2025). Physiological signal analysis using explainable artificial intelligence: A systematic review. NEUCOM, 618, 128920. (www)

  @article{ , abstract = {With the continuous development of wearable sensors, it has become increasingly convenient to collect various physiological signals from the human body. The combination of Artificial Intelligence (AI) technology and various physiological signals has significantly improved people’s awareness of their psychological and physiological states, thus promoting substantial progress in the medical and health industries. However, most current research on physiological signal modeling does not consider the issue of interpretability, which poses a significant challenge for clinical diagnosis and treatment support. Interpretability refers to the explanation of the internal workings of AI models when generating decision results and is regarded as an important foundation for understanding model operations. Despite substantial progress made in this field in recent years, there remains a lack of systematic discussion regarding interpretable AI in physiological signal modeling, resulting in researchers having difficulty comprehensively grasping the latest developments and emerging trends in the field. Therefore, this paper provides a systematic review of interpretable AI technologies in the domain of physiological signals. Based on the scope of interpretability, these technologies are divided into two categories: global and local interpretability, and we conduct an in-depth analysis and comparison of these two types of technologies. Subsequently, we explore the potential applications of interpretable physiological signal modeling in areas such as medicine and healthcare. Finally, we summarize the key challenges of interpretable AI in the context of physiological signals and discuss future research directions. This study aims to provide researchers with a systematic framework to better understand and apply interpretable AI technologies and lay the foundation for future research.}, archive = {J_NEUCOM}, author = {Jian Shen and Jinwen Wu and Huajian Liang and Zeguang Zhao and Kunlin Li and Kexin Zhu and Kang Wang and Yu Ma and Wenbo Hu and Chenxu Guo and Yanan Zhang and Bin Hu}, doi = {10.1016/j.neucom.2024.128920}, journal = {Neurocomputing}, month = {2}, pages = {128920}, shortjournal = {Neurocomputing}, title = {Physiological signal analysis using explainable artificial intelligence: A systematic review}, volume = {618}, year = {2025}, }
• (2025). MAGO: Multi-knowledge aware and global strategy sequence optimizing network for emotional support conversation. NEUCOM, 618, 128888. (www)

  @article{ , abstract = {Emotional Support Conversation (ESConv) task aims to mitigate the emotional distress of help-seekers by providing supportive responses, which is facilitated by two main sub-tasks: selecting an appropriate supportive strategy and generating effective supportive responses to the help-seeker. However, the existing methodologies exhibit two primary shortcomings. Firstly, they primarily utilize commonsense knowledge such as using the COMET , ignoring the conceptual facts which can enhance the user-utterance understanding and improve the supportive response generation quality. Secondly, they rely excessively on dialogue history or user emotional feedbacks for strategy selection, neglecting the broader characteristics and interrelationships among various strategies globally. In this paper, we introduce the Multi-Knowledge Aware and Global Strategy Sequence Optimizing Network (MAGO). MAGO employs a Multi-Knowledge Aware Integrator to enrich the context understanding and improve response generation from the perspective of commonsense knowledge and conceptual facts. Additionally, MAGO globally incorporates the estimation of optimal strategy tag sequence by using a Strategy-Constrained Conditional Random Field (SCRF). Experimental results demonstrate MAGO’s superior performance in both strategy prediction (significantly increasing accuracy by 18%) and response generation, and analyses illustrate the importance of selecting strategies globally.}, archive = {J_NEUCOM}, author = {Qijun Xie and Wei Peng}, doi = {10.1016/j.neucom.2024.128888}, journal = {Neurocomputing}, month = {2}, pages = {128888}, shortjournal = {Neurocomputing}, title = {MAGO: Multi-knowledge aware and global strategy sequence optimizing network for emotional support conversation}, volume = {618}, year = {2025}, }
• (2025). Controllable instance synthesis with hierarchical regularization for semi-supervised pedestrian detection. NEUCOM, 618, 128831. (www)

  @article{ , abstract = {Generative Adversarial Network (GAN)-based pedestrian image synthesis opens up the possibility of using synthesized data with sufficient diversity to cover unseen variations, which facilitates semi-supervised pedestrian detection. To improve synthesized image quality and diversity in the semi-supervised setting, we propose a Hierarchical Regularized GAN (HiR-GAN), which allows a controllable instance generation process. In the proposed model, a pedestrian instance is synthesized and stitched on top of a background image. By imposing shape-based regularization on intermediate results and complete pedestrian images, a generator learns to disentangle pedestrian shape and appearance without any supervision. As a result, we achieve effective control of background, shape, and appearance via explicit latent codes. Further, the generator competes with two discriminators, which focus more on pedestrian shape and image quality, respectively. By jointly training a background-pedestrian classifier, the generator is guided to capture more precise semantics of real pedestrian data, while at the same time the classifier is able to produce more accurate annotations of unlabeled data. Experiment results on multiple benchmarks demonstrate the superiority of HiR-GAN over previous state-of-the-arts in semi-supervised pedestrian image synthesis and the effectiveness of high-fidelity synthesized data in reducing the dependence on labeled data in the pedestrian detection task.}, archive = {J_NEUCOM}, author = {Tianyi Chen and Gaozhe Li and Xiangyu Sai and Lianxin Xie and Yunfei Zhang and Si Wu}, doi = {10.1016/j.neucom.2024.128831}, journal = {Neurocomputing}, month = {2}, pages = {128831}, shortjournal = {Neurocomputing}, title = {Controllable instance synthesis with hierarchical regularization for semi-supervised pedestrian detection}, volume = {618}, year = {2025}, }
• (2025). Defender of privacy and fairness: Tiny but reversible generative model via mutually collaborative knowledge distillation. NEUCOM, 618, 128822. (www)

  @article{ , abstract = {Sharing vast amounts of data to train powerful artificial intelligence (AI) models raises public interest concerns such as privacy and fairness. While reversible anonymization techniques are very effective for privacy preservation and fairness enhancement, these methods rely on heavy reversible generative models, making them only suitable to run in the cloud or on a server independent from the image source. For example, data transmission might be under the privacy threats such as channel eavesdropping. Therefore, we propose a novel mutually collaborative knowledge distillation strategy to train a tiny and reversible generative model. This enables us to build a synthesis-based privacy and fairness protection system in embedded devices for anonymizing privacy-sensitive data and thus improve security protection capabilities from the source. The proposed mutually collaborative knowledge distillation method exploits the reversibility of the generative model. By pairing the teacher encoder (decoder) with the student decoder (encoder), we train the student decoder (encoder) by reconstructing the image space (latent space) from the prior image space (latent space). This results in tiny-size student models that can be embedded into devices. We deploy and evaluate our system on NVIDIA Jetson TX2 devices, which operate in real-time. Extensive experiments demonstrate that our system effectively anonymizes face images and thus protects privacy and also improves fairness while minimizing the impact on downstream tasks. Our code will be publicly available on GitHub.}, archive = {J_NEUCOM}, author = {Sissi Xiaoxiao Wu and Zehong Huang and Zhicong Liang and Lin Gu and Tatsuya Harada and Zheng Li and Yingying Zhu}, doi = {10.1016/j.neucom.2024.128822}, journal = {Neurocomputing}, month = {2}, pages = {128822}, shortjournal = {Neurocomputing}, title = {Defender of privacy and fairness: Tiny but reversible generative model via mutually collaborative knowledge distillation}, volume = {618}, year = {2025}, }
• (2025). RLBoost: Boosting supervised models using deep reinforcement learning. NEUCOM, 618, 128815. (www)

  @article{ , abstract = {Data quality or evaluation can sometimes be a task as important as collecting a large volume of data when it comes to generating accurate artificial intelligence models. Being able to evaluate the data can lead to a larger database that is better suited to a particular problem because we have the ability to filter out data of dubious quality obtained automatically. In this paper, we present RLBoost (Anguiano et al., 2024) [1] , an algorithm that uses deep reinforcement learning strategies to evaluate a particular dataset and obtain a model capable of estimating the quality of any new data in order to improve the final predictive quality of a supervised learning model. This solution has the advantage of being agnostic regarding the supervised model used and, through multi-attention strategies, takes into account the data in its context and not only individually. The results of the article show that this model obtains better and more stable results than other state-of-the-art algorithms such as Leave One Out, Shapley or Data Valuation using Reinforcement Learning.}, archive = {J_NEUCOM}, author = {Eloy Anguiano Batanero and Ángela Fernández Pascual and Álvaro Barbero Jiménez}, doi = {10.1016/j.neucom.2024.128815}, journal = {Neurocomputing}, month = {2}, pages = {128815}, shortjournal = {Neurocomputing}, title = {RLBoost: Boosting supervised models using deep reinforcement learning}, volume = {618}, year = {2025}, }
• (2025). DESRGAN: Detail-enhanced generative adversarial networks for small sample single image super-resolution. NEUCOM, 617, 129121. (www)

  @article{ , abstract = {Image super-resolution involves reconstructing a blurry, low-resolution image with limited information into a clear, high-resolution image containing more detailed information. The images generated by super-resolution reconstruction can enhance the performance of downstream computer vision tasks, and hold wide application prospects in fields such as industrial fault detection, plant phenotype parameter extraction, medical imaging, and more. High-frequency components in images, such as edges and texture details, typically require more attention. However, when the training samples are limited, effectively recovering clear high-frequency details of images becomes highly challenging. Therefore, this paper proposes a single-image super-resolution method based on generative adversarial networks, named DESRGAN. Compared to existing methods, DESRGAN achieves better reconstruction of image details even with a limited number of training samples. DESRGAN introduces several key innovations: a shallow generator structure to address overfitting issues in small sample scenarios, a dual-stream feature extraction network with dilated convolutions to capture multi-scale contextual information and expand the receptive field, and an artifact loss designed to eliminate artifacts and preserve the true high-frequency details of the super-resolved images. Extensive ablation experiments and comparative studies with multiple state-of-the-art models are conducted on two small sample datasets, "Root" and "Leaves," as well as five publicly available datasets. The results demonstrate that the proposed DESRGAN achieves superior performance in small sample single image super-resolution tasks, with improvements of 1.39 dB in PSNR and 0.013 in SSIM. The generated high-resolution images exhibit clear texture and edge structures, presenting favorable subjective visual effects. Moreover, the model displays strong generalization capabilities.}, archive = {J_NEUCOM}, author = {Congcong Ma and Jiaqi Mi and Wanlin Gao and Sha Tao}, doi = {10.1016/j.neucom.2024.129121}, journal = {Neurocomputing}, month = {2}, pages = {129121}, shortjournal = {Neurocomputing}, title = {DESRGAN: Detail-enhanced generative adversarial networks for small sample single image super-resolution}, volume = {617}, year = {2025}, }
• (2025). Agricultural surface water extraction in environmental remote sensing: A novel semantic segmentation model emphasizing contextual information enhancement and foreground detail attention. NEUCOM, 617, 129110. (www)

  @article{ , abstract = {The intelligent and efficient analysis of agricultural surface water is essential for agricultural irrigation, flood prevention, and resource utilization. However, accurate extraction from high-resolution optical remote sensing images is challenging due to agricultural heterogeneous terrains, dynamic water body scales, and diverse surrounding vegetation. To address these challenges, this paper introduces a novel architecture called the Complex Rural Water Extraction Network (CRWENet). The CRWENet is designed to enhance the understanding of contextual information and suppress random background noise, which comprises four key components including feature encoder, feature decoder, context information enhancement module, and foreground detail attention module. The context information enhancement module harnesses the long-range dependency-capturing capabilities of Transformer, merging multi-scale features to enhance global information perception. Meanwhile, the foreground detail attention module facilitates adaptive suppression of intricate background noise and accurate delineation of agricultural surface water boundaries by harmonizing interactions across channel, height, and width dimensions. Extensive experimentation on the LoveDA dataset and the GLH-water dataset validates the effectiveness of CRWENet. The results demonstrate that the proposed CRWENet outperforms the existing classical semantic segmentation networks and the mainstream water body extraction methods, achieving Intersection over Union (IoU) scores of 70.93 % and 80.55 % respectively.}, archive = {J_NEUCOM}, author = {Pengyu Lei and Jizheng Yi and Sijia Li and Yasheng Li and Hui Lin}, doi = {10.1016/j.neucom.2024.129110}, journal = {Neurocomputing}, month = {2}, pages = {129110}, shortjournal = {Neurocomputing}, title = {Agricultural surface water extraction in environmental remote sensing: A novel semantic segmentation model emphasizing contextual information enhancement and foreground detail attention}, volume = {617}, year = {2025}, }
• (2025). An enhanced few-shot text classification approach by integrating topic modeling and prompt-tuning. NEUCOM, 617, 129082. (www)

  @article{ , abstract = {The prompt-tuning method demonstrates remarkable effectiveness in few-shot text classification tasks. However, the performance of the method depends to some extent on a well-constructed label word space. In this paper, we propose a topic and prompt-tuning approach (TPT) for few-shot text classification tasks. Our method leverages a topic model with prior knowledge to expand label words, and incorporates label words into the verbalizer to enhance its performance, enabling twenty label words to encompass the vast majority of text classes. Experimental results on five public datasets demonstrate that the proposed TPT approach exhibits more stable performance compared to existing prompt-tuning methods.}, archive = {J_NEUCOM}, author = {Yinghui Zhang and Yichun Xu and Fangmin Dong}, doi = {10.1016/j.neucom.2024.129082}, journal = {Neurocomputing}, month = {2}, pages = {129082}, shortjournal = {Neurocomputing}, title = {An enhanced few-shot text classification approach by integrating topic modeling and prompt-tuning}, volume = {617}, year = {2025}, }
• (2025). A user behavior-aware multi-task learning model for enhanced short video recommendation. NEUCOM, 617, 129076. (www)

  @article{ , abstract = {In the rapidly evolving landscape of digital media consumption, accurately predicting user preferences and behaviors is critical for the effectiveness of recommendation systems, especially for short video content. Traditional recommendation methods often ignore the association between multiple user behavior types and struggle with dynamically adapting to user behavior changes, leading to suboptimal personalization and user engagement. To address these issues, this paper introduces a user behavior-aware multi-task learning model for enhanced short video recommendation (UBA-SVR) by leveraging insights into dynamic user interactions. In our approach, we construct a user behavior-aware transformer to comprehensively capture users’ dynamic interests and generate the fusion feature representation. Subsequently, we introduce a hierarchical knowledge extraction framework to process features in multi-stage and adopt a task-aware attention mechanism within the tower network structure to facilitate effective information sharing and differentiation among tasks. Furthermore, we employ a dynamic joint loss optimization strategy to adaptively adjust the loss weights for different tasks to promote collaborative enhancement. Extensive experiments on two real-world datasets demonstrate that the proposed method achieves significant improvements in multiple prediction tasks simultaneously.}, archive = {J_NEUCOM}, author = {Yuewei Wu and Ruiling Fu and Tongtong Xing and Zhenyu Yu and Fulian Yin}, doi = {10.1016/j.neucom.2024.129076}, journal = {Neurocomputing}, month = {2}, pages = {129076}, shortjournal = {Neurocomputing}, title = {A user behavior-aware multi-task learning model for enhanced short video recommendation}, volume = {617}, year = {2025}, }
• (2025). Fixed/preassigned-time non-chattering synchronization of nonlinear coupled Cohen–Grossberg neural networks via event-triggered control. NEUCOM, 617, 129069. (www)

  @article{ , abstract = {In this paper, the studies on fixed-time and preassigned-time synchronization for a kind of nonlinear coupled Cohen–Grossberg neural networks (CGNNs) are researched by event-triggered non-chattering control. Firstly, differing from some existed fixed-time synchronization control strategies, a power-law feedback controller without sign function is adopted to achieve non-chattering control effect. To save communication and computational resources, an event-triggered mechanism (ETM) is introduced, which can naturally avoid Zeno behavior. By designing preassigned time parameters as controller parameters, the preassigned-time synchronization of dynamic networks is studied. Then, based on the designed control mechanism, sufficient conditions for fixed-time and preassigned-time synchronization of the nonlinear coupled CGNNs are given. Finally, the validity and generality of the obtained results are verified by numerical simulation.}, archive = {J_NEUCOM}, author = {Shuang Qing and Huamin Wang and Shiping Wen}, doi = {10.1016/j.neucom.2024.129069}, journal = {Neurocomputing}, month = {2}, pages = {129069}, shortjournal = {Neurocomputing}, title = {Fixed/preassigned-time non-chattering synchronization of nonlinear coupled Cohen–Grossberg neural networks via event-triggered control}, volume = {617}, year = {2025}, }
• (2025). Integrated adaptive communication in multi-agent systems: Dynamic topology, frequency, and content optimization for efficient collaboration. NEUCOM, 617, 129068. (www)

  @article{ , abstract = {In multi-agent systems (MAS), effective communication is essential for coordination and achieving common goals, especially in complex environments. However, existing communication methods face significant challenges, such as high resource consumption and limited adaptability to dynamic environments. To address these, we propose the integrated adaptive communication network (IACN) built on multi-agent proximal policy optimization (MAPPO), which enhances communication efficiency and adaptability in MAS. IACN dynamically adjusts communication topology using a learnable graph and optimizes content based on task relevance. Additionally, it incorporates an adaptive frequency adjustment mechanism to balance communication demands based on task urgency and environmental changes. Experiments in multi-agent particle environments demonstrate that IACN significantly outperforms existing methods in terms of overall performance, coordination effectiveness, and adaptability.}, archive = {J_NEUCOM}, author = {Jianrui Wang and Yi Li and Yitian Hong and Yang Tang}, doi = {10.1016/j.neucom.2024.129068}, journal = {Neurocomputing}, month = {2}, pages = {129068}, shortjournal = {Neurocomputing}, title = {Integrated adaptive communication in multi-agent systems: Dynamic topology, frequency, and content optimization for efficient collaboration}, volume = {617}, year = {2025}, }
• (2025). Transformer-guided feature pyramid network for multi-view stereo. NEUCOM, 617, 129066. (www)

  @article{ , abstract = {Feature Pyramid Network (FPN) is widely used in Multi-View Stereo (MVS) to extract multi-scale features, effectively enhancing both the quality and efficiency of reconstruction. However, the multi-scale features are simply fused by element-wise addition, ignoring the contextual relationship between them. To solve this problem, a Transformer-guided Feature Pyramid Network (TFPN-MVSNet) is proposed, which uses the self-attention mechanism to aggregate the long-range context information between multi-scale features. Firstly, the feature enhancement module is designed based on internal attention mechanism to calculate the importance of different parts within the feature sequence, which pays more attention to the features related to reconstruction. The feature aggregation module is designed based on cross attention mechanism to focus on the contextual information of multi-scale features. Secondly, the two modules are introduced into the multi-scale feature extraction network to enhance the three-stage features from coarse to fine. Finally, the Spatial Weighted Cost Volume Aggregation method (SWCA)uses the spatial attention mechanism to calculate the pixel weights at the same position in different views, which suppresses the negative impact of occluded pixels. Evaluated on the DTU dataset, the overall error is reduced by 0.062 mm compared to CasMVSNet, and by 0.012 mm and 0.019 mm compared to similar methods TransMVSNet and CostFormer, respectively. Our method also achieves competitive results on the Tanks & Temples and ETH3D datasets.}, archive = {J_NEUCOM}, author = {Lina Wang and Jiangfeng She and Zhao Qiang and Xiang Wen and Yuzheng Guan}, doi = {10.1016/j.neucom.2024.129066}, journal = {Neurocomputing}, month = {2}, pages = {129066}, shortjournal = {Neurocomputing}, title = {Transformer-guided feature pyramid network for multi-view stereo}, volume = {617}, year = {2025}, }
• (2025). Promoting camouflaged object detection through novel edge–target interaction and frequency-spatial fusion. NEUCOM, 617, 129064. (www)

  @article{ , abstract = {The realm of camouflage object detection endeavors to precisely discern concealed targets that blend seamlessly into their surroundings. Recently, numerous research endeavors have converged on unraveling the disguise tactics employed by such objects, striving for efficient and accurate detection results. While these methods have garnered notable advancements in unveiling camouflaged targets, they still grapple with challenges posed by the high similarity between objects and their backgrounds, resulting in misidentifications, detection oversights, and the loss of intricate details. In light of these challenges, this paper introduces an innovative EF Net, aimed at elevating the integrity of target recognition and the finesse of detail capture through deepened interactions between E dge–target features and F requency-spatial information. Notably, recognizing that the edges of camouflaged objects harbor a wealth of intrinsic details and serve as crucial delimiters distinguishing them from their surroundings, we have devised a dual-path architecture, where an auxiliary Edge Detection (ED) branch collaborates seamlessly with the primary Object Segmentation (OS) branch. Within the OS branch, we meticulously crafted the Edge-induced Segmentation Refinement module (ESR), which ingeniously harnesses the refined edge information provided by the ED branch to bolster target segmentation accuracy and augment detail fidelity. Moreover, in the supporting ED branch, we propose the Structure-induced Edge Enhancement module (SEE), designed to accentuate edge information while suppressing irrelevant noise with the assistance of the OS branch, thereby ensuring the precise extraction of edge information. To further enrich the information on dual-path, we innovatively introduce a Frequency Decomposition Module (FDM), which decomposes images into low- and high-frequency components, tailored to enhance target segmentation and edge detection, respectively. Additionally, to ensure seamless integration of frequency-domain and spatial-domain information, we devised the versatile Frequency-Spatial Domain Mixer (FSDM), achieving precise information alignment and fusion of these two domains. Quantitative and qualitative experimental results demonstrate that our proposed EFNet significantly outperforms existing state-of-the-art methods on four widely used benchmark datasets.}, archive = {J_NEUCOM}, author = {Juwei Guan and Weiqi Qian and Tongxin Zhu and Xiaolin Fang}, doi = {10.1016/j.neucom.2024.129064}, journal = {Neurocomputing}, month = {2}, pages = {129064}, shortjournal = {Neurocomputing}, title = {Promoting camouflaged object detection through novel edge–target interaction and frequency-spatial fusion}, volume = {617}, year = {2025}, }
• (2025). Enhancing federated learning-based social recommendations with graph attention networks. NEUCOM, 617, 129045. (www)

  @article{ , abstract = {Social recommendations are crucial in information services, particularly for mitigating information overload on the Internet. Graph Attention Networks (GATs) have emerged as an effective solution for the social recommendation problem, harnessing the benefits of representation learning from user–item and user–user graphs. However, traditional GAT-based social recommendation models often necessitate the centralized storage and aggregation of all the users’ social connections and item interactions. This centralization raises significant privacy concerns. To address these issues, we propose a federated learning-based social recommendation system using GATs. Specifically, we first design a structured-enhanced graph clustering algorithm to obtain soft clustering assignments for items. Based on these assignments, the similarity between pairwise items and their corresponding correlations are computed. Second, we employ graph attention networks to infer the local heterogeneity of social relationships, user–item interactions, and item relationships within the federated learning framework for each client. Using the soft clustering assignments, we derive hard clustering indicators and sample pseudo items. Local gradients of embedding and model parameters are then computed and uploaded to the central server for aggregation under the local differential privacy mechanism. Experimental results over real-world data sets indicate that the proposed model is more effective in privacy protection while maintaining acceptable performance compared with the central learning framework. Additionally, the proposed model is more stable in addressing the cold-start issue for items than state-of-the-art federated learning-based social recommendation models.}, archive = {J_NEUCOM}, author = {Zhihui Xu and Bing Li and Wenming Cao}, doi = {10.1016/j.neucom.2024.129045}, journal = {Neurocomputing}, month = {2}, pages = {129045}, shortjournal = {Neurocomputing}, title = {Enhancing federated learning-based social recommendations with graph attention networks}, volume = {617}, year = {2025}, }
• (2025). Overcoming the challenges of multi-modal medical image sharing: A novel data distillation strategy via contrastive learning. NEUCOM, 617, 129043. (www)

  @article{ , abstract = {Multi-modal medical images serve as a critical basis for disease diagnosis. However, sharing multi-modal medical data between hospitals is a significant challenge due to privacy protection concerns as well as the substantial costs associated with data transfer and storage. These hurdles are further compounded by two primary issues: (1) the complexity involved in sharing data features, which arises from the intricate process of multi-modal image registration; (2) the difficulty and high cost of acquiring accurate data labels. To address these issues, we propose a novel multi-modal medical dataset distillation method based on contrastive learning. We consider each modality as a unique approach to data augmentation for the corresponding imaging region, offering a unique perspective on the same anatomical region, which is then fused to collectively contribute to a comprehensive global representation of the medical image. Validation results on two datasets indicate our method is capable of reducing the original dataset to just 5% of its original size while retaining classification performance comparable to that of the full dataset, and notably, this efficiency is achieved using only 30% of the total labels, which significantly lowers the cost and effort associated with dataset labeling.}, archive = {J_NEUCOM}, author = {Taoli Du and Wenhui Li and Zeyu Wang and Feiyang Yang and Peihong Teng and Xingcheng Yi and Hongyu Chen and Zixuan Wang and Ping Zhang and Tianyang Zhang}, doi = {10.1016/j.neucom.2024.129043}, journal = {Neurocomputing}, month = {2}, pages = {129043}, shortjournal = {Neurocomputing}, title = {Overcoming the challenges of multi-modal medical image sharing: A novel data distillation strategy via contrastive learning}, volume = {617}, year = {2025}, }
• (2025). Low light recognition of traffic police gestures based on lightweight extraction of skeleton features. NEUCOM, 617, 129042. (www)

  @article{ , abstract = {With the increasing demand for information interaction between intelligent vehicles and traffic managers, understanding traffic police commands has become an essential element in achieving human–computer interaction. The aim of this study is to investigate the low light traffic police gesture recognition based on lightweight extraction of skeleton features. A two-stage recognition framework for eight gesture commands is proposed. The first stage utilizes convolutional neural networks to acquire the human skeleton under low light. The second stage utilizes a recurrent neural network to construct a lightweight feature extraction model for gesture recognition. HRnet acquires spatial features of the action, high-resolution and multi-scale fusion circumvents the effect of low-light intensity. Self-attention (SA) mechanism and MobileNet lightweight feature extraction have addressed the time-consuming issue in the feature extraction phase. The improved GRU network is used to extract time domain features, the fusion of spatio-temporal features improves the recognition accuracy. For training and validation, a gesture dataset (TPGD) containing 3391 gesture instances is constructed. The proposed method obtains 94.75% accuracy in the TPGD test set, demonstrating its effectiveness. This study holds significant importance for enhancing the level of intelligence in traffic management, promoting the development of intelligent transportation systems, and ensuring road traffic safety.}, archive = {J_NEUCOM}, author = {Mengying Chang and Huizhi Xu and Yuanming Zhang}, doi = {10.1016/j.neucom.2024.129042}, journal = {Neurocomputing}, month = {2}, pages = {129042}, shortjournal = {Neurocomputing}, title = {Low light recognition of traffic police gestures based on lightweight extraction of skeleton features}, volume = {617}, year = {2025}, }
• (2025). Monocular thermal SLAM with neural radiance fields for 3D scene reconstruction. NEUCOM, 617, 129041. (www)

  @article{ , abstract = {Visual simultaneous localization and mapping (SLAM) faces significant challenges in environments with variable lighting and smoke. Excelling in such visually degraded settings, thermal imaging captures scene radiance effectively. To address the limitations of traditional thermal SLAM in 3D scene reconstruction, we propose ThermalSLAM-NeRF, a novel integration of thermal SLAM with neural radiance fields (NeRF). This method significantly enhances the quality of high dynamic range thermal images by improving their signal-to-noise ratio, contrast, and detail. It also employs online photometric calibration to ensure grayscale consistency between consecutive frames. We utilize a sparse direct method for pose estimation, selecting keyframes based on photometric error and tracking quality. The NeRF map is reconstructed using a multi-view keyframe sequence. Our evaluations on datasets containing over 15,000 thermal images show that ThermalSLAM-NeRF achieves an average improvement of 59.30% in trajectory accuracy over existing state-of-the-art SLAM methods. This approach uniquely tracks all sequences and constructs comprehensive NeRF maps, enabling robust and precise pose estimation without the need for extensive pre-training.}, archive = {J_NEUCOM}, author = {Yuzhen Wu and Lingxue Wang and Lian Zhang and Mingkun Chen and Wenqu Zhao and Dezhi Zheng and Yi Cai}, doi = {10.1016/j.neucom.2024.129041}, journal = {Neurocomputing}, month = {2}, pages = {129041}, shortjournal = {Neurocomputing}, title = {Monocular thermal SLAM with neural radiance fields for 3D scene reconstruction}, volume = {617}, year = {2025}, }
• (2025). MSEB: Plug and play multi-scale image embedding block for vision backbone. NEUCOM, 617, 129040. (www)

  @article{ , abstract = {Mainstream backbone networks in computer vision, leveraging sophisticated CNN or Transformer architectures, have achieved remarkable performance across various tasks. However, these works primarily focus on downsample ranges of 4 × to 32 × on input images, with limited exploration into the embedding portion of the network. In this paper, we propose a novel Multi-Scale Embedding Block (MSEB) to enhance embedding features. MSEB is constructed through a combination of Self-Calibrated Channel Dilate Unit (SCDU) and Multi-Scale Encoder Unit (MSEU). SCDU projects images into high-dimensional space, incorporating channel self-calibration to preserve fine-grained information in shallow feature maps, thereby providing a more comprehensive representation for subsequent network layers. MSEU employs a multi-scale cascade design with implicit feature sharing characteristics, expanding the receptive field scale range to enhance embedding representation. Extensive experiments on mainstream benchmarks demonstrate that our proposed multi-scale embedding block can be seamlessly integrated into most popular CNN and Transformer architectures such as ResNet, Res2Net, Swin-Transformer, and MViT. By replacing the existing embedding structure in these methods, significant improvements on ImageNet-1k datasets are observed, e.g., CNN-based ResNet-50 and further adopting a multi-scale design Res2Net-50 achieve Top-1 accuracy improvements of 0.47% and 1.22%, respectively. Attention-based Swin-Tiny and MViT-Tiny, both adopting multi-scale design, achieved Top-1 accuracy improvements of 2.57% and 1.05%, respectively.}, archive = {J_NEUCOM}, author = {Hao Yuan and Bin Zhang and Yachuan Wang}, doi = {10.1016/j.neucom.2024.129040}, journal = {Neurocomputing}, month = {2}, pages = {129040}, shortjournal = {Neurocomputing}, title = {MSEB: Plug and play multi-scale image embedding block for vision backbone}, volume = {617}, year = {2025}, }
• (2025). STRAT: Image style transfer with region-aware transformer. NEUCOM, 617, 129039. (www)

  @article{ , abstract = {Style transfer methods render various artistic styles to a natural image through the extraction and transfer of textural features. Existing neural style transfer methods often rely on CNNs to extract image features and tend to suffer from feature leakage and content distortion due to limited receptive fields. Transformer-based style transfer methods outperform CNN-based methods by learning the global information of image through self-attention mechanism. However, local features are ignored and details are lost since the semantic information of images is not taken into account. To address this critical issue, this paper proposes a novel style transfer framework based on region-aware transformer (STRAT). We integrate the CNN based short-range branch with the transformer-based long-range branch to extract both local and non-local features to achieve region-adaptive texture transfer with two region-aware attention modules, respectively. Specifically, we utilize the SNR metric and masks as guide to propose the SNR-guided attention module and mask-guided cross attention module to enable region-varying feature extraction and adaptive texture transfer, respectively. Extensive experimental results demonstrate that our proposed method outperforms the state-of-the-arts methods in terms of subjective and objective results.}, archive = {J_NEUCOM}, author = {Na Qi and Yezi Li and Rao Fu and Qing Zhu}, doi = {10.1016/j.neucom.2024.129039}, journal = {Neurocomputing}, month = {2}, pages = {129039}, shortjournal = {Neurocomputing}, title = {STRAT: Image style transfer with region-aware transformer}, volume = {617}, year = {2025}, }
• (2025). IncARMAG: A convolutional neural network with multi-level autoregressive moving average graph convolutional processing framework for medical image classification. NEUCOM, 617, 129038. (www)

  @article{ , abstract = {Effective computer-aided detection and diagnosis algorithms are being sought to carry out complex pattern recognition with great precision while increasing efficiency and robustness through automation. In this study, we leveraged self-constructing graph (SCG) and autoregressive moving average (ARMA) graph convolution to postprocess multi-level feature maps obtained from an Inception V3 network for medical image classification. The adopted SCG learns latent embeddings from the feature maps, which are subsequently used to construct graph topology based on embedding similarity. The ARMA convolution then learns to encode more flexible spectral filter responses that simultaneously takes into account long-range dependencies from the SCG graph. The proposed IncARMAG model was evaluated on 2D MedMNIST datasets, which consist of several imaging modalities and classification tasks in the medical domain. Results from our experiments on MedMNIST datasets showed that IncARMAG outperformed state-of-the-art models on many cases. IncARMAG has also achieved the highest accuracy based on evaluation of various benchmark CNN, transformers, and hybrid models on four selected datasets. These results highlight how IncARMAG is a strong candidate in medical image analysis involving classification tasks.}, archive = {J_NEUCOM}, author = {Adrian S. Remigio}, doi = {10.1016/j.neucom.2024.129038}, journal = {Neurocomputing}, month = {2}, pages = {129038}, shortjournal = {Neurocomputing}, title = {IncARMAG: A convolutional neural network with multi-level autoregressive moving average graph convolutional processing framework for medical image classification}, volume = {617}, year = {2025}, }
• (2025). Learning a more compact representation for low-rank tensor completion. NEUCOM, 617, 129036. (www)

  @article{ , abstract = {Transform-based tensor nuclear norm (TNN) methods have gained considerable attention for their effectiveness in addressing tensor recovery challenges. The integration of deep neural networks as nonlinear transforms has been shown to significantly enhance their performance. Minimizing transform-based TNN is equivalent to minimizing the ℓ 1 norm of singular values in the transformed domain, which can be interpreted as finding a sparse representation with respect to the bases supported by singular vectors. This work aims to advance deep transform-based TNN methods by identifying a more compact representation through learnable bases, ultimately improving recovery accuracy. We specifically employ convolutional kernels as these learnable bases, demonstrating their capability to generate more compact representation, i.e., sparser coefficients of real-world tensor data compared to singular vectors. Our proposed model consists of two key components: a transform component, implemented through fully connected neural networks (FCNs), and a convolutional component that replaces traditional singular matrices. Then, this model is optimized using the ADAM algorithm directly on the incomplete tensor in a zero-shot manner, meaning all learnable parameters within the FCNs and convolution kernels are inferred solely from the observed data. Experimental results indicate that our method, with this straightforward yet effective modification, outperforms state-of-the-art approaches on video and multispectral image recovery tasks.}, archive = {J_NEUCOM}, author = {Xi-Zhuo Li and Tai-Xiang Jiang and Liqiao Yang and Guisong Liu}, doi = {10.1016/j.neucom.2024.129036}, journal = {Neurocomputing}, month = {2}, pages = {129036}, shortjournal = {Neurocomputing}, title = {Learning a more compact representation for low-rank tensor completion}, volume = {617}, year = {2025}, }
• (2025). Global span semantic dependency awareness and filtering network for nested named entity recognition. NEUCOM, 617, 129035. (www)

  @article{ , abstract = {Span-based methods for nested named entity recognition (NER) are effective in handling the complexities of nested entities with hierarchical structures. However, these methods often overlook valid semantic dependencies among global spans, resulting in a partial loss of semantic information. To address this issue, we propose the Global Span Semantic Dependency Awareness and Filtering Network (GSSDAF). Our model begins with BERT for initial sentence encoding. Following this, a span semantic representation matrix is generated using a multi-head biaffine attention mechanism. We introduce the Global Span Dependency Awareness (GSDA) module to capture valid semantic dependencies among all spans, and the Local Span Dependency Enhancement (LSDE) module to selectively enhance key local dependencies. The enhanced span semantic representation matrix is then decoded to classify the spans. We evaluated our model on seven public datasets. Experimental results demonstrate that our model effectively handles nested NER, achieving higher F1 scores compared to baselines. Ablation experiments confirm the effectiveness of each module. Further analysis indicates that our model can learn valid semantic dependencies between global spans, significantly improving the accuracy of nested entity recognition. Our code is available at https://github.com/Shaun-Wong/GSSDAF .}, archive = {J_NEUCOM}, author = {Yunlei Sun and Xiaoyang Wang and Haosheng Wu and Miao Hu}, doi = {10.1016/j.neucom.2024.129035}, journal = {Neurocomputing}, month = {2}, pages = {129035}, shortjournal = {Neurocomputing}, title = {Global span semantic dependency awareness and filtering network for nested named entity recognition}, volume = {617}, year = {2025}, }
• (2025). Towards the characterization of representations learned via capsule-based network architectures. NEUCOM, 617, 129027. (www)

  @article{ , abstract = {Capsule Neural Networks (CapsNets) have been re-introduced as a more compact and interpretable alternative to standard deep neural networks. While recent efforts have proved their compression capabilities, to date, their interpretability properties have not been fully assessed. Here, we conduct a systematic and principled study towards assessing the interpretability of these types of networks. We pay special attention towards analyzing the level to which part-whole relationships are encoded within the learned representation. Our analysis in the MNIST, SVHN, CIFAR-10, and CelebA datasets on several capsule-based architectures suggest that the representations encoded in CapsNets might not be as disentangled nor strictly related to parts-whole relationships as is commonly stated in the literature.}, archive = {J_NEUCOM}, author = {Saja Tawalbeh and José Oramas}, doi = {10.1016/j.neucom.2024.129027}, journal = {Neurocomputing}, month = {2}, pages = {129027}, shortjournal = {Neurocomputing}, title = {Towards the characterization of representations learned via capsule-based network architectures}, volume = {617}, year = {2025}, }
• (2025). Closed-loop seizure modulation via extreme learning machine supervisor based sliding mode disturbance rejection control. NEUCOM, 617, 129026. (www)

  @article{ , abstract = {Neuromodulation is a low-risk and high-efficient therapy to treat epilepsy. In clinic, there is an urgent need for a regulation strategy that is adaptable to unknown nonlinearities and strong robust to kinds of disturbances and uncertainties. Linear active disturbance rejection control (LADRC) can adapt to complex epileptic dynamics and improve the epilepsy modulation, even if little model information is available, various uncertainties and external disturbances exist. However, a proportional plus derivative controller in the LADRC is weak to resist external disturbances that are not addressed by an extended state observer. In addition, the phase delay of the input and output lowers the speed of modulation. An extreme learning machine (ELM) based supervisor can get an inversion of the plant timelier and more accurately, and an ELM supervisor based sliding mode disturbance rejection control (ESSMDRC) is proposed to improve both speed and robustness of the modulation. Closed-loop stability and the phase-leading property are analysed. Numerical results show that the proposed ESSMDRC guarantees a more satisfactory closed-loop neuromodulation.}, archive = {J_NEUCOM}, author = {Wei Wei and Zijin Wang}, doi = {10.1016/j.neucom.2024.129026}, journal = {Neurocomputing}, month = {2}, pages = {129026}, shortjournal = {Neurocomputing}, title = {Closed-loop seizure modulation via extreme learning machine supervisor based sliding mode disturbance rejection control}, volume = {617}, year = {2025}, }
• (2025). LGAT: A novel model for multivariate time series anomaly detection with improved anomaly transformer and learning graph structures. NEUCOM, 617, 129024. (www)

  @article{ , abstract = {Time series anomaly detection involves identifying data points in continuously collected datasets that deviate from normal patterns. Given that real-world systems often consist of multiple variables, detecting anomalies in multivariate datasets has become a key focus of current research. This challenge has wide-ranging applications across various industries for system maintenance, such as in water treatment and distribution networks, transportation, and autonomous vehicles, thus driving active research in the field of time series anomaly detection. However, traditional methods primarily address this issue by predicting and reconstructing input time steps, but they still suffer from problems of overgeneralization and inconsistency in providing high performance for reasoning about complex dynamics. In response, we propose a novel unsupervised model called LGAT, which can automatically learn graph structures and leverage an enhanced Anomaly Transformer architecture to capture temporal dependencies. Moreover, the model features a new encoder–decoder architecture designed to enhance context extraction capabilities. In particular, the model calculates anomaly scores for multivariate time series anomaly detection by combining the reconstruction of input time series with the model’s computed prior associations and sequential correlations. This model captures inter-variable relationships and exhibit stronger context extraction abilities, making it more sensitive to anomaly detection. Extensive experiments on six common anomaly detection benchmarks further demonstrate the superiority of our approach over other state-of-the-art methods, with an improvement of approximately 1.2% across various metrics.}, archive = {J_NEUCOM}, author = {Mi Wen and ZheHui Chen and Yun Xiong and YiChuan Zhang}, doi = {10.1016/j.neucom.2024.129024}, journal = {Neurocomputing}, month = {2}, pages = {129024}, shortjournal = {Neurocomputing}, title = {LGAT: A novel model for multivariate time series anomaly detection with improved anomaly transformer and learning graph structures}, volume = {617}, year = {2025}, }
• (2025). Distributed continuous-time algorithm for nonsmooth aggregative optimization over weight-unbalanced digraphs. NEUCOM, 617, 129022. (www)

  @article{ , abstract = {This paper studies the problem of distributed continuous-time aggregative optimization with set constraints under a weight-unbalanced digraph, where the nonsmooth objective function of each agent relies both on its own decision and on the aggregation of all agents’ decisions. To eliminate the impact of unbalanced digraphs, a consensus-based estimator that tracks the aggregation information is designed through a gradient rescaling technique. Considering that cost functions are nondifferentiable in many scenarios, such as electric power management that takes price caps into account, a novel distributed continuous-time optimization algorithm via generalized gradient is presented in a two-time scale. Moreover, the convergence of the algorithm is established through nonsmooth analysis and singular perturbation theory. Compared to the existing results, which depend on undirected graphs, the proposed strategy is applicable to general digraphs, which may be weight-unbalanced. Further, the assumption on the differentiability of objective functions is relaxed. Finally, two numerical examples are provided to verify the findings.}, archive = {J_NEUCOM}, author = {Zheng Zhang and Guang-Hong Yang}, doi = {10.1016/j.neucom.2024.129022}, journal = {Neurocomputing}, month = {2}, pages = {129022}, shortjournal = {Neurocomputing}, title = {Distributed continuous-time algorithm for nonsmooth aggregative optimization over weight-unbalanced digraphs}, volume = {617}, year = {2025}, }
• (2025). Data-specific activation function learning for constructive neural networks. NEUCOM, 617, 129020. (www)

  @article{ , abstract = {Activation functions play a crucial role in learning and expressive capabilities of advanced neural networks due to their non-linear or non-saturated properties. However, how to determine the appropriate activation function from various candidates is a challenging yet not well-addressed topic. To address the issue, a novel self-learning approach, called as data-specific activation function learning (DSAFL) algorithm, is proposed to establish constructive neural network on one-time by adaptively selecting appropriate activation function based on the specific data characteristics. To assess the space dimension mapping abilities of different activation functions, the configuration probabilities are used to guide the generation of various candidate activation functions and corresponding candidate hidden node. In the learning stage, an exploration-exploitation mechanism composed of the random algorithm and the greedy strategy is developed to obtain the influence of different candidate activation functions, thereby avoiding configuration probabilities falling into local optimum. A reward-penalty mechanism is built to update the configuration probabilities and enhance the robustness of network by integrating the simulated annealing strategy. In final, the activation function with the highest configuration probability, as the best one, is used to reconstruct the neural network. Experimental results on both regression and classification tasks demonstrate the efficiency and effectiveness of DSAFL in the activation function selection problems of a class of constructive neural networks.}, archive = {J_NEUCOM}, author = {Zhenxing Xia and Wei Dai and Xin Liu and Haijun Zhang and Xiaoping Ma}, doi = {10.1016/j.neucom.2024.129020}, journal = {Neurocomputing}, month = {2}, pages = {129020}, shortjournal = {Neurocomputing}, title = {Data-specific activation function learning for constructive neural networks}, volume = {617}, year = {2025}, }
• (2025). Unsupervised fuzzy temporal knowledge graph entity alignment via joint fuzzy semantics learning and global structure learning. NEUCOM, 617, 129019. (www)

  @article{ , abstract = {Temporal Knowledge Graph Entity Alignment (TKGEA) aims to identify the equivalent entities between different Temporal Knowledge Graphs (TKGs), which is important to knowledge fusion. The current mainstream TKGEA models are supervised embedding-based models that rely on pre-aligned seeds and implicitly encode structural information into entity embedding space for identifying equivalent entities. To deal with the TKGs structural information, some models use Graph Neural Network (GNN) encoding. But they ignore the design of decoders, failing to fully leverage the TKGs structural information. In addition, they primarily focus on crisp TKGs with clear entity semantics. However, many real-world TKGs exhibit fuzzy semantics. This fuzzy information makes existing TKGEA models face the challenge of handling the fuzzy semantics when aligning the equivalent fuzzy entities. To solve the above problems, we propose a novel unsupervised F uzzy T emporal Knowledge Graphs Entity Alignment (EA) framework that jointly performs F uzzy Semantics Learning and Global S tructure Learning, namely FTFS. In this framework, we convert the EA task into an unsupervised optimal transport task between two intra-graph matrices, eliminating the necessity for pre-aligned seeds and thereby avoiding intensive labor. Since we further consider the relation between graph structure and entities during the optimal-transport-based decoder module, it can make better use of the global structural information rather than simply encoding it implicitly into the embedding space. Moreover, unlike TKGEA models, which use binary classification to represent temporal relational facts, we introduce fuzzy semantics learning to embed membership degrees of fuzzy temporal relational facts. Extensive experiments on five FTKG datasets show that our unsupervised method is superior to the state-of-the-art EA methods.}, archive = {J_NEUCOM}, author = {Jingni Song and Luyi Bai and Xuanxuan An and Longlong Zhou}, doi = {10.1016/j.neucom.2024.129019}, journal = {Neurocomputing}, month = {2}, pages = {129019}, shortjournal = {Neurocomputing}, title = {Unsupervised fuzzy temporal knowledge graph entity alignment via joint fuzzy semantics learning and global structure learning}, volume = {617}, year = {2025}, }
• (2025). An adaptation of hybrid binary optimization algorithms for medical image feature selection in neural network for classification of breast cancer. NEUCOM, 617, 129018. (www)

  @article{ , abstract = {The performance of neural network is largely dependent on their capability to extract very discriminant features supporting the characterization of abnormalities in the medical image. Several benchmark architectures have been proposed and the use of transfer learning has further made these architectures return good performances. Study has shown that the use of optimization algorithms for selection of relevant features has improved classifiers. However continuous optimization algorithms have mostly been used though it allows variables to take value within a range of values. The advantage of binary optimization algorithms is that it allows variables to be assigned only two states, and this have been sparsely applied to medical image feature optimization. This study therefore proposes hybrid binary optimization algorithms to efficiently identify optimal features subset in medical image feature sets. The binary dwarf mongoose optimizer (BDMO) and the particle swarm optimizer (PSO) were hybridized with the binary Ebola optimization search algorithm (BEOSA) on new nested transfer functions. Medical images passed through convolutional neural networks (CNN) returns extracted features into a continuous space which are piped through these new hybrid binary optimizers. Features in continuous space a mapped into binary space for optimization, and then mapped back into the continuous space for classification. Experimentation was conducted on medical image samples using the Curated Breast Imaging Subset of Digital Database for Screening Mammography (DDSM+CBIS). Results obtained from the evaluation of the hybrid binary optimization methods showed that they yielded outstanding classification accuracy, fitness, and cost function values of 0.965, 0.021 and 0.943. To investigate the statistical significance of the hybrid binary methods, the analysis of variance (ANOVA) test was conducted based on the two-factor analysis on the classification accuracy, fitness, and cost metrics. Furthermore, results returned from application of the binary hybrid methods medical image analysis showed classification accuracy of 0.8286, precision of 0.97, recall of 0.83, and F1-score of 0.99, AUC of 0.8291. Findings from the study showed that contrary to the popular approach of using continuous metaheuristic algorithms for feature selection problem, the binary metaheuristic algorithms are well suitable for handling the challenge. Complete source code can be accessed from: https://github.com/NathanielOy/hybridBinaryAlgorithm4FeatureSelection}, archive = {J_NEUCOM}, author = {Olaide N. Oyelade and Enesi Femi Aminu and Hui Wang and Karen Rafferty}, doi = {10.1016/j.neucom.2024.129018}, journal = {Neurocomputing}, month = {2}, pages = {129018}, shortjournal = {Neurocomputing}, title = {An adaptation of hybrid binary optimization algorithms for medical image feature selection in neural network for classification of breast cancer}, volume = {617}, year = {2025}, }
• (2025). A new visual-inertial odometry scheme for unmanned systems in unified framework of zeroing neural networks. NEUCOM, 617, 129017. (www)

  @article{ , abstract = {In recent years, multi-sensor fusion has gained significant attention from researchers and is used extensively in simultaneous localization and mapping (SLAM) applications, such as visual-inertial odometry (VIO). This technology primarily utilizes visual and odometry measurements for unmanned aerial vehicles (UAVs) to estimate their position, orientation, and environment. However, in most previous works, the input error data of sensors in the system were considered independent. To improve system precision and fully utilize sensor data, a new method called Multi-State Constraint Kalman Filter with NearSAC (MSCKF-NearSAC), based on the MSCKF, is proposed. This method eliminates outliers by limiting the range of selected points, which significantly improves the success rate of feature point matching in the front-end. Furthermore, the MSCKF-ZNN method is proposed for the back-end, and combines zeroing neural network (ZNN) (originated from the Hopfield-type neural network) and error state, resulting in an exponentially converging output trajectory error, thus improving the trajectory precision of the SLAM system. The proposed algorithms, MSCKF-NearSAC and MSCKF-ZNN, are used in the excellent work of the stereo multi-state constraint Kalman filter system (S-MSCKF). A plethora of comparison experiments, utilizing precise measurement and calibration techniques, are conducted on open-source datasets and real-world environments. Experimental results demonstrate that the introduced approach exhibits higher stability in contrast to other algorithms.}, archive = {J_NEUCOM}, author = {Dechao Chen and Jianan Jiang and Zhixiong Wang and Shuai Li}, doi = {10.1016/j.neucom.2024.129017}, journal = {Neurocomputing}, month = {2}, pages = {129017}, shortjournal = {Neurocomputing}, title = {A new visual-inertial odometry scheme for unmanned systems in unified framework of zeroing neural networks}, volume = {617}, year = {2025}, }
• (2025). An HVS-derived network for assessing the quality of camouflaged targets with feature fusion. NEUCOM, 617, 129016. (www)

  @article{ , abstract = {High-value assets on the battlefield typically require adequate camouflage to evade detection and annihilation by enemy scouts. Consequently, artificial camouflage technology is extensively acknowledged and utilized as a crucial defensive tactic in the military sphere. The quality of camouflage performance was assessed by military observers through the human visual system (HVS). This method involved locating the camouflaged objects and rating the camouflaged degree against the background. Current camouflage assessment methods typically involved the manual extraction and aggregation of objective features throughout an image. These approaches fall short in constructing a correlation mapping between objective features and subjective perceptions of camouflaged objects, culminating in imprecise assessments and discrepancies. To address these issues, this paper presents the first three-stage full-reference learning framework for locating camouflaged objects, extracting camouflage features, and assessing camouflage quality. Given the lack of datasets specifically designed for evaluating camouflage quality, we have contributed a datasets focused on human-camouflaged targets. The experimental results show that the three-stage framework is remarkably accurate in assessing the camouflage quality, leading to an explainable network. The camouflaged people quality assessment(CPQA) dataset is available at http://github.com/samsunq/CPQA_Datasets.git .}, archive = {J_NEUCOM}, author = {Qiyang Sun and Xia Wang and Changda Yan and Xin Zhang and Shiwei Xu}, doi = {10.1016/j.neucom.2024.129016}, journal = {Neurocomputing}, month = {2}, pages = {129016}, shortjournal = {Neurocomputing}, title = {An HVS-derived network for assessing the quality of camouflaged targets with feature fusion}, volume = {617}, year = {2025}, }
• (2025). Swin-MSTP: Swin transformer with multi-scale temporal perception for continuous sign language recognition. NEUCOM, 617, 129015. (www)

  @article{ , abstract = {Continuous sign language recognition (CSLR) aims to recognize and interpret sequences of sign language gestures in videos. Currently, most CSLR frameworks combine spatial feature extractors based on convolutional neural networks (CNNs) with temporal convolutional networks (TCNs) for sequence learning. However, CNN-based spatial feature extractors apply the same convolutional kernel uniformly across all regions of an image, which limits their capacity to extract complex details such as fingers and facial features, which are essential for CSLR. In addition, sign languages include signs of varying lengths that cannot be accurately modeled using a fixed-size TCN. To address these issues, we present the Swin multiscale temporal perception (Swin-MSTP) framework, in which the Swin Transformer is utilized as the spatial feature extractor, capable of capturing fine spatial details and providing a stronger contextual understanding between sign language elements in video frames. The Swin Transformer was integrated with the MSTP module to extract time-wise features. Experimental results show that our single-modality system outperformed existing methods on the CSL dataset, including multimodal frameworks. The model also achieved competitive performance on the Phoenix2014, Phoenix2014-T, and CSL-Daily datasets. The code is available at https://github.com/snalyami/Swin-MSTP .}, archive = {J_NEUCOM}, author = {Sarah Alyami and Hamzah Luqman}, doi = {10.1016/j.neucom.2024.129015}, journal = {Neurocomputing}, month = {2}, pages = {129015}, shortjournal = {Neurocomputing}, title = {Swin-MSTP: Swin transformer with multi-scale temporal perception for continuous sign language recognition}, volume = {617}, year = {2025}, }
• (2025). View-channel mixer network for double incomplete multi-view multi-label learning. NEUCOM, 617, 129013. (www)

  @article{ , abstract = {In reality, multi-view multi-label learning has a richer descriptive power than single-view single-label learning. However, due to multiple uncertainties in practical situations, the process of collecting data often results in missing views and labels. Therefore, the double incomplete multi-view multi-label classification task emerged. In this complex field, two issues can be characterized as follows: i) How to embed complementary learning objectives in the view feature space? ii) How to jointly learn the consistency and complementarity of views? To this end, this paper proposes a View-Channel Mixer Network (VCMN) based on the mixing operations of a multilayer perceptron to address this complex problem. First, we designed a View-Channel Mixer module to model the complementary relationships in the high-dimensional feature space of the views. Second, cross-view and cross-sample contrastive loss is used to enhance the network’s ability to learn consistent representations across multiple views, enabling the joint learning of complementarity and consistency. Our method demonstrates competitive experimental results compared to other state-of-the-art methods on five widely used multi-view multi-label datasets.}, archive = {J_NEUCOM}, author = {Kaibo Duan and Shaodong Cui and Hiroyuki Shinnou and Shi Bao}, doi = {10.1016/j.neucom.2024.129013}, journal = {Neurocomputing}, month = {2}, pages = {129013}, shortjournal = {Neurocomputing}, title = {View-channel mixer network for double incomplete multi-view multi-label learning}, volume = {617}, year = {2025}, }
• (2025). Prototype matching-based meta-learning model for few-shot fault diagnosis of mechanical system. NEUCOM, 617, 129012. (www)

  @article{ , abstract = {The efficacy of advanced deep-learning diagnostic methods is contingent mainly upon sufficient trainable data for each fault category. However, gathering ample data in real-world scenarios is often challenging, rendering these deep-learning techniques ineffective. This paper introduces a novel Prototype Matching-based Meta-Learning (PMML) approach to address the few-shot fault diagnosis under constrained data conditions. Initially, the PMML’s feature extractor is meta-trained within the Model-Agnostic Meta-Learning framework, utilizing multiple fault classification tasks from known operational conditions in the source domain to acquire prior meta-knowledge for fault diagnosis. Subsequently, the trained feature extractor is employed to derive meta-features from few-shot samples in the target domain, and metric learning is conducted to facilitate swift and precise few-shot fault diagnosis, leveraging meta-knowledge and similarity information across sample sets. Moreover, instead of utilizing all target domain samples, the prototype of each fault category is used to capture similarity information between support and query samples. Concurrently, BiLSTM is employed to selectively embed the meta-feature prototype, enabling the extraction of more distinguishable metric features for enhanced metric learning. Finally, the effectiveness of the proposed PMML is validated through a series of comparative experiments on two fault datasets, demonstrating its outstanding performance in addressing both zero-shot and few-shot fault diagnosis challenges.}, archive = {J_NEUCOM}, author = {Lin Lin and Sihao Zhang and Song Fu and Yikun Liu and Shiwei Suo and Guolei Hu}, doi = {10.1016/j.neucom.2024.129012}, journal = {Neurocomputing}, month = {2}, pages = {129012}, shortjournal = {Neurocomputing}, title = {Prototype matching-based meta-learning model for few-shot fault diagnosis of mechanical system}, volume = {617}, year = {2025}, }
• (2025). Towards robust DeepFake distortion attack via adversarial autoaugment. NEUCOM, 617, 129011. (www)

  @article{ , abstract = {Face forgery by DeepFake is posing a potential threat to society. Previous studies have shown that adversarial examples can effectively disrupt DeepFake models. However, the practical application of adversarial examples to defend against DeepFake is limited due to the existence of various input transformations. To address this issue, we propose a R obust D eepFake D istortion A ttack (RDDA) method from the perspective of data augmentation, which uses adversarial autoaugment to generate robust and generalized adversarial examples to disrupt DeepFake. Specifically, we design an adversarial autoaugment module to synthesize diverse and challenging input transformations. Through coping with these transformations, the robustness and generalization ability of the adversarial examples in disrupting DeepFake models are greatly enhanced. In addition, we further improve the generalization ability of adversarial examples in handling specific input transformations by incremental learning. With RDDA and incremental learning, our generated adversarial examples can effectively protect personal privacy from being violated by DeepFake. Extensive experiments on public benchmarks demonstrate that our DeepFake defense method has better robustness and generalization ability than state-of-the-arts.}, archive = {J_NEUCOM}, author = {Qi Guo and Shanmin Pang and Zhikai Chen and Qing Guo}, doi = {10.1016/j.neucom.2024.129011}, journal = {Neurocomputing}, month = {2}, pages = {129011}, shortjournal = {Neurocomputing}, title = {Towards robust DeepFake distortion attack via adversarial autoaugment}, volume = {617}, year = {2025}, }
• (2025). A novel multi-morphological representation approach for multi-source EEG signals. NEUCOM, 617, 129010. (www)

  @article{ , abstract = {Advances in artificial intelligence have significantly enhanced intelligent assistance and rehabilitation medicine by leveraging electroencephalogram (EEG) signal recognition. Nevertheless, eliminating cross-subject variability remains a significant challenge in expending the application of EEG signal recognition to the broader society. The transfer learning strategy has been utilized to address this issue; however, multi-source domains are often treated as a single entity in transfer learning, leading to underutilization of the information from multiple sources. Furthermore, many EEG signal transfer approaches overlook the low-dimensional structural information and multivariate statistical features inherent in EEG signals, leading to inadequate interpretability and suboptimal performance. Thus, in this study, a novel multi-morphological representation approach (MMRA) was proposed for multi-source EEG signal recognition to address these issues. MMRA utilized multi-manifold mapping to extract the common invariant representation shared between the multi-source domains and target domain. It took into account the low-dimensional structure and multivariate statistical features of EEG signals to enhance the acquisition of high-quality common invariant representations. Subsequently, the multi-source domains were decomposed to extract one-to-one features. The Maximum Mean Discrepancy (MMD) loss was further applied to guide the model in obtaining high-quality private invariant representations. The performance of the proposed MMRA method was evaluated using three publicly available motor imagery datasets and a driving fatigue dataset. Experimental results demonstrated that our proposed MMRA method outperformed other state-of-the-art methods in scenarios involving multiple subjects. In conclusion, the MMRA method developed in this study can serve as a novel tool offering enhanced performance to analyze EEG signals across various subjects.}, archive = {J_NEUCOM}, author = {Yunyuan Gao and Yici Liu and Ming Meng and Feng Fang and Michael Houston and Yingchun Zhang}, doi = {10.1016/j.neucom.2024.129010}, journal = {Neurocomputing}, month = {2}, pages = {129010}, shortjournal = {Neurocomputing}, title = {A novel multi-morphological representation approach for multi-source EEG signals}, volume = {617}, year = {2025}, }
• (2025). Soft prompt-tuning for unsupervised domain adaptation via self-supervision. NEUCOM, 617, 129008. (www)

  @article{ , abstract = {Unsupervised domain adaptation methods aim to facilitate learning tasks in unlabeled target domains using labeled information from related source domains. Recently, prompt-tuning has emerged as a powerful instrument to incorporate templates that reformulate input examples into equivalent cloze-style phrases. However, there are still two great challenges for domain adaptation: (1) Existing prompt-tuning methods only rely on the general knowledge distributed in upstream pre-trained language models to alleviate the domain discrepancy. How to incorporate specific features in the source and target domains into prompt-tuning model is still divergent and under-explored; (2) In the prompt-tuning, either the crafted template methods are time-consuming and labor-intensive, or automatic prompt generation methods cannot achieve satisfied performance. To address these issues, in this paper, we propose an innovative Soft Prompt-tuning method for Unsupervised Domain Adaptation via Self-Supervision, which combines two novel ideas: Firstly, instead of only stimulating knowledge distributed in the pre-trained model, we further employ hierarchically clustered optimization strategies in a self-supervised manner to retrieve knowledge for the verbalizer construction in prompt-tuning. Secondly, we construct prompts with the special designed verbalizer that facilitate the transfer of learning representations across domains, which can consider both the automatic template generation and cross-domain classification performance. Extensive experimental results demonstrate that our method even outperforms SOTA baselines that utilize external open knowledge with much less computational time.}, archive = {J_NEUCOM}, author = {Yi Zhu and Shuqin Wang and Yun Li and Yunhao Yuan and Jipeng Qiang}, doi = {10.1016/j.neucom.2024.129008}, journal = {Neurocomputing}, month = {2}, pages = {129008}, shortjournal = {Neurocomputing}, title = {Soft prompt-tuning for unsupervised domain adaptation via self-supervision}, volume = {617}, year = {2025}, }
• (2025). CoATR: A convolutional autoregressive tensor-ring decomposition method for sparse spatio-temporal traffic data. NEUCOM, 617, 129006. (www)

  @article{ , abstract = {Spatio-temporal traffic data collected by various sensing systems have chronic issues of missing and corruption, thus accurate data imputation and prediction have been extensively researched. Previous most methods can be divided into two categories: model-driven methods with physical explanations and data-driven deep learning methods. These methods all achieve promising results due to their respective advantages. However they have some limitations: model-driven methods are often linear and may not accurately model the spatio-temporal complexity, while deep learning methods often lack the physical reality making them probe to over-fitting. Inspired by both, we propose a new Co nvolutional-based generalized A utoregressive T ensor- R ing decomposition method (CoATR) for the completion of spatio-temporal data. CoATR not only retains the advantages of the tensor-ring (TR) decomposition model for global modeling of spatio-temporal data but also exploits the ability of deep networks to model nonlinear features. To be specific, we introduce TR decomposition to capture the global low-rankness and employ multilayer convolutional neural networks to model the global complex interactions among the TR factors for exploring the nonlinear features of the spatio-temporal data. Moreover, we design a new autoregressive network to further explore the local temporal variation in the data. Extensive experiments on a variety of common traffic datasets have validated the effectiveness and superiority of the CoATR over classical model-driven methods and other state-of-the-art data-driven deep learning methods.}, archive = {J_NEUCOM}, author = {Tianchi Liao and Lei Zhang and Jinghua Yang and Chuan Chen and Zibin Zheng}, doi = {10.1016/j.neucom.2024.129006}, journal = {Neurocomputing}, month = {2}, pages = {129006}, shortjournal = {Neurocomputing}, title = {CoATR: A convolutional autoregressive tensor-ring decomposition method for sparse spatio-temporal traffic data}, volume = {617}, year = {2025}, }
• (2025). Enhancing cooperative multi-agent reinforcement learning through the integration of R-STDP and federated learning. NEUCOM, 617, 129005. (www)

  @article{ , abstract = {This paper introduces a novel approach to enhance the stability and efficiency of R-STDP in the context of federated learning. The primary objective is to stabilize the unbounded growth of R-STDP and make it more responsive to real-time changes. The methodology involves integrating R-STDP with Spiking Neural Networks and employing the norm of the neural network model for adjusting weighted aggregation in federated learning systems. The proposed method incorporates a mechanism where weights decay over time, depending on the duration since the agent last published its model. Additionally, the sampling time is dynamically adjusted based on the Euclidean norm, which measures the distance between the weight matrices of the agents and the server. The results demonstrate that the proposed event-triggered federated learning method significantly enhances learning speed and performance. At the same time, the dynamic aggregation interval efficiently reduces communication between the agents and the central server, especially after model convergence. This research presents a significant advancement in federated learning and offers a more stable, responsive, and efficient learning process.}, archive = {J_NEUCOM}, author = {Mohammad Tayefe Ramezanlou and Howard Schwartz and Ioannis Lambadaris and Michel Barbeau}, doi = {10.1016/j.neucom.2024.129005}, journal = {Neurocomputing}, month = {2}, pages = {129005}, shortjournal = {Neurocomputing}, title = {Enhancing cooperative multi-agent reinforcement learning through the integration of R-STDP and federated learning}, volume = {617}, year = {2025}, }
• (2025). Point cloud feature consistency learning for incomplete 3D face recognition. NEUCOM, 617, 129000. (www)

  @article{ , abstract = {Point cloud-based 3D face recognition has emerged as an exciting research topic due to the availability of 3D facial structures and detailed surface information. Existing approaches have primarily focused on complete facial point clouds and have achieved remarkable results. However, in real-world applications, the collected facial point clouds are often incomplete due to factors such as various poses, occlusion, and noise, posing significant challenges to face recognition tasks. In this paper, a feature consistency learning framework is proposed to improve incomplete 3D face recognition. The feature gap between incomplete and complete data is filled through joint optimization of completion and supervised contrastive learning. Specifically, to maintain and enhance the structure of incomplete point clouds, we introduce a structure-enhanced representation method for neighboring points that incorporates positional information residuals during the formation of point proxies. Additionally, a simple and effective dynamic input approach within the point proxy completion process is designed to alleviate concerns related to density disparities and detail loss in point clouds that exhibit relatively minor degrees of incompleteness. Extensive experiments on four datasets demonstrate our proposed method outperforms state-of-the-art methods on both inherent and artificially constructed incomplete data. Moreover, it also achieves comparable results on complete 3D face recognition. Overall, this work represents an early exploration into the realm of point cloud-based incomplete 3D face recognition through feature consistency learning, providing a promising approach for practical applications.}, archive = {J_NEUCOM}, author = {Faxiu Huang and Yanqiong Guo and Zhu Xu and Zhisheng You and Xiao Yang}, doi = {10.1016/j.neucom.2024.129000}, journal = {Neurocomputing}, month = {2}, pages = {129000}, shortjournal = {Neurocomputing}, title = {Point cloud feature consistency learning for incomplete 3D face recognition}, volume = {617}, year = {2025}, }
• (2025). Depth-wise convolutions in vision transformers for efficient training on small datasets. NEUCOM, 617, 128998. (www)

  @article{ , abstract = {The Vision Transformer (ViT) leverages the Transformer’s encoder to capture global information by dividing images into patches and achieves superior performance across various computer vision tasks. However, the self-attention mechanism of ViT captures the global context from the outset, overlooking the inherent relationships between neighboring pixels in images or videos. Transformers mainly focus on global information while ignoring the fine-grained local details. Consequently, ViT lacks inductive bias during image or video dataset training. In contrast, convolutional neural networks (CNNs), with their reliance on local filters, possess an inherent inductive bias, making them more efficient and quicker to converge than ViT with less data. In this paper, we present a lightweight Depth-Wise Convolution module as a shortcut in ViT models, bypassing entire Transformer blocks to ensure the models capture both local and global information with minimal overhead. Additionally, we introduce two architecture variants, allowing the Depth-Wise Convolution modules to be applied to multiple Transformer blocks for parameter savings, and incorporating independent parallel Depth-Wise Convolution modules with different kernels to enhance the acquisition of local information. The proposed approach significantly boosts the performance of ViT models on image classification, object detection, and instance segmentation by a large margin, especially on small datasets, as evaluated on CIFAR-10, CIFAR-100, Tiny-ImageNet and ImageNet for image classification, and COCO for object detection and instance segmentation. The source code can be accessed at https://github.com/ZTX-100/Efficient_ViT_with_DW .}, archive = {J_NEUCOM}, author = {Tianxiao Zhang and Wenju Xu and Bo Luo and Guanghui Wang}, doi = {10.1016/j.neucom.2024.128998}, journal = {Neurocomputing}, month = {2}, pages = {128998}, shortjournal = {Neurocomputing}, title = {Depth-wise convolutions in vision transformers for efficient training on small datasets}, volume = {617}, year = {2025}, }
• (2025). Do inpainting yourself: Generative facial inpainting guided by exemplars. NEUCOM, 617, 128996. (www)

  @article{ , abstract = {We present EXE-GAN, a novel exemplar-guided facial inpainting framework using generative adversarial networks. Our approach not only preserves the quality of the input facial image but also completes the image with exemplar-like facial attributes. We achieve this by simultaneously leveraging the global style of the input image, the stochastic style generated from the random latent code, and the exemplar style of the exemplar image. We introduce a novel attribute similarity metric to encourage the networks to learn the style of facial attributes from the exemplar in a self-supervised way. To guarantee the natural transition across the boundaries of inpainted regions, we introduce a novel spatial variant gradient backpropagation technique to adjust the loss gradients based on the spatial location. We extensively evaluate EXE-GAN on public CelebA-HQ and FFHQ datasets with practical applications, which demonstrates the superior visual quality of facial inpainting. The source code is available at https://github.com/LonglongaaaGo/EXE-GAN .}, archive = {J_NEUCOM}, author = {Wanglong Lu and Hanli Zhao and Xianta Jiang and Xiaogang Jin and Yong-Liang Yang and Kaijie Shi}, doi = {10.1016/j.neucom.2024.128996}, journal = {Neurocomputing}, month = {2}, pages = {128996}, shortjournal = {Neurocomputing}, title = {Do inpainting yourself: Generative facial inpainting guided by exemplars}, volume = {617}, year = {2025}, }
• (2025). Global context volume construction and semantics-guided disparity refinement for stereo matching. NEUCOM, 617, 128995. (www)

  @article{ , abstract = {The accuracy of stereo matching has been greatly improved with the advent of convolutional neural networks. However, existing methods still perform poorly in regions of low-texture and reflections due to insufficient matching information. In this paper, we propose a novel semantic stereo matching network named GSSNet, which combines global context information with high-level semantic clues to further enrich the matching features. Specifically, we employ Swin Transformer as a joint feature extractor to capture multi-scale global features for both semantic segmentation and disparity estimation. By fusing low-resolution multi-scale feature maps progressively, we construct a Global Context Volume (GCV) representation covering long-range receptive field for initial disparity estimation. Initial disparities are further refined with a high-level embedding learned from the semantic segmentation branch of the network through residual learning. During training, due to the lack of joint annotations, we also propose a simple and effective pseudo-labeling strategy. Comprehensive experimental results demonstrated on various datasets manifest the effectiveness of the proposed GSSNet. Among all published methods as of 9, February 2024, our approach ranks 1st on KITTI 2012 leaderboard and 3rd on KITTI 2015 leaderboard, and produces competitive results on Scene Flow and Cityscapes datasets. Code will be available at https://github.com/Twil-7/GSSNet .}, archive = {J_NEUCOM}, author = {Fudong Xu and Lin Wang and Huibin Li and Zhipeng Zhu and Fan Jia}, doi = {10.1016/j.neucom.2024.128995}, journal = {Neurocomputing}, month = {2}, pages = {128995}, shortjournal = {Neurocomputing}, title = {Global context volume construction and semantics-guided disparity refinement for stereo matching}, volume = {617}, year = {2025}, }
• (2025). RTA: A reinforcement learning-based temporal knowledge graph question answering model. NEUCOM, 617, 128994. (www)

  @article{ , abstract = {Temporal Knowledge Graph Question Answering (TKGQA) is crucial research, focusing on finding an entity or a timestamp to answer temporal questions in the corresponding temporal knowledge graph. Currently, the main challenge in the temporal KGQA task is answering complex temporal questions, often necessitating complex multi-hop temporal reasoning in the TKG. In this paper, we propose a method for the TKGQA task called Reinforcement learning Temporal knowledge graph question A nswering ( RTA ). First, in the question understanding stage, our model extracts context information to select topic entities of the given question, which can effectively deal with scenarios involving multiple entities in complex temporal questions. Furthermore, reasoning complexity escalates significantly with complex temporal questions, as varying timestamps alter the relations between entities. Therefore, we introduce reinforcement learning into the reasoning process. In the policy network, a dynamic path-matching module is specifically included to aggregate the features of relational paths to effectively capture the dynamic changes of the relations between entities on the reasoning paths. At the same time, the weights are calculated to obtain the degree of attention of each candidate action. Then the score of each candidate action is obtained through a weighted summation mechanism which helps the agent learn the optimal path reasoning policy for effective exploration. Finally, we evaluate our method on the CRONQUESTIONS dataset and validate its superiority over all baseline methods. Specifically, our approach proves effective in handling complex temporal questions.}, archive = {J_NEUCOM}, author = {Yu Zhu and Tinghuai Ma and Shengjie Sun and Huan Rong and Yexin Bian and Kai Huang}, doi = {10.1016/j.neucom.2024.128994}, journal = {Neurocomputing}, month = {2}, pages = {128994}, shortjournal = {Neurocomputing}, title = {RTA: A reinforcement learning-based temporal knowledge graph question answering model}, volume = {617}, year = {2025}, }
• (2025). Unmasking deception: A comprehensive survey on the evolution of face anti‐spoofing methods. NEUCOM, 617, 128992. (www)

  @article{ , abstract = {With the growing popularity of facial recognition (FR) in access control systems, there has been a corresponding increase in presentation attacks (PAs) to gain unauthorized access. The sophistication of these attacks has been exacerbated by advancements in technology and a reduction in computation costs, thus posing significant security threats to small- and large-scale deployments alike. To address this issue, Face Anti-Spoofing (FAS) and Presentation Attack Detection (PAD) have garnered significant interest from the research community in recent years. This paper provides a comprehensive review of the state-of-the-art works published over the past decade and discusses the temporal evolution of the FAS/PAD field. It reviews different types of attacks against facial authentication systems and covers key features used for FAS models. It also discusses the FAS design approaches followed by the backbone architectures used to design these methods. It also discusses publicly available databases for FAS models, standard protocols, and benchmarking methods. An extensive comparative analysis of experimental results from different PAD methods over the past decade is provided, highlighting limitations and current challenges. It observes a lack of a robust, large scale general dataset for FAS and underscores the need for new developments in the field.}, archive = {J_NEUCOM}, author = {Aashania Antil and Chhavi Dhiman}, doi = {10.1016/j.neucom.2024.128992}, journal = {Neurocomputing}, month = {2}, pages = {128992}, shortjournal = {Neurocomputing}, title = {Unmasking deception: A comprehensive survey on the evolution of face anti‐spoofing methods}, volume = {617}, year = {2025}, }
• (2025). DMANet: Dual-branch multiscale attention network for real-time semantic segmentation. NEUCOM, 617, 128991. (www)

  @article{ , abstract = {The dual-branch structure has made impressive achievements in the field of semantic segmentation. However, the size of feature map tends to be kept constant in the single branch of dual-branch network, which leads to the lack of multiscale feature information. Previous self-attention mechanisms were often limited to processing singlescale feature maps, which restricted the ability of multiscale feature extraction. To alleviate these two problems, in this paper we propose a dual-branch multiscale attention network (DMANet) for real-time semantic segmentation. In order to make full use of the multiscale feature information, our DMANet utilizes the multiscale external attention block (EAB) to handle feature maps at different scales. In addition, our proposed dual-branch cascade attention (DCA) module can capture more local information in some sense and enable information interaction across branches. Our DCA module is built and based on the intra-branch cascade fusion (IBCF) structure. Extensive experiments demonstrate the effectiveness of our proposed DMANet. On Cityscapes and CamVid datasets, DMANet achieves 78.5% and 78.7% mIoU respectively and the inference speed reaches the real-time requirement.}, archive = {J_NEUCOM}, author = {Yongsheng Dong and Chongchong Mao and Lintao Zheng and Qingtao Wu}, doi = {10.1016/j.neucom.2024.128991}, journal = {Neurocomputing}, month = {2}, pages = {128991}, shortjournal = {Neurocomputing}, title = {DMANet: Dual-branch multiscale attention network for real-time semantic segmentation}, volume = {617}, year = {2025}, }
• (2025). A neurodynamic approach with fixed-time convergence for complex-variable pseudo-monotone variational inequalities. NEUCOM, 617, 128988. (www)

  @article{ , abstract = {Based on Wirtinger calculus, this paper proposes a complex-valued projection neural network (CPNN) designed to address complex-variables variational inequality problems. The global convergence of the CPNN is established under the assumptions of pseudomonotonicity and Lipschitz continuity. We demonstrate that the CPNN achieves convergence within a fixed-time, which is unaffected by the initial conditions and converges towards the optimal solution of the constrained optimization problem. And this result is distinct from asymptotic or exponential convergence that depend on initial condition. Furthermore, the CPNN shows utility in tackling diverse related problems, encompassing variational inequalities, pseudo-convex optimization problems, linear and nonlinear complementarity problems, as well as linear and convex quadratic programming problems. The efficacy of the proposed CPNN is substantiated through numerical simulations.}, archive = {J_NEUCOM}, author = {Jinlan Zheng and Xingxing Ju and Naimin Zhang and Dongpo Xu}, doi = {10.1016/j.neucom.2024.128988}, journal = {Neurocomputing}, month = {2}, pages = {128988}, shortjournal = {Neurocomputing}, title = {A neurodynamic approach with fixed-time convergence for complex-variable pseudo-monotone variational inequalities}, volume = {617}, year = {2025}, }
• (2025). Scalable kernel logistic regression with nyström approximation: Theoretical analysis and application to discrete choice modelling. NEUCOM, 617, 128975. (www)

  @article{ , abstract = {The application of kernel-based Machine Learning (ML) techniques to discrete choice modelling using large datasets often faces challenges due to memory requirements and the considerable number of parameters involved in these models. This complexity hampers the efficient training of large-scale models. This paper addresses these problems of scalability by introducing the Nyström approximation for Kernel Logistic Regression (KLR) on large datasets. The study begins by presenting a theoretical analysis in which: (i) the set of KLR solutions is characterised, (ii) an upper bound to the solution of KLR with Nyström approximation is provided, and finally (iii) a specialisation of the optimisation algorithms to Nyström KLR is described. After this, the Nyström KLR is computationally validated. Four landmark selection methods are tested, including basic uniform sampling, a k -means sampling strategy, and two non-uniform methods grounded in leverage scores. The performance of these strategies is evaluated using large-scale transport mode choice datasets and is compared with traditional methods such as Multinomial Logit (MNL) and contemporary ML techniques. The study also assesses the efficiency of various optimisation techniques for the proposed Nyström KLR model. The performance of gradient descent, Momentum, Adam, and L-BFGS-B optimisation methods is examined on these datasets. Among these strategies, the k -means Nyström KLR approach emerges as a successful solution for applying KLR to large datasets, particularly when combined with the L-BFGS-B and Adam optimisation methods. The results highlight the ability of this strategy to handle datasets exceeding 200,000 observations while maintaining robust performance.}, archive = {J_NEUCOM}, author = {José Ángel Martín-Baos and Ricardo García-Ródenas and Luis Rodriguez-Benitez and Michel Bierlaire}, doi = {10.1016/j.neucom.2024.128975}, journal = {Neurocomputing}, month = {2}, pages = {128975}, shortjournal = {Neurocomputing}, title = {Scalable kernel logistic regression with nyström approximation: Theoretical analysis and application to discrete choice modelling}, volume = {617}, year = {2025}, }
• (2025). Training-free prior guided diffusion model for zero-reference low-light image enhancement. NEUCOM, 617, 128974. (www)

  @article{ , abstract = {Images captured under poor illumination not only struggle to provide satisfactory visual information but also adversely affect high-level visual tasks. Therefore, we delve into low-light image enhancement. We mainly focus on two practical challenges: (1) previous methods predominantly require supervised training with paired data, tending to learn mappings specific to the training data, which limits their generalization ability on unseen images. (2) existing unsupervised methods usually yield sub-optimal image quality due to the insufficient utilization of image priors. To address these challenges, we propose a training-free Prior Guided Diffusion model, namely PGDiff , for zero-reference low-light image enhancement. Specifically, to leverage the implicit information within the degraded image, we propose a frequency-guided mechanism to obtain low-frequency features through bright channel prior, which combined with the generative prior of the pre-trained diffusion model to recover high-frequency details. To improve the quality of generated images, we further introduce the gradient guidance based on image exposure and color priors. Benefiting from this dual-guided mechanism, PGDiff can produce high-quality restoration results without requiring tedious training or paired reference images. Extensive experiments on paired and unpaired datasets show that our training-free method achieves competitive performance against existing learning-based methods, surpassing the state-of-the-art method QuadPrior by 0.25 dB in PSNR on the LOL dataset.}, archive = {J_NEUCOM}, author = {Kai Shang and Mingwen Shao and Chao Wang and Yuanjian Qiao and Yecong Wan}, doi = {10.1016/j.neucom.2024.128974}, journal = {Neurocomputing}, month = {2}, pages = {128974}, shortjournal = {Neurocomputing}, title = {Training-free prior guided diffusion model for zero-reference low-light image enhancement}, volume = {617}, year = {2025}, }
• (2025). ParaDC: Parallel-learning-based dynamometer cards augmentation with diffusion models in sucker rod pump systems. NEUCOM, 617, 128973. (www)

  @article{ , abstract = {The accurate fault diagnosis of sucker rod pump systems (SRPs) is crucial for the sustainable development of oil & gas. Currently, dynamometer cards (DCs) are widely employed to evaluate the working condition of SRPs, framing fault diagnosis as a pattern recognition problem. While significant attention has been dedicated to enhancing the performance of diagnostic algorithms, the critical role of high-quality DC datasets in improving diagnostic accuracy has been comparatively underexplored. To address issues of incomplete and imbalanced data distribution in existing DC datasets, this paper introduces ParaDC, a novel data augmentation mechanism grounded in parallel learning, to facilitate the transition of DCs from “small data” to “big data” and ultimately realize “deep intelligence”. Under this mechanism, wave equations representing the behavior of SRPs are first utilized to construct the customized “small” DC datasets. Diffusion models are then incorporated to augment the “big” DC datasets and enhance data diversity. Additionally, iterative training combined with human feedback is introduced to optimize and improve the quality of generated DCs, accelerating the pathway towards “deep intelligence”. To further validate the feasibility and effectiveness of ParaDC, extensive computational experiments are conducted to demonstrate its outstanding generative performance. Finally, the potential of intelligent diagnostic systems, supported by digital workers, is discussed in the context of Industry 5.0, which is believed to be indispensable in the future industrial diagnostic paradigm.}, archive = {J_NEUCOM}, author = {Xingxia Wang and Yuhang Liu and Xiang Cheng and Yutong Wang and Yonglin Tian and Fei-Yue Wang}, doi = {10.1016/j.neucom.2024.128973}, journal = {Neurocomputing}, month = {2}, pages = {128973}, shortjournal = {Neurocomputing}, title = {ParaDC: Parallel-learning-based dynamometer cards augmentation with diffusion models in sucker rod pump systems}, volume = {617}, year = {2025}, }
• (2025). Filter competition results in more robust convolutional neural networks. NEUCOM, 617, 128972. (www)

  @article{ , abstract = {Convolutional layers, one of the basic building blocks of deep learning architectures, contain numerous trainable filters for feature extraction. These filters operate independently which can result in distinct filters learning similar weights and extracting similar features. In contrast, competition mechanisms in the brain contribute to the sharpening of the responses of activated neurons, enhancing the contrast and selectivity of individual neurons towards specific stimuli, and simultaneously increasing the diversity of responses across the population of neurons. Inspired by this observation, this paper proposes a novel convolutional layer based on the theory of predictive coding, in which each filter effectively tries to block other filters from responding to the input features which it represents. In this way, filters learn to become more distinct which increases the diversity of the extracted features. When replacing standard convolutional layers with the proposed layers the performance of classification networks is not only improved on ImageNet but also significantly boosted on eight robustness benchmarks, as well as on downstream detection and segmentation tasks. Most notably, ResNet50/101/152 robust accuracy increases by 15.9%/20.0%/20.9% under FGSM attack, and by 10.5%/14.7%/15.0% under PGD attack.}, archive = {J_NEUCOM}, author = {Bo Gao and Michael W. Spratling}, doi = {10.1016/j.neucom.2024.128972}, journal = {Neurocomputing}, month = {2}, pages = {128972}, shortjournal = {Neurocomputing}, title = {Filter competition results in more robust convolutional neural networks}, volume = {617}, year = {2025}, }
• (2025). DRL for handover in 6G-vehicular networks: A survey. NEUCOM, 617, 128971. (www)

  @article{ , abstract = {3GPP is working on technology improvements related to sixth-generation (6G) wireless communication networks to keep pace. 6G networks are being developed as the next phase forward. Compared to their predecessor wireless technologies, 6G networks are predicted to offer better coverage and flexibility by supporting higher throughput, faster velocities, lower latency, and higher capacity. 6G targets the health, education, industry, and transport sectors. The transport field is undergoing rapid change. 6G and artificial intelligence (AI) will have an essential impact in this area, bringing users new services and functionalities. Within this field, the handover (HO) mechanism remains a concern that researchers must consider for achieving excellent communication quality since HO technology is crucial for ensuring seamless connectivity during user transfers between cells. Numerous proposed Machine Learning (ML) approaches, including Deep Reinforcement Learning (DRL), were discussed to solve HO issues. Recently, DRL methods have garnered significant interest in prospective wireless networks. They can surmount the escalating obstacles of the wireless environment and the constraints of conventional approaches. Moreover, DRL is crucial in wireless networks because of its capability to overcome the specific threats and dynamic nature of wireless settings. Elaborating on a comprehensive survey of these approaches related to HO and DRL can provide a unified analysis of the current advancements. This overview will help to improve understanding of this topic. This survey provides an overview of requirements and usage scenarios for 6G. It highlights the impact of this new wireless technology on the transportation field or the Vehicle-to-everything (V2X). In addition, we provide a deep study of HO management in 6G networks and elaborate on the various DRL literature solutions for HO in mobile and vehicular networks.}, archive = {J_NEUCOM}, author = {Arwa Amaira and Hend Koubaa and Faouzi Zarai}, doi = {10.1016/j.neucom.2024.128971}, journal = {Neurocomputing}, month = {2}, pages = {128971}, shortjournal = {Neurocomputing}, title = {DRL for handover in 6G-vehicular networks: A survey}, volume = {617}, year = {2025}, }
• (2025). DouN-GNN: Double nodes graph neural network for few-shot learning. NEUCOM, 617, 128970. (www)

  @article{ , abstract = {In recent years, graph neural networks (GNNs) for few-shot learning have garnered significant attention due to their powerful learning capabilities. However, previous methods typically construct nodes using single-modal samples, often overlooking additional information (e.g., high-frequency details information) that can be provided by other modalities, which may limit model performance. To fully leverage multi-dimensional information from various sample modalities, we propose a novel double-node graph neural network (DouN-GNN). In our approach, each node comprises two sub-nodes, with each sub-node representing a different modality of the sample image. To address the issue of information redundancy between modalities when constructing sub-nodes, we introduce an orthogonal transformation method to orthogonalize the sub-node features. Additionally, we develop a graph update module for double-nodes, which alternately updates the nodes and edges of the graph to facilitate the aggregation of multi-dimensional information from multi-modal images. As the number of graph update layers increases, the edge features become more reliable, further enhancing performance. Extensive experiments on the miniImageNet, TieredImageNet, and CUB-200-2011 datasets demonstrate that our method outperforms existing state-of-the-art approaches.}, archive = {J_NEUCOM}, author = {Yan Zhang and Xudong Zhou and Nian Wang and Jun Tang and Tao Xuan}, doi = {10.1016/j.neucom.2024.128970}, journal = {Neurocomputing}, month = {2}, pages = {128970}, shortjournal = {Neurocomputing}, title = {DouN-GNN: Double nodes graph neural network for few-shot learning}, volume = {617}, year = {2025}, }
• (2025). Enhancing object recognition: The role of object knowledge decomposition and component-labeled datasets. NEUCOM, 617, 128969. (www)

  @article{ , abstract = {Deep learning models’ decision-making processes can be elusive, often raising concerns about their reliability. To address this, we have introduced the Object Knowledge Decomposition and Components Label Dataset (OKD-CL), designed to improve the interpretability and accuracy of object recognition models. This dataset includes 99 categories from PartImageNet, each detailed with clear physical structures that align with human visual concepts. In a hierarchical structure, every category is described by Abstract Component Knowledge (ACK) descriptions and each image instance comes with Explicit Visual Knowledge (EVK) masks, highlighting the visual components’ appearance. By evaluating multiple deep neural networks guided with ACK and EVK (dual-knowledge-guidance approach), we saw better accuracy and a higher Foreground Reasoning Ratio (FRR), confirming our knowledge-guided method’s effectiveness. When used on the Hard-ImageNet dataset, this approach reduced the model’s reliance on incorrect feature assumptions without sacrificing classification accuracy. This hierarchical comprehension encouraged by OKD-CL is crucial in minimizing incorrect feature associations and strengthening model robustness. The entire code and dataset are available on: https://github.com/XiGuaBo/OKD-CL .}, archive = {J_NEUCOM}, author = {Nuoye Xiong and Ning Wang and Hongsheng Li and Guangming Zhu and Liang Zhang and Syed Afaq Ali Shah and Mohammed Bennamoun}, doi = {10.1016/j.neucom.2024.128969}, journal = {Neurocomputing}, month = {2}, pages = {128969}, shortjournal = {Neurocomputing}, title = {Enhancing object recognition: The role of object knowledge decomposition and component-labeled datasets}, volume = {617}, year = {2025}, }
• (2025). Sparse cascade: Multi-stage cascade fusion and shape dictionary guided network for segmentation of microscopic cytology images. NEUCOM, 617, 128968. (www)

  @article{ , abstract = {Microscopic cytology image instance segmentation is critical for clinical diagnosis, treatment planning, and biomedical research. However, challenges such as cell overlapping, external interferences, and irregular shapes hinder accurate segmentation. In this paper, a microscopic cytology image instance segmentation algorithm based on a multi-stage cascaded fusion and shape-guided strategy is proposed to improve instance segmentation accuracy. To mitigate the issue of external interference and cell overlap, a feedforward fusion mechanism is introduced in the multi-stage cascade architecture to effectively integrate the mask features from multiple stages. To make the shape of the segmentation result more consistent with the shape of the cell, and to alleviate the issues of unclear boundaries and incomplete shape segmentation caused by cellular overlap and irregular shapes, the sparse code guidance branch based on the explicit shape guidance strategy is introduced. Experiments on the private embryonic cell dataset and the ISBI2014 dataset indicate that the network has achieved improvements over the Cascade Mask R-CNN, with enhancements of 2.51% and 3.92% in bounding boxes mean Average Precision (mAP), and 2.82% and 4.70% in segmentation mAP, respectively. The experimental results indicate that the network proposed in this paper can output more accurate bounding boxes and instance masks.}, archive = {J_NEUCOM}, author = {Ao Liu and Yuan Wang and Jie Ding and Xinyu Ma and Zhenying Zhao and Haixuan Sun and Jian Zheng}, doi = {10.1016/j.neucom.2024.128968}, journal = {Neurocomputing}, month = {2}, pages = {128968}, shortjournal = {Neurocomputing}, title = {Sparse cascade: Multi-stage cascade fusion and shape dictionary guided network for segmentation of microscopic cytology images}, volume = {617}, year = {2025}, }
• (2025). EMPViT: Efficient multi-path vision transformer for security risks detection in power distribution network. NEUCOM, 617, 128967. (www)

  @article{ , abstract = {To maintain the safe operation of power distribution network (PDN) equipment, it is important to accurately and promptly identify security risks. However, conventional drone-based object detection methods face challenges due to noise and similarity features in risk targets, as well as limited computing resources of unmanned aerial vehicles (UAVs). To address these challenges, an efficient embedding-based multi-path fusion architecture is proposed. This architecture uses a re-parameterized depthwise block to embed local context information at different scales, enhancing the extraction of tiny features while preserving inference speed. Additionally, a coordinated self-attention module is proposed to reduce computational complexity while maintaining the performance of global information. By fusing fine and coarse feature representations without requiring a lot of computation, this module efficiently learns from both local and global features from images. The goal is to create an efficient multi-path vision transformer (EMPViT) architecture that achieves a balance between accuracy and efficiency. The proposed EMPViT has been evaluated on two different drone image dataset, demonstrating better performance compared to other architectures. Specifically, the EMPViT-S improves the detection mAP by 1.2%, and the inference speed is improved to 1.24 times on average on Drone-PDN dataset. It has achieved the same performance improvement on VisDrone-DET2019 dataset, gaining detection performance by 1.3% and 1.2 times acceleration on average.}, archive = {J_NEUCOM}, author = {Pan Li and Xiaofang Yuan and Haozhi Xu and Jinlei Wang and Yaonan Wang}, doi = {10.1016/j.neucom.2024.128967}, journal = {Neurocomputing}, month = {2}, pages = {128967}, shortjournal = {Neurocomputing}, title = {EMPViT: Efficient multi-path vision transformer for security risks detection in power distribution network}, volume = {617}, year = {2025}, }
• (2025). Improving generalization performance of adaptive gradient method via bounded step sizes. NEUCOM, 617, 128966. (www)

  @article{ , abstract = {While adaptive gradient methods such as Adam have been widely used in the training of deep neural networks, a recent study has provided a synthetic function that shows the non-convergence problem of Adam. This issue stems from the existence of extreme gradients and the mismatch between the first and second moments. Several adaptive optimizers have been continuously developed. However, designing a fast optimizer with excellent generalization capability is still challenging. We propose an adaptive method with bounded step sizes, named AdaBS, which removes the extreme step sizes and ensures that it appropriately adjusts adaptive step sizes to mitigate the over-adaptation of step sizes in Adam. In particular, AdaBS effectively clips step sizes that are too large or too small by using two static bounds with a predetermined boundary to control updates. When determining the step size, static bound clipping will be used if the preconditioner is outside the modest boundary, and vanilla Adam will be used if the preconditioner is inside the boundary. AdaBS establishes a trust region around the basic step size and obtains benefits of both Adam and SGD, i.e. fast convergence and better generalization. Finally, we conduct extensive experiments on a variety of practical tasks with benchmark datasets, including image classification and modeling language tasks. Empirical results demonstrate AdaBS’s promising performance with remarkably fast convergence, superior generalization, and robustness.}, archive = {J_NEUCOM}, author = {Yangchuan Wang and Lianhong Ding and Peng Shi and Juntao Li and Ruiping Yuan}, doi = {10.1016/j.neucom.2024.128966}, journal = {Neurocomputing}, month = {2}, pages = {128966}, shortjournal = {Neurocomputing}, title = {Improving generalization performance of adaptive gradient method via bounded step sizes}, volume = {617}, year = {2025}, }
• (2025). Content-aware preserving image generation. NEUCOM, 617, 128965. (www)

  @article{ , abstract = {Remarkable progress has been achieved in image generation with the introduction of generative models. However, precisely controlling the content in generated images remains a challenging task due to their fundamental training objective. This paper addresses this challenge by proposing a novel image generation framework explicitly designed to incorporate desired content in output images. The framework utilizes advanced encoding techniques, integrating subnetworks called content fusion and frequency encoding modules. The frequency encoding module first captures features and structures of reference images by exclusively focusing on selected frequency components. Subsequently, the content fusion module generates a content-guiding vector that encapsulates desired content features. During the image generation process, content-guiding vectors from real images are fused with projected noise vectors. This ensures the production of generated images that not only maintain consistent content from guiding images but also exhibit diverse stylistic variations. To validate the effectiveness of the proposed framework in preserving content attributes, extensive experiments are conducted on widely used benchmark datasets, including Flickr-Faces-High Quality, Animal Faces High Quality, and Large-scale Scene Understanding datasets.}, archive = {J_NEUCOM}, author = {Giang H. Le and Anh Q. Nguyen and Byeongkeun Kang and Yeejin Lee}, doi = {10.1016/j.neucom.2024.128965}, journal = {Neurocomputing}, month = {2}, pages = {128965}, shortjournal = {Neurocomputing}, title = {Content-aware preserving image generation}, volume = {617}, year = {2025}, }
• (2025). The impact of multi-class information decoupling in latent space on skin lesion segmentation. NEUCOM, 617, 128962. (www)

  @article{ , abstract = {Noise, out-of-distribution (OOD) challenges, and boundary ambiguity are not only challenges in skin disease segmentation, but also difficulties in the field of medical image segmentation. Although various medical image segmentation networks have emerged, there are certain limitations in addressing these issues because previous research has neglected the expression of medical image features in latent space. In this article, we study the relationship between latent space information and medical image segmentation. We demonstrate that the decoupling of multi-class information in latent space is closely related to the accuracy of medical image segmentation. Therefore, this study proposes a new segmentation scheme - Contextual directional decoupling network (CodeNet), which aims to enhance feature representation by decoupling multi-class information in latent space, enabling the model to successfully address noise, artifacts, OOD challenges, and boundary ambiguity. Comprehensive experiments on four public datasets and one private dataset show that our proposed CodeNet is superior to other state-of-the-art models and achieves good performance on multiple commonly used metrics. This demonstrates the effectiveness of our method in skin lesion segmentation.}, archive = {J_NEUCOM}, author = {Qibo Zhang and Chengfei Li and Song Zuo and Yufei Cai and Ao Xu and Haijian Huang and Shiqin Zhou}, doi = {10.1016/j.neucom.2024.128962}, journal = {Neurocomputing}, month = {2}, pages = {128962}, shortjournal = {Neurocomputing}, title = {The impact of multi-class information decoupling in latent space on skin lesion segmentation}, volume = {617}, year = {2025}, }
• (2025). TCM: An efficient lightweight MLP-based network with affine transformation for long-term time series forecasting. NEUCOM, 617, 128960. (www)

  @article{ , abstract = {Time series forecasting (TSF) involves extracting underlying patterns from past information to predict future sequences over a specific period. Extending the prediction length of time series and improving the prediction accuracy have always been challenging tasks. Autoregressive prediction methods based on Markov chains tend to accumulate errors over time. Although Transformer-based methods with various self-attention mechanisms have shown some improvements, they require higher memory and computational resources. In this work, we present an effective MLP-based TSF framework named TCM, which models the sequence and channel dependencies separately using Token MLP and Channel MLP. Additionally, we employ the Affine Transformation to replace layer normalization or batch normalization, leading to substantial enhancements in both accuracy and inference speed. Compared to current state-of-the-art long-term time series forecasting models, TCM achieves 6.0% relative improvement on seven real-world datasets, including electricity, weather, and illness domains. The TCM model, characterized by its efficiency and lightweight architecture, also makes it suitable for scenarios with high real-time requirements.}, archive = {J_NEUCOM}, author = {Hongwei Jiang and Dongsheng Liu and Xinyi Ding and Yaning Chen and Hongtao Li}, doi = {10.1016/j.neucom.2024.128960}, journal = {Neurocomputing}, month = {2}, pages = {128960}, shortjournal = {Neurocomputing}, title = {TCM: An efficient lightweight MLP-based network with affine transformation for long-term time series forecasting}, volume = {617}, year = {2025}, }
• (2025). Accuracy-preassigned fixed-time synchronization of switched inertial neural networks with time-varying distributed, leakage and transmission delays. NEUCOM, 617, 128958. (www)

  @article{ , abstract = {In this paper, the accuracy-preassigned fixed-time synchronization problem of a class of switched inertial neural networks with time-varying distributed, leakage and transmission delays is studied. To this end, a parameterized system solution-based direct analysis method is proposed for the first time. Unlike existing works, this method sets out from the definition of accuracy-preassigned fixed-time synchronization, and does not require variable substitution for inertial item or the construction of any Lyapunov–Krasovskii functional. This not only simplifies the proof process, but also reduces the computational complexity for solving synchronization conditions. Significantly, this paper introduced the time-varying leakage delay into switched inertial neural networks for the first time. Furthermore, the approach utilized in this manuscript stands apart from all previous techniques for achieving fixed-time synchronization. Finally, the reliability of the theoretical results is verified by numerical simulation.}, archive = {J_NEUCOM}, author = {Shilei Yuan and Yantao Wang and Xiaona Yang and Xian Zhang}, doi = {10.1016/j.neucom.2024.128958}, journal = {Neurocomputing}, month = {2}, pages = {128958}, shortjournal = {Neurocomputing}, title = {Accuracy-preassigned fixed-time synchronization of switched inertial neural networks with time-varying distributed, leakage and transmission delays}, volume = {617}, year = {2025}, }
• (2025). Fully distributed adaptive optimization event-triggered/self-triggered synchronization for multi-agent systems. NEUCOM, 617, 128954. (www)

  @article{ , abstract = {This paper investigates the distributed adaptive optimization synchronization problem of multi-agent systems (MASs) with general linear dynamics on undirected graphs. The goal is to fulfill the synchronization among agents and synergistically optimize the team cost function formed by a family of local convex functions. The tracking servo signal is first generated by sampling the implicit state of the axillary system, and its sampling events are governed by the triggering mechanism I. Meanwhile, the disagreement vector is sampled if the triggering mechanism II is violated. An adaptive event-triggered scheme is then constructed by the gradient term whose input is the tracking servo signal and the relative sampling information among followers, which fulfills the synchronization as the desired one and minimizes the team cost function. It proves that Zeno behavior is excluded under the triggering mechanisms I and II, respectively. Moreover, a self-triggered strategy is leveraged that depends only on the partial derivative of the local cost function in the implicit state sampling and the relative sampling information of itself and its neighbors; thus, continuously monitoring the information of neighbors is avoided. It is noted that the proposed scheme incorporates adaptive event-triggered control, which makes it possible to implement the fully distributed control manner. The efficacy and advantage of the presented theoretical results are finally demonstrated using a non-trivial simulation example.}, archive = {J_NEUCOM}, author = {Lina Xia and Qing Li and Ruizhuo Song}, doi = {10.1016/j.neucom.2024.128954}, journal = {Neurocomputing}, month = {2}, pages = {128954}, shortjournal = {Neurocomputing}, title = {Fully distributed adaptive optimization event-triggered/self-triggered synchronization for multi-agent systems}, volume = {617}, year = {2025}, }
• (2025). Spatiotemporal subspace variational autoencoder with repair mechanism for traffic data imputation. NEUCOM, 617, 128948. (www)

  @article{ , abstract = {High-quality spatial–temporal traffic data is crucial for the functioning of modern smart transportation systems. However, the collection and storage of traffic data in real-world scenarios are often hindered by many factors, causing data loss that greatly affects decision-making. Different modes of data absence result in varying degrees of information loss, which introduces considerable challenges to the precise imputation of traffic data. Many existing studies are concentrate on two main aspects: the examination of data distribution and the extraction of spatiotemporal relationships. On the one hand, methods that focus on distribution fitting do not require a large volume of observational data but often fail to capture spatial–temporal relationships, leading to overly smooth results. On the other hand, methods that aim to identify spatial–temporal relationships, while offering higher accuracy in fitting, demand a substantial amount of high-quality historical data. Taking into account the merits and demerits of both two paradigm, we developed a novel unsupervised two-stage model simultaneously takes into account the spatiotemporal distribution and relationships, termed Spatiotemporal Subspace Variational Autoencoder with Repair Mechanism (SVAE-R). In stage one, we introduced the concept of spatiotemporal subspace, which not only mitigates the noise impact caused by data sparsity but also reduces the cost for the model to find the distribution. In stage two, we designed a simple repair structure to capture spatial–temporal relationships among data through graph convolution network(GCN) and gated recurrent units(GRU), revising the details of the data. We have evaluated our model on two authentic datasets, and it has exhibited a high degree of robustness, maintaining effective performance even under extreme data loss conditions.}, archive = {J_NEUCOM}, author = {Jialong Qian and Shiqi Zhang and Yuzhuang Pian and Xinyi Chen and Yonghong Liu}, doi = {10.1016/j.neucom.2024.128948}, journal = {Neurocomputing}, month = {2}, pages = {128948}, shortjournal = {Neurocomputing}, title = {Spatiotemporal subspace variational autoencoder with repair mechanism for traffic data imputation}, volume = {617}, year = {2025}, }
• (2025). Graph convolutional network for fast video summarization in compressed domain. NEUCOM, 617, 128945. (www)

  @article{ , abstract = {Video summarization is the process of generating a concise and representative summary of a video by selecting its most important frames. It plays a vital role in the video streaming industry, allowing users to quickly understand the overall content of a video without watching it in its entirety. Most existing video summarization methods require fully decoding the video stream and extracting the features with a pre-trained deep learning model in the pixel domain, which is time-consuming and computationally expensive. To address this issue, this paper proposes a novel method called Graph Convolutional Network-based Compressed-domain Video Summarization (GCNCVS), which directly exploits the compressed-domain information and leverages graph convolutional network to learn temporal relationships between frames, thereby enhancing its ability to capture contextual and valuable information when generating summarized videos. To evaluate the performance of GCNCVS, we conduct experiments on two benchmark datasets, SumMe and TVSum. Experimental results demonstrate that our method outperforms existing methods, achieving an average F-score of 53.5% on the SumMe dataset and 72.3% on the TVSum dataset. Additionally, the proposed method shows Kendall's τ correlation coefficient of 0.157 and Spearman's ρ correlation coefficient of 0.205 on the TVSum dataset. Our method also significantly reduces computational time, which enhances the feasibility of video summarization in video streaming environments.}, archive = {J_NEUCOM}, author = {Chia-Hung Yeh and Chih-Ming Lien and Zhi-Xiang Zhan and Feng-Hsu Tsai and Mei-Juan Chen}, doi = {10.1016/j.neucom.2024.128945}, journal = {Neurocomputing}, month = {2}, pages = {128945}, shortjournal = {Neurocomputing}, title = {Graph convolutional network for fast video summarization in compressed domain}, volume = {617}, year = {2025}, }
• (2025). Multimodal sentiment analysis based on disentangled representation learning and cross-modal-context association mining. NEUCOM, 617, 128940. (www)

  @article{ , abstract = {Multimodal sentiment analysis aims to extract sentiment information expressed by users from multimodal data, including linguistic, acoustic, and visual cues. However, the heterogeneity of multimodal data leads to disparities in modal distribution, thereby impacting the model’s ability to effectively integrate complementarity and redundancy across modalities. Additionally, existing approaches often merge modalities directly after obtaining their representations, overlooking potential emotional correlations between them. To tackle these challenges, we propose a Multiview Collaborative Perception (MVCP) framework for multimodal sentiment analysis. This framework consists primarily of two modules: Multimodal Disentangled Representation Learning (MDRL) and Cross-Modal Context Association Mining (CMCAM). The MDRL module employs a joint learning layer comprising a common encoder and an exclusive encoder. This layer maps multimodal data to a hypersphere, learning common and exclusive representations for each modality, thus mitigating the semantic gap arising from modal heterogeneity. To further bridge semantic gaps and capture complex inter-modal correlations, the CMCAM module utilizes multiple attention mechanisms to mine cross-modal and contextual sentiment associations, yielding joint representations with rich multimodal semantic interactions. In this stage, the CMCAM module only discovers the correlation information among the common representations in order to maintain the exclusive representations of different modalities. Finally, a multitask learning framework is adopted to achieve parameter sharing between single-modal tasks and improve sentiment prediction performance. Experimental results on the MOSI and MOSEI datasets demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Zuhe Li and Panbo Liu and Yushan Pan and Weiping Ding and Jun Yu and Haoran Chen and Weihua Liu and Yiming Luo and Hao Wang}, doi = {10.1016/j.neucom.2024.128940}, journal = {Neurocomputing}, month = {2}, pages = {128940}, shortjournal = {Neurocomputing}, title = {Multimodal sentiment analysis based on disentangled representation learning and cross-modal-context association mining}, volume = {617}, year = {2025}, }
• (2025). Deep echo state network with projection-encoding for multi-step time series prediction. NEUCOM, 617, 128939. (www)

  @article{ , abstract = {To fully utilize the advantage of reservoir computing in deep network modeling, a deep echo state network with projection-encoding (DEESN) is newly proposed for multi-step time series prediction in this paper. DEESN integrates multiple echo state network (ESN) modules and extreme learning machine (ELM) encoders in series arrays. Firstly, the k th ESN in DEESN learner is responsible for k th step ahead prediction. The forecast output and encoded reservoir states of the previous ESN module are concatenated with the input variable to form the new input signals for the next adjacent module. Therefore, the temporal dependency among future time steps can be learned, which contributes the performance improvement. Secondly, the ELM encoder is used to optimize the reservoir states for time consumption reduction. Finally, the effectiveness of DEESN is evaluated in artificial chaos benchmarks and real-world applications. Experimental results on six different datasets and comparative models demonstrate that the proposed DEESN has excellent accuracy and robust generalization for multi-step time series prediction.}, archive = {J_NEUCOM}, author = {Tao Li and Zhijun Guo and Qian Li}, doi = {10.1016/j.neucom.2024.128939}, journal = {Neurocomputing}, month = {2}, pages = {128939}, shortjournal = {Neurocomputing}, title = {Deep echo state network with projection-encoding for multi-step time series prediction}, volume = {617}, year = {2025}, }
• (2025). Temporal convolution derived multi-layered reservoir computing. NEUCOM, 617, 128938. (www)

  @article{ , abstract = {The prediction of time series is a challenging task relevant in such diverse applications as analyzing financial data, forecasting flow dynamics or understanding biological processes. Especially chaotic time series that depend on a long history pose an exceptionally difficult problem. While machine learning has shown to be a promising approach for predicting such time series, it either demands long training time and much training data when using deep Recurrent Neural Networks. Alternative, when using a Reservoir Computing approach it comes with high uncertainty and typically a high number of random initializations and extensive hyper-parameter tuning. In this paper, we focus on the Reservoir Computing approach and propose a new mapping of input data into the reservoir’s state space. Furthermore, we incorporate this method in two novel network architectures increasing parallelizability, depth and predictive capabilities of the neural network while reducing the dependence on randomness. For the evaluation, we approximate a set of time series from the Mackey–Glass equation, inhabiting non-chaotic as well as chaotic behavior as well as the SantaFe Laser dataset and compare our approaches in regard to their predictive capabilities to Echo State Networks, Autoencoder connected Echo State Networks and Gated Recurrent Units. For the chaotic time series, we observe an error reduction of up to 85.45% compared to Echo State Networks and 90.72% compared to Gated Recurrent Units. Furthermore, we also observe tremendous improvements for non-chaotic time series of up to 99.99% in contrast to the existing approaches.}, archive = {J_NEUCOM}, author = {Johannes Viehweg and Dominik Walther and Patrick Mäder}, doi = {10.1016/j.neucom.2024.128938}, journal = {Neurocomputing}, month = {2}, pages = {128938}, shortjournal = {Neurocomputing}, title = {Temporal convolution derived multi-layered reservoir computing}, volume = {617}, year = {2025}, }
• (2025). Physics embedded neural network: Novel data-free approach towards scientific computing and applications in transfer learning. NEUCOM, 617, 128936. (www)

  @article{ , abstract = {This paper introduces a novel approach for solving partial differential equations using neural networks for scientific computing applications. The proposed approach is referred to as physics embedded neural network (or PENN) due to its unique modified architecture designed to contain the partial differential equation information within the neural network, which is built upon a feed-forward neural network (NN) with fully connected layers. The key aspect of PENN is the numerical embedding of a differential equation associated with physical problem in the penultimate layer of the neural network. Here we propose that through such integration of PDE to the neural network’s penultimate layer, we can produce a fast numerical solution of the PDE, and with time performance comparable with Finite Element Method (FEM). Due to direct solution of the PDE, the NN solution occurs without training and testing phases which are normally applied when using NN in a data-based training. To demonstrate the effectiveness of the proposed architecture, we investigate variety of cases focused around second order in-homogeneous PDE in 2 dimensional (2D) problems. Additionally, we demonstrate the effectiveness of our approach by applying it to solve electromagnetic problem to illustrate its engineering applications. In this study, we also conducted a comparison of solution time and error achieved using our proposed network with a classical finite element method (FEM) solver. The convergence was found exponential with number of epochs proving an error significantly lower error compared to FEM and reported in other NN based algorithms such as Physics Informed Neural Networks. Furthermore, we explored the application of PENN in transfer learning. Our findings showed that by using transfer learning, up to 5 times additional reduction in simulation time is possible as compared to a direct solution through PENN. Overall, the study contributes to the development of efficient and accurate methods for solving PDEs through neural network optimization and provides important insights into the potential benefits of incorporating transfer learning techniques in numerical modeling of physical phenomenology.}, archive = {J_NEUCOM}, author = {Pawan Gaire and Shubhendu Bhardwaj}, doi = {10.1016/j.neucom.2024.128936}, journal = {Neurocomputing}, month = {2}, pages = {128936}, shortjournal = {Neurocomputing}, title = {Physics embedded neural network: Novel data-free approach towards scientific computing and applications in transfer learning}, volume = {617}, year = {2025}, }
• (2025). Detection of video transcoding from AVC to HEVC based on intra prediction feature maps. NEUCOM, 617, 128935. (www)

  @article{ , abstract = {As the High Efficiency Video Coding (HEVC) standard gains popularity, forgers are more inclined to transcode videos into HEVC format from the previous Advanced Video Coding (AVC) format. To verify the originality and authenticity of videos, it is crucial to propose a method for transcoded HEVC video detection. In this paper, a novel method is proposed to detect video transcoding from AVC to HEVC (AVC-HEVC). Analysis shows that the intra prediction mode is sensitive to spatial loss introduced by previous AVC encoding in the transcoding process. Thus, the intra prediction modes are extracted from the luminance and chrominance components to create Intra Prediction Feature Maps (IPFMs). Subsequently, a Dual-flow Attention-based MobileNet (DAM-Net) is introduced to learn the deep representation of AVC-HEVC transcoding artifacts. Finally, video level results are derived from the frame level analysis provided by DAM-Net. Extensive experiment results demonstrate that the performance of the proposed method outperforms the existing methods in the detection of AVC-HEVC transcoding.}, archive = {J_NEUCOM}, author = {Yueneng Wang and Zhongjie Mi and Xinghao Jiang and Tanfeng Sun}, doi = {10.1016/j.neucom.2024.128935}, journal = {Neurocomputing}, month = {2}, pages = {128935}, shortjournal = {Neurocomputing}, title = {Detection of video transcoding from AVC to HEVC based on intra prediction feature maps}, volume = {617}, year = {2025}, }
• (2025). Novelty-aware concept drift detection for neural networks. NEUCOM, 617, 128933. (www)

  @article{ , abstract = {Neural network models are widely adopted in real-world applications for processing streaming data. However, these applications often face challenges in terms of accuracy degradation, caused by changes in the data distribution of the stream data compared to the training data. Two underlying reasons contribute to these changes. The first, known as the concept drift problem, occurs when there is a change in the correlation between the input data and the prediction output, making the models trained on the training data no longer suitable for the new data. The second reason, known as the novelty problem, arises when real-world data contains unexpected data categories that were not present in the training data, resulting in incorrect predictions. The research community has divided into different groups and each developed various methods to detect either concept drift or novelty distribution changes. However, these methods only address one aspect of the problem and are unable to distinguish between them. This leads to an inappropriate allocation of model maintenance resources, including the high cost of model retraining and the acquisition of true label data. In this study, we aim to address this gap by proposing a novel concept drift detection method that is capable of distinguishing between known labeled concept drift and novelty. Our method is also more efficient than existing drift detection methods, making it suitable for applications on neural networks.}, archive = {J_NEUCOM}, author = {Dan Shang and Guangquan Zhang and Jie Lu}, doi = {10.1016/j.neucom.2024.128933}, journal = {Neurocomputing}, month = {2}, pages = {128933}, shortjournal = {Neurocomputing}, title = {Novelty-aware concept drift detection for neural networks}, volume = {617}, year = {2025}, }
• (2025). SAM-assisted temporal-location enhanced transformer segmentation for object tracking with online motion inference. NEUCOM, 617, 128914. (www)

  @article{ , abstract = {Current transformer-based trackers typically represent targets using bounding boxes. However, bounding boxes do not accurately describe the target and uncontrollably contain most background pixels. This paper proposes a Segment Anything Model (SAM)-Assisted Temporal-Location Enhanced Transformer Segmentation for Object Tracking with Online Motion Inference. First, a novel transformer-based temporal-location enhanced segmentation method is proposed. The target temporal features are clustered into foreground–background tokens utilizing a mask to capture discriminative information distribution. Then, the suitable positional prompts are learned in the proposed mask prediction head to establish the mapping between target features and localization, which enhances the specific foreground weights for precise mask generation. Second, a temporal-based motion inference module is proposed. It fully utilizes the target temporal state to construct an online displacement model inferring motion relationships of the target between frames and providing robust position prompts for the segmentation process. We also introduce SAM for initial mask generation. Precise pixel-level object tracking is achieved by combining segmentation and localization within a unified process. Experimental results demonstrate that the proposed method yields competitive performance compared to existing approaches.}, archive = {J_NEUCOM}, author = {Huanlong Zhang and Xiangbo Yang and Xin Wang and Weiqiang Fu and Bineng Zhong and Jianwei Zhang}, doi = {10.1016/j.neucom.2024.128914}, journal = {Neurocomputing}, month = {2}, pages = {128914}, shortjournal = {Neurocomputing}, title = {SAM-assisted temporal-location enhanced transformer segmentation for object tracking with online motion inference}, volume = {617}, year = {2025}, }
• (2025). TF4TF: Multi-semantic modeling within the time–frequency domain for long-term time-series forecasting. NEUCOM, 617, 128913. (www)

  @article{ , abstract = {Long-term Time Series Forecasting (LTSF) plays a crucial role in real-world applications for early warning and decision-making. Time series inherently embody complex semantic information, including segment semantics, global–local semantics, and multi-view semantics, the thorough mining of which can significantly enhance the accuracy. Previous works have not been able to simultaneously address all of the semantic information mentioned above. Meanwhile, the thorough mining of semantic information introduces additional computational complexity, resulting in inefficiency issues for existing multi-semantic information mining methods. Considering the aforementioned situation, we propose a multi-semantic method within the Time–Frequency domain For long-term Time-series Forecasting (TF4TF), which can balance complex semantic information mining and efficiency. For sequences with segment semantics following patching process, mining is conducted from both time and frequency domain perspectives to extract Multi-View Semantics. Within this framework, Progressive Local Windows (PLW) blocks and Global Frequency Filtering (GFF) blocks are specifically designed, which achieve efficient mining of multi-scale information while maintaining lower complexity. Ultimately, forecasting is achieved by integrating the semantic information outlined above. Our proposed method, TF4TF, has achieved state-of-the-art (SOTA) results on seven real-world time series forecasting datasets.}, archive = {J_NEUCOM}, author = {Xueer Zhang and Jing Wang and Yunfei Bai and Lu Zhang and Youfang Lin}, doi = {10.1016/j.neucom.2024.128913}, journal = {Neurocomputing}, month = {2}, pages = {128913}, shortjournal = {Neurocomputing}, title = {TF4TF: Multi-semantic modeling within the time–frequency domain for long-term time-series forecasting}, volume = {617}, year = {2025}, }
• (2025). Diffusion models for image super-resolution: State-of-the-art and future directions. NEUCOM, 617, 128911. (www)

  @article{ , abstract = {The single image super-resolution (SISR) task has received much attention due to the wide range of applications in many tasks. The progress in this SISR is mainly based on deep learning methods. In recent years, many methods have been developed based on using diffusion models to solve the SISR task. The performance of the diffusion-based model depends on many factors, such as the type of diffusion, noise schedule, and loss function. Based on this limitation, we developed a review paper about the state-of-the-art diffusion models, especially ones that were developed for SR. In this survey, the SR-based diffusion models are classified based on four metrics, namely the diffusion model type, the noise schedule, the loss function, and the used datasets. Then, we discuss the main categories in each metric and the developed diffusion models related to each category. Three main diffusion model types are considered, including denoising diffusion probabilistic models (DDPM), noise-conditioned score networks (NCSN), and stochastic differential equations models (SDEM), in addition to other diffusion models. In the end, we discuss the limitations of these diffusion models and show our expectations for future research in this domain.}, archive = {J_NEUCOM}, author = {Garas Gendy and Guanghui He and Nabil Sabor}, doi = {10.1016/j.neucom.2024.128911}, journal = {Neurocomputing}, month = {2}, pages = {128911}, shortjournal = {Neurocomputing}, title = {Diffusion models for image super-resolution: State-of-the-art and future directions}, volume = {617}, year = {2025}, }
• (2025). IIRNet: Infinite impulse response inspired network for compressed video quality enhancement. NEUCOM, 617, 128909. (www)

  @article{ , abstract = {Deep learning approaches have demonstrated notable advancements in a compressed video quality enhancement (CVQE) task. Nevertheless, they inadequately account for the loss of correlation detail information between features caused by the large motion in the video sequence and fail to comprehensively address the impact of diverse compression artifacts on the inaccurate capture of inter-frame dependent information. To address these issues, we find that the process of enhancing the input videos to restore high-quality video in the CVQE task can be analogized to the process of enhancing the input signals to emphasize useful signals in the digital signal processing (DSP) domain. In this paper, we present an infinite impulse response inspired network (IIRNet) for CVQE. Specifically, we propose to employ a design scheme inspired by the five-point difference format (FDF) of IIR and develop a multilevel 3D-FDF cycle module to obtain rich detailed feature information at multiple levels to maintain a strong correlation between features. Furthermore, we devise a cross-domain point fusion attention module to extract dissimilar information from the frequency domain of adjacent frames by non-local dissimilarity technique, especially in the region prone to artifacts, for efficient processing. The proposed IIRNet is evaluated on the public MFQE 2.0 dataset, demonstrating effectiveness in enhancing the quality of compressed videos by removing compression artifacts and reducing noise. Compared with the state-of-the-art (SOTA) methods, the proposed method achieves better performance in both objective metrics and visual quality.}, archive = {J_NEUCOM}, author = {Mingjin Zhang and Lingping Zheng and Yunsong Li and Xinbo Gao}, doi = {10.1016/j.neucom.2024.128909}, journal = {Neurocomputing}, month = {2}, pages = {128909}, shortjournal = {Neurocomputing}, title = {IIRNet: Infinite impulse response inspired network for compressed video quality enhancement}, volume = {617}, year = {2025}, }
• (2025). Identifying local useful information for attribute graph anomaly detection. NEUCOM, 617, 128900. (www)

  @article{ , abstract = {Graph anomaly detection primarily relies on shallow learning methods based on feature engineering and deep learning strategies centred on autoencoder-based reconstruction. However, these methods frequently fail to harness the local attributes and structural information within graph data, making it challenging to capture the underlying distribution in scenarios with class-imbalanced graph anomalies, which can result in overfitting. To deal with the above issue, this paper proposes a new anomaly detection method called LIAD (Identifying L ocal Useful I nformation for A ttribute Graph Anomaly D etection), which learns the data’s underlying distribution and captures richer local information. First, LIAD employs data augmentation techniques to create masked graphs and pairs of positive and negative subgraphs. Then, LIAD leverages contrastive learning to derive rich embedding representations from diverse local structural information. Additionally, LIAD utilizes a variational autoencoder (VAE) to generate new graph data, capturing the neighbourhood distribution within the masked graph. During the training process, LIAD aligns the generated graph data with the original to deepen its comprehension of local information. Finally, anomaly scoring is achieved by comparing the discrimination and reconstruction scores of the contrastive pairs, enabling effective anomaly detection. Extensive experiments on five real-world datasets demonstrate the effectiveness of LIAD compared to state-of-the-art methods. Comprehensive ablation studies and parametric analyses further affirm the robustness and efficacy of our model.}, archive = {J_NEUCOM}, author = {Penghui Xi and Debo Cheng and Guangquan Lu and Zhenyun Deng and Guixian Zhang and Shichao Zhang}, doi = {10.1016/j.neucom.2024.128900}, journal = {Neurocomputing}, month = {2}, pages = {128900}, shortjournal = {Neurocomputing}, title = {Identifying local useful information for attribute graph anomaly detection}, volume = {617}, year = {2025}, }
• (2025). Cross-modal attention and geometric contextual aggregation network for 6DoF object pose estimation. NEUCOM, 617, 128891. (www)

  @article{ , abstract = {The availability of affordable RGB-D sensors has made it more suitable to use RGB-D images for accurate 6D pose estimation, which allows for precise 6D parameter prediction using RGB-D images while maintaining a reasonable cost. A crucial research challenge is effectively exploiting adaptive feature extraction and fusion from the appearance information of RGB images and the geometric information of depth images. Moreover, previous methods have neglected the spatial geometric relationships of local position and the properties of point features, which are beneficial for tackling pose estimation in occlusion scenarios. In this work, we propose a cross-attention fusion framework for learning 6D pose estimation from RGB-D images. During the feature extraction stage, we design a geometry-aware context network that encodes local geometric properties of objects in point clouds using dual criteria, distance, and geometric angles. Moreover, we propose a cross-attention framework that combines spatial and channel attention in a cross-modal attention manner. This innovative framework enables us to capture the correlation and importance between RGB and depth features, resulting in improved accuracy in pose estimation, particularly in complex scenes. In the experimental results, we demonstrated that the proposed method outperforms state-of-the-art methods on four challenging benchmark datasets: YCB-Video, LineMOD, Occlusion LineMOD, and MP6D. Video is available at https://youtu.be/4mgdbQKaHOc .}, archive = {J_NEUCOM}, author = {Yi Guo and Fei Wang and Hao Chu and Shiguang Wen}, doi = {10.1016/j.neucom.2024.128891}, journal = {Neurocomputing}, month = {2}, pages = {128891}, shortjournal = {Neurocomputing}, title = {Cross-modal attention and geometric contextual aggregation network for 6DoF object pose estimation}, volume = {617}, year = {2025}, }
• (2025). Locality-constrained robust discriminant non-negative matrix factorization for depression detection: An fNIRS study. NEUCOM, 617, 128887. (www)

  @article{ , abstract = {Major depressive disorder (MDD) is having an increasingly severe impact worldwide, which creates a pressing need for an efficient and objective method of depression detection. Functional near-infrared spectroscopy (fNIRS), which directly monitors changes in cerebral oxygenation, has become an important tool in depression research. Currently, feature extraction methods based on multi-channel fNIRS data often overlook the local structure of the data and the subsequent classification cost. To address these challenges, we introduce an innovative feature extraction algorithm, namely locality-constrained robust discriminant non-negative matrix factorization (LRDNMF). The algorithm incorporates ℓ 2 , 1 regularization, local coordinate constraints, within-class scatter distance, and total scatter distance, achieving a fusion of robustness, locality, and discrimination. LRDNMF enhances feature representation, reduces noise impact, and significantly boosts classification ability. Based on experimental results from 56 participants, LRDNMF achieves an accuracy of 90.55%, a recall of 91.48%, a precision of 90.46%, and an F1 score of 0.91 under full stimuli. These results outperform existing algorithms, validating the effectiveness of LRDNMF and demonstrating its significant potential in auxiliary diagnosis of depression.}, archive = {J_NEUCOM}, author = {Yushan Wu and Jitao Zhong and Lu Zhang and Hele Liu and Shuai Shao and Bin Hu and Hong Peng}, doi = {10.1016/j.neucom.2024.128887}, journal = {Neurocomputing}, month = {2}, pages = {128887}, shortjournal = {Neurocomputing}, title = {Locality-constrained robust discriminant non-negative matrix factorization for depression detection: An fNIRS study}, volume = {617}, year = {2025}, }
• (2025). Viewpoint‐Based collaborative Feature‐Weighted Multi‐View intuitionistic fuzzy clustering using neighborhood information. NEUCOM, 617, 128884. (www)

  @article{ , abstract = {This paper presents an intuitionistic fuzzy c -means-based clustering algorithm for multi-view clustering, addressing key challenges such as noise sensitivity, outlier influence, and the distinct importance of views, features, and samples. Our proposed approach incorporates view weights, feature weights, sample weights, and neighborhood information into a novel objective function. Additionally, we introduce an effective initial cluster center selection strategy that enhances clustering robustness. The efficiency of the proposed method is evaluated using various clustering criteria ( AR, NMI, RI, FMI, and JI ). Moreover, the effect of each module of the algorithm on the general clustering performance is examined exclusively. Experimental results on various benchmark multi-view datasets demonstrate that our algorithm outperforms state-of-the-art methods in terms of clustering accuracy and stability. The source code of the proposed method is accessible at https://github.com/Amin-Golzari-Oskouei/VCoFWMVIFCM.}, archive = {J_NEUCOM}, author = {Amin Golzari Oskouei and Negin Samadi and Jafar Tanha and Asgarali Bouyer and Bahman Arasteh}, doi = {10.1016/j.neucom.2024.128884}, journal = {Neurocomputing}, month = {2}, pages = {128884}, shortjournal = {Neurocomputing}, title = {Viewpoint‐Based collaborative Feature‐Weighted Multi‐View intuitionistic fuzzy clustering using neighborhood information}, volume = {617}, year = {2025}, }
• (2025). Dynamic event triggering output feedback synchronization for markov jump neural networks with mode detection information. NEUCOM, 617, 128872. (www)

  @article{ , abstract = {This article investigates the synchronization control problem of discrete-time Markov jump neural networks. Because of the possible mismatch of controller mode information and the difficulty in obtaining neuron information in practical environments, a hidden Markov model is introduced, which contains a partially unknown detection probability matrix and a partially unknown transition probability matrix. To overcome the unpredictability of the system state and enhance the effective utilization of communication resources, a static output feedback controller based on a dynamic event triggering strategy is designed. Moreover, the conservatism of theoretical derivation is further reduced through the activation function division. Finally, numerical examples are used to verify the reliability of the above results, which are then applied to image encryption.}, archive = {J_NEUCOM}, author = {Cheng Fan and Ling Jin and Lei Su and Xihong Fei}, doi = {10.1016/j.neucom.2024.128872}, journal = {Neurocomputing}, month = {2}, pages = {128872}, shortjournal = {Neurocomputing}, title = {Dynamic event triggering output feedback synchronization for markov jump neural networks with mode detection information}, volume = {617}, year = {2025}, }
• (2025). GPTGAN: Utilizing the GPT language model and GAN to enhance adversarial text generation. NEUCOM, 617, 128865. (www)

  @article{ , abstract = {Training generative models that can generate high-quality and diverse text remains a significant challenge in the field of natural language generation (NLG). Recently, the emergence of large language models (LLMs) like GPT has enabled the generation of text with remarkable quality and diversity. However, building these models from scratch is both time-consuming and resource-intensive, making their comprehensive training practically unfeasible. Nonetheless, LLMs utility extends to addressing issues in other models. For instance, generative adversarial models often grapple with the well-known problem of mode collapse during training, leading to a trade-off between text quality and diversity. This means that these models tend to favor quality over diversity. In this study, we introduce a novel approach designed to enhance adversarial text generation by striking a balance between the quality and diversity of generated text, leveraging the capabilities of the GPT language model and other LLMs. To achieve this, we propose an enhanced generator that is guided by the GPT model. Essentially, the GPT model functions as a mentor to the generator, influencing its outputs. To achieve this guidance, we employ discriminators of varying scales on both real data and the texts generated by GPT. Experimental results underscore a substantial enhancement in the quality and diversity of outcomes across two benchmark datasets. Also the results demonstrate the generator’s ability to assimilate the output domain of the GPT language model. Furthermore, the proposed model exhibits superior performance in human evaluations when compared to other existing adversarial methods.}, archive = {J_NEUCOM}, author = {Omid Hajipoor and Ahmad Nickabadi and Mohammad Mehdi Homayounpour}, doi = {10.1016/j.neucom.2024.128865}, journal = {Neurocomputing}, month = {2}, pages = {128865}, shortjournal = {Neurocomputing}, title = {GPTGAN: Utilizing the GPT language model and GAN to enhance adversarial text generation}, volume = {617}, year = {2025}, }
• (2025). Distantly supervised relation extraction with a meta-relation enhanced contrastive learning framework. NEUCOM, 617, 128864. (www)

  @article{ , abstract = {Distantly supervised relation extraction employs the alignment of unstructured corpora with knowledge bases to automatically generate labeled data. This method, however, often introduces significant label noise. To address this, multi-instance learning has been widely utilized over the past decade, aiming to extract reliable features from a bag of sentences. Yet, multi-instance learning struggles to effectively distinguish between clean and noisy instances within a bag, thereby hindering the full utilization of informative instances and the reduction of the impact of incorrectly labeled instances. In this paper, we propose a new Meta-Relation enhanced Contrastive learning based method for distantly supervised Relation Extraction named MRConRE. Specifically, we generate a “meta relation pattern” ( MRP ) for each bag, based on its semantic content, to differentiate between clean and noisy instances. Noisy instances are then transformed into beneficial bag-level instances through relabeling. Subsequently, contrastive learning is employed to develop precise sentence representations, forming the overall representation of the bag. Finally, we utilize a mixup strategy to integrate bag-level information for model training. Our method’s effectiveness is validated through experiments on various benchmarks.}, archive = {J_NEUCOM}, author = {Chuanshu Chen and Shuang Hao and Jian Liu}, doi = {10.1016/j.neucom.2024.128864}, journal = {Neurocomputing}, month = {2}, pages = {128864}, shortjournal = {Neurocomputing}, title = {Distantly supervised relation extraction with a meta-relation enhanced contrastive learning framework}, volume = {617}, year = {2025}, }
• (2025). Unifying domain gap in federated learning: A geometric approach. NEUCOM, 617, 128813. (www)

  @article{ , abstract = {Federated Learning (FL) is a framework designed to enable distributed and privacy-preserving learning. It typically involves a central server coordinating multiple clients to collaboratively train a global model without the need for sharing private data. Recent studies have revealed that this parameter-only communication strategy can be hindered by non-independent and identically distributed (non-iid) data across clients, leading to statistical heterogeneity. This issue necessitates the development of Personalized Federated Learning (PFL), which aims to enhance the performance of individual clients while still allowing for the learning of a global model. Most existing work on PFL has focused on label shift in supervised tasks, with relatively little attention given to feature shift, another common problem in real-world applications of federated learning. For instance, client data may vary in terms of illumination, angles, and quality for natural images, or in imaging protocols for medical images. To address this issue, we propose a general yet powerful framework that leverages the geometric interpretation of deep neural networks. Geometrically, a deep neural network partitions the feature space into a Power Diagram (PD) or Voronoi Diagram (VD). Regularized losses are introduced to mitigate feature shift from the perspective of PD and VD. Additionally, a personalized generator network is formulated for each client to rectify feature distribution and adapt the input data from its own domain to an ”averaged” or unified domain that generalizes well across all clients. Our method demonstrates competitive performance compared to current state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Mingxi Lei and Meng Ding and Chunwei Ma and Ziyun Huang and Mingchen Gao and Jinhui Xu}, doi = {10.1016/j.neucom.2024.128813}, journal = {Neurocomputing}, month = {2}, pages = {128813}, shortjournal = {Neurocomputing}, title = {Unifying domain gap in federated learning: A geometric approach}, volume = {617}, year = {2025}, }
• (2025). Finite-time-convergent support vector neural dynamics for classification. NEUCOM, 617, 128810. (www)

  @article{ , abstract = {Support vector machine (SVM) is a popular binary classification algorithm widely utilized in various fields due to its accuracy and versatility. However, most of the existing research involving SVMs stays at the application level, and there is few research on optimizing the support vector solving process. Therefore, it is an alternative way to optimize the support vector solving process for improving the classification performance via constructing new solving methods. Recent research has demonstrated that neural dynamics exhibit robust solving performance and high accuracy. Motivated by this inspiration, this paper leverages neural dynamics to improve the accuracy and robustness of SVM solutions. Specifically, this paper models the solving process of SVM as a standard quadratic programming (QP) problem. Then, a support vector neural dynamics (SVND) model is specifically developed to provide the optimal solution to the aforementioned QP problem, with theoretical analysis confirming its ability to achieve global convergence. Datasets of varying sizes from various sources are employed to validate the effectiveness of the designed SVND model. Experimental results show that the designed SVND model demonstrates superior classification accuracy and robustness compared to other classical machine learning algorithms. The source code is available at https://github.com/LongJin-lab/NC_SVND .}, archive = {J_NEUCOM}, author = {Mei Liu and Qihai Jiang and Hui Li and Xinwei Cao and Xin Lv}, doi = {10.1016/j.neucom.2024.128810}, journal = {Neurocomputing}, month = {2}, pages = {128810}, shortjournal = {Neurocomputing}, title = {Finite-time-convergent support vector neural dynamics for classification}, volume = {617}, year = {2025}, }
• (2025). DCM_MCCKF: A non-gaussian state estimator with adaptive kernel size based on CS divergence. NEUCOM, 617, 128809. (www)

  @article{ , abstract = {In practical industrial system models, noise is better characterized by heavy-tailed non-Gaussian distributions. For state estimation in systems with heavy-tailed non-Gaussian noise, the maximum correntropy criterion (MCC) based on information theoretic learning (ITL) is widely adopted, achieving good filtering performance. The performance of MCC-based filtering depends on the selection of the kernel function and its parameters. To overcome the sensitivity of the Gaussian kernel to its parameters and the limitation of a single kernel function in comprehensively reflecting the characteristics of complex heterogeneous data, a double-Cauchy mixture-based MCC Kalman Filtering (DCM_MCCKF) algorithm is proposed. This selection uses a mixture of two Cauchy kernel functions, using their heavy-tailed properties to better handle large errors and reduce sensitivity to kernel size variations. As a result, it improves the robustness and flexibility of MCC-based filtering. The kernel size should adapt to changes in signal distribution. To address the limitation of fixed kernel size, an adaptive kernel size update rule is designed by comprehensively considering system models, accessible measurements, CS divergence between noise distributions, and covariance propagation. Simulation examples of target tracking validate that the proposed DCM_MCCKF algorithm, under the adaptive kernel size updating rule, effectively handles complex data and achieves superior filtering performance in heavy-tailed non-Gaussian noise scenarios. This algorithm outperforms traditional Kalman filters (KF) based on the mean square error (MSE) criterion, Gaussian sum filtering (GSF), particle filtering (PF), and Maximum Correntropy Criterion Kalman filters (MCCKF) with a single Gaussian kernel (G_MCCKF), a double-Gaussian mixture kernel (DGM_MCCKF), and a Gaussian-Cauchy mixture kernel (GCM_MCCKF). Consequently, the DCM_MCCKF algorithm significantly enhances the applicability and robustness of MCC-based filtering methods.}, archive = {J_NEUCOM}, author = {Xuefei Bai and Quanbo Ge and Pingliang(Peter) Zeng}, doi = {10.1016/j.neucom.2024.128809}, journal = {Neurocomputing}, month = {2}, pages = {128809}, shortjournal = {Neurocomputing}, title = {DCM_MCCKF: A non-gaussian state estimator with adaptive kernel size based on CS divergence}, volume = {617}, year = {2025}, }
• (2025). A collaborative filtering recommender systems: Survey. NEUCOM, 617, 128718. (www)

  @article{ , abstract = {In the current digital landscape, both information consumers and producers encounter numerous challenges, underscoring the importance of recommender systems (RS) as a vital tool. Among various RS techniques, collaborative filtering (CF) has emerged as a highly effective method for suggesting products and services. However, traditional CF methods face significant obstacles in the era of big data, including issues related to data sparsity, accuracy, cold start problems, and high dimensionality. This paper offers a comprehensive survey of CF-based RS enhanced by machine learning (ML) and deep learning (DL) algorithms. It aims to serve as a valuable resource for both novice and experienced researchers in the field of RS. The survey is structured into two main sections: the first elucidates the fundamental concepts of RS, while the second delves into solutions for CF-based RS challenges, examining the specific tasks addressed by various studies, as well as the metrics and datasets employed.}, archive = {J_NEUCOM}, author = {Mohammed Fadhel Aljunid and Manjaiah D.H. and Mohammad Kazim Hooshmand and Wasim A. Ali and Amrithkala M. Shetty and Sadiq Qaid Alzoubah}, doi = {10.1016/j.neucom.2024.128718}, journal = {Neurocomputing}, month = {2}, pages = {128718}, shortjournal = {Neurocomputing}, title = {A collaborative filtering recommender systems: Survey}, volume = {617}, year = {2025}, }
• (2025). Deep learning and pre-training technology for encrypted traffic classification: A comprehensive review. NEUCOM, 617, 128444. (www)

  @article{ , abstract = {Network traffic classification has long been a pivotal topic in network security. In the past two decades, methods like port-based classification, deep packet inspection, and machine learning approaches have significantly progressed. Still, they are now facing reduced effectiveness due to the evolving complexity of the Internet, new encryption protocols, and advanced defense strategies. Given the problem that traditional models cannot efficiently generalize encrypted traffic, two promising technology paths are currently: deep learning and pre-training. On the one hand, deep learning-based methods effectively dissect complex network structures and unearth pivotal relational patterns. These approaches excel due to the neural networks’ robust generalization capabilities, significantly boosting the accuracy and efficiency of recognition processes. Graph representation learning stands out as the most compelling contemporary model for such intricate analysis, adeptly revealing the critical relationships within network communication structures. We emphatically introduce mainstream deep learning-based methods, and the mechanism and scenarios are also analyzed. On the other hand, recognizing that although the analysis based on large models is the trend of the field, the application is truly limited now, we underscore the importance of pre-training, which aligns with the future trajectory towards the adoption of large-scale models in encrypted traffic analysis. The pre-trained model can overcome various defects of previous models and achieve more remarkable performance through its low labeled data dependency and strong scenario adaptability. We provide a comprehensive overview of existing pre-training-based approaches from the three stages of operation: pre-processing, pre-training, fine-tuning, and comparing representative relevant work. Finally, because of the current needs and the improvement space of the existing pre-training methods in the field, we synthetically analyze the challenges and opportunities for interested researchers to explore.}, archive = {J_NEUCOM}, author = {Wenqi Dong and Jing Yu and Xinjie Lin and Gaopeng Gou and Gang Xiong}, doi = {10.1016/j.neucom.2024.128444}, journal = {Neurocomputing}, month = {2}, pages = {128444}, shortjournal = {Neurocomputing}, title = {Deep learning and pre-training technology for encrypted traffic classification: A comprehensive review}, volume = {617}, year = {2025}, }
• (2025). Fewer interpretable bands via self-supervision for hyperspectral anomaly detection. NEUCOM, 616, 128964. (www)

  @article{ , abstract = {Hyperspectral band selection (BS) algorithms aim to identify the most important bands relevant to model decisions. While existing methods can effectively assign importance to different bands, they face challenges in unsupervised hyperspectral anomaly detection (AD) due to the lack of predefined labels to guide the selection process. This paper addresses these challenges by proposing a self-supervised approach for identifying fewer interpretable bands that best separate background and anomalous data. By formulating band selection as a subset selection problem using pseudo-anomalies, bands are chosen based on high confidence, rich feature diversity, and inter-collaboration. Specifically, the pseudo-anomalies and the corresponding map are utilized to identify bands with high confidence through a self-supervised strategy. To further enhance feature diversity and collaboration among bands, we impose a feature diversity constraint on the selection of subsets and assess the collaboration ability of various subsets. A novel evaluation function is designed to discover more useful bands for AD. Experiments demonstrate the effectiveness of each module and show that the proposed method selects bands that are beneficial for AD across four different datasets. The code is released at https://github.com/rk-rkk/Fewer-Interpretable-Bands-for-Hyperspectral-Anomaly-Detection .}, archive = {J_NEUCOM}, author = {Ruike Wang and Jing Hu}, doi = {10.1016/j.neucom.2024.128964}, journal = {Neurocomputing}, month = {2}, pages = {128964}, shortjournal = {Neurocomputing}, title = {Fewer interpretable bands via self-supervision for hyperspectral anomaly detection}, volume = {616}, year = {2025}, }
• (2025). 10-minute forest early wildfire detection: Fusing multi-type and multi-source information via recursive transformer. NEUCOM, 616, 128963. (www)

  @article{ , abstract = {Forest wildfire has great impacts on both nature and human society. While disrupts the ecosystems, wildfire leads to significant economic loss and poses a threat to local communities. To detect forest wildfire, remote sensing technology has become an essential and powerful tool. Compared with polar-orbiting satellite, the new generation of geostationary satellite provides higher temporal resolution and faster response capability. In this study, we utilize the near real-time data of Himawari-8/9 satellite, to achieve 10-min forest early wildfire detection. A recursive transformer model is proposed in this work. It fuses multi-type and multi-source information for Himawari-8/9 satellite. By leveraging the spectral, temporal and spatial features of fire pixels and considering land cover information, the proposed method reduces interference factors like cloud and terrain, resulting in minute-level and near real-time detection of forest wildfire. In 21 ground truth forest wildfire scenarios and MODIS-based cross-validation dataset, the proposed method achieves better results compared to the JAXA wildfire product, in terms of overall fire detection accuracy, early fire detection rate, omission rate, and real-time performance. Furthermore, the proposed framework effectively lowers the emergency response time for early forest wildfire detection, thereby reducing the loss caused by forest wildfire.}, archive = {J_NEUCOM}, author = {Qiang Zhang and Jian Zhu and Yushuai Dong and Enyu Zhao and Meiping Song and Qiangqiang Yuan}, doi = {10.1016/j.neucom.2024.128963}, journal = {Neurocomputing}, month = {2}, pages = {128963}, shortjournal = {Neurocomputing}, title = {10-minute forest early wildfire detection: Fusing multi-type and multi-source information via recursive transformer}, volume = {616}, year = {2025}, }
• (2025). EEG-based driver drowsiness detection based on simulated driving environment. NEUCOM, 616, 128961. (www)

  @article{ , abstract = {One of the primary causes of traffic accidents is drowsy driving. Therefore, it is of paramount importance to detect driver drowsiness promptly through accurate electroencephalography (EEG) methods. Nevertheless, some existing drowsy driving detection experiments are overly idealized in design, resulting in significant discrepancies between the collected data and the actual driving environment. Additionally, the data granularity is relatively coarse. This paper proposes an EEG data acquisition and driver drowsiness detection scheme that simulates the real driving environment. First, by simulating the real driving environment, more realistic EEG data are successfully collected when the driver is performing a continuous concentration task. Second, a convolutional neural network (CNN) is employed, integrating the attention mechanism and multi-anchor attentive fusion technology. To address the issue of data imbalance resulting from subject-specific differences, a novel loss function was devised in conjunction with Focal Loss. This approach enables the model to focus its attention on a smaller number of samples, thereby enhancing the classification performance and robustness of the model. Furthermore, a contribution-based Gumbel channel selection module was introduced to reduce the computational cost while minimizing the impact of feature reduction on performance. To assess the efficacy of the proposed approach, extensive experiments were conducted on both public and self-built datasets. The experimental results demonstrate that the scheme is capable of effectively simulating real driving situations, accurately capturing drivers' EEG signals, and detecting their drowsy states. This approach may prove valuable in preventing traffic accidents caused by driver drowsiness. Data: https://github.com/jiki-mhe0509/driver_drowsy_data}, archive = {J_NEUCOM}, author = {Xuefen Lin and Zhiqi Huang and Weifeng Ma and Wei Tang}, doi = {10.1016/j.neucom.2024.128961}, journal = {Neurocomputing}, month = {2}, pages = {128961}, shortjournal = {Neurocomputing}, title = {EEG-based driver drowsiness detection based on simulated driving environment}, volume = {616}, year = {2025}, }
• (2025). Newton cooling theorem-based local overlapping regions cleaning and oversampling techniques for imbalanced datasets. NEUCOM, 616, 128959. (www)

  @article{ , abstract = {Imbalanced datasets pose significant challenges to machine learning tasks because traditional classifiers tend to favor the majority class. While numerous methods have been proposed to balance data distribution, recent studies have identified that imbalanced classification is also hindered by other data characteristics. Among these factors, the joint effects of class overlap and within-class imbalance are particularly harmful to classification performance. To the end, we propose a novel algorithm called Newton Cooling Theorem-Based Local Overlapping Regions Cleaning and Oversampling (NCLO-SMOTE). This method employs an adaptive semi-supervised clustering algorithm, which divides the minority class into several clusters without requiring a pre-set number of clusters. It quantifies both the overall and local overlapping degrees of the dataset based on the number of clusters and their local information. Additionally, it uses Newton’s Cooling Theorem to clean these overlapping regions and a cluster-weighted oversampling strategy to address within-class imbalance. Comparative experiments were conducted between NCLO-SMOTE and ten state-of-the-art sampling methods on 48 real-world imbalanced datasets. The experimental results demonstrate that our proposed method not only achieves superior performance but also exhibits strong robustness and versatility in handling the joint effects of class overlap and imbalance.}, archive = {J_NEUCOM}, author = {Liangliang Tao and Qingya Wang and Fen Yu and Hui Cao and Yage Liang and Huixia Luo and Jinghui Guo}, doi = {10.1016/j.neucom.2024.128959}, journal = {Neurocomputing}, month = {2}, pages = {128959}, shortjournal = {Neurocomputing}, title = {Newton cooling theorem-based local overlapping regions cleaning and oversampling techniques for imbalanced datasets}, volume = {616}, year = {2025}, }
• (2025). DSAFuse: Infrared and visible image fusion via dual-branch spatial adaptive feature extraction. NEUCOM, 616, 128957. (www)

  @article{ , abstract = {By exploiting the thermal radiation information from infrared images and the detailed texture information from visible light images, image fusion technology enables more accurate target identification. However, most current image fusion methods primarily rely on convolutional neural networks for cross-modal local feature extraction and do not fully utilize long-range contextual information, resulting in limited performance in complex scenarios. To address this issue, this paper proposes an infrared and visible light image fusion method termed DSAFuse, which is based on dual-branch spatially adaptive feature extraction. Specifically, a unimodal feature mixing module is used for multi-scale spatially adaptive feature extraction on both modal images with shared weights. The extracted features are then inputted into a dual-branch feature extraction module comprising flatten transformer blocks and vanilla blocks, which extract low-frequency texture features and high-frequency local detail features, respectively. Subsequently, features from both modalities are concatenated, and a bimodal feature mixing module reconstructs the fused image to generate semantically rich fusion results. Additionally, to achieve end-to-end unsupervised training, a loss function consisting of decomposition loss, gradient loss, and structural similarity loss is designed. Qualitative and quantitative experimental results demonstrate that our DSAFuse outperforms the state-of-the-art IVIF methods across various benchmark datasets. It effectively preserves the texture details and target features of the source images, producing satisfactory fusion results even in harsh environments and enhancing downstream visual tasks.}, archive = {J_NEUCOM}, author = {Shixian Shen and Yong Feng and Nianbo Liu and Ming Liu and Yingna Li}, doi = {10.1016/j.neucom.2024.128957}, journal = {Neurocomputing}, month = {2}, pages = {128957}, shortjournal = {Neurocomputing}, title = {DSAFuse: Infrared and visible image fusion via dual-branch spatial adaptive feature extraction}, volume = {616}, year = {2025}, }
• (2025). Multi-modal degradation feature learning for unified image restoration based on contrastive learning. NEUCOM, 616, 128955. (www)

  @article{ , abstract = {In this paper, we address the unified image restoration challenge by reframing it as a contrastive learning-based classification problem. Despite the significant strides made by deep learning methods in enhancing image restoration quality, their limited capacity to generalize across diverse degradation types and intensities necessitates the training of separate models for each specific degradation scenario. We proposes an all-encompassing approach that can restore images from various unknown corruption types and levels. We devise a method that learns representations of the latent sharp image’s degradation and accompanying textual features (such as dataset categories and image content descriptions), converting these into prompts which are then embedded within a reconstruction network model to enhance cross-database restoration performance. This culminates in a unified image reconstruction framework. The study involves two stages: In the first stage, we design a MultiContentNet that learns multi-modal features (MMFs) of the latent sharp image. This network encodes the visual degradation expressions and contextual text features into latent variables, thereby exerting a guided classification effect. Specifically, MultiContentNet is trained as an auxiliary controller capable of taking the degraded input image and, through contrastive learning, extracts MMFs of the latent target image. This effectively generates natural classifiers tailored for different degradation types. The second phase integrates the learned MMFs into an image restoration network via cross-attention mechanisms. This guides the restoration model to learn high-fidelity image recovery. Experiments conducted on six blind image restoration tasks demonstrate that the proposed method achieves state-of-the-art performance, highlighting the potential significance of large-scale pretrained vision-language models’ MMFs in advancing high-quality unified image reconstruction.}, archive = {J_NEUCOM}, author = {Lei Chen and Qingbo Xiong and Wei Zhang and Xiaoli Liang and Zhihua Gan and Liqiang Li and Xin He}, doi = {10.1016/j.neucom.2024.128955}, journal = {Neurocomputing}, month = {2}, pages = {128955}, shortjournal = {Neurocomputing}, title = {Multi-modal degradation feature learning for unified image restoration based on contrastive learning}, volume = {616}, year = {2025}, }
• (2025). Working condition decoupling adversarial network: A novel method for multi-target domain fault diagnosis. NEUCOM, 616, 128953. (www)

  @article{ , abstract = {In the practical application of rotating machinery, the change of working conditions can meet different manufacturing requirements. When fault diagnosis is performed on monitoring data with different working conditions, the change of data distribution will bring interference information which is highly related to working conditions and inconsistent matching problems in the process of multi-target domain transfer. In order to solve these problems, a working condition decoupling adversarial network (WCDAN) is proposed for multi-target domain fault diagnosis. Specifically, the prototype discrepancy alignment module is constructed following a weight-shared wavelet convolution feature extractor to ensure a clear prototype representation boundary. Then, the adaptive domain discriminator weight, along with the acquired multi-domain discrepancy, are utilized to decouple the working conditions. This process filters out interference information that highly associated with the source domain working conditions while preserving the inherent fault characteristics. Furthermore, the strategy of multi-domain hybrid alignment aims to minimize the disparity between different domains and solve the inconsistent matching issue. Based on two gearbox fault datasets under stable and unstable conditions, the comparative experimental results show that the WCDAN can be generalized from a single source domain to multiple target domains at the same time and achieve excellent fault diagnosis performance.}, archive = {J_NEUCOM}, author = {Xuepeng Zhang and Jinrui Wang and Xue Jiang and Zongzhen Zhang and Baokun Han and Huaiqian Bao and Xingxing Jiang}, doi = {10.1016/j.neucom.2024.128953}, journal = {Neurocomputing}, month = {2}, pages = {128953}, shortjournal = {Neurocomputing}, title = {Working condition decoupling adversarial network: A novel method for multi-target domain fault diagnosis}, volume = {616}, year = {2025}, }
• (2025). Gossip-based asynchronous algorithms for distributed composite optimization. NEUCOM, 616, 128952. (www)

  @article{ , abstract = {The distributed composite optimization problem associated a multi-agent network is investigated in this paper. Different from conventional optimization issues, the cost function of composite optimization consists of a convex function and a regularization function (possibly nonsmooth). The gossip protocol is also introduced to enhance the robustness of the network, and a gossip-based distributed composite mirror descent algorithm is presented to deal with the previous problem, which adopts the asynchronous communication method. Moreover, the algorithm performance is analyzed and the theoretical results on the corresponding error bounds are obtained. Finally, the distributed logistic regression is provided as an example to validate the practicability of the proposed algorithm.}, archive = {J_NEUCOM}, author = {Xianju Fang and Baoyong Zhang and Deming Yuan}, doi = {10.1016/j.neucom.2024.128952}, journal = {Neurocomputing}, month = {2}, pages = {128952}, shortjournal = {Neurocomputing}, title = {Gossip-based asynchronous algorithms for distributed composite optimization}, volume = {616}, year = {2025}, }
• (2025). Stable transfer learning-based control: An off-dynamics adaptive approach for unknown nonlinear systems. NEUCOM, 616, 128951. (www)

  @article{ , abstract = {Although data-driven control can achieve satisfactory control performance without relying on the system’s mathematical model (i.e., model-free), extensive online interaction with the real system remains necessary. This interaction-intensive nature has become a barrier to the widespread use of data-driven control. In fact, training a controller on the system from scratch is usually infeasible due to limited data availability and safety reasons. This paper aims to mitigate this drawback while still preserving the model-free feature of data-driven control with help of transfer learning. Specifically, a controller transfer learning problem is considered for nonlinear systems with unknown dynamics, a topic underexplored in the literature. The proposed method can transfer the controller trained on the source system directly to the target system without further tuning. It is firstly be revealed that the transfer map for this problem is a dynamic system coupled with the control law, which is learned by constructing an auxiliary system in a data-driven manner. Then, the off-dynamics adaptive law is designed to train the controller using the Lyapunov stability theory. Furthermore, our method is extended to a class of higher-order nonlinear systems with stability guarantees. Finally, the stability and reliability of the method are investigated through both rigorous theoretical analysis and illustrative examples. The results demonstrate that our method successfully transfers the tracking controller between two unknown and different nonlinear systems with stability guarantees.}, archive = {J_NEUCOM}, author = {Haoran Zhang and Chunhui Zhao}, doi = {10.1016/j.neucom.2024.128951}, journal = {Neurocomputing}, month = {2}, pages = {128951}, shortjournal = {Neurocomputing}, title = {Stable transfer learning-based control: An off-dynamics adaptive approach for unknown nonlinear systems}, volume = {616}, year = {2025}, }
• (2025). Rethinking residual connection in training large-scale spiking neural networks. NEUCOM, 616, 128950. (www)

  @article{ , abstract = {Spiking Neural Network (SNN) is known as the most famous brain-inspired model, but the non-differentiable spiking mechanism makes it hard to train large-scale SNNs. To facilitate the training of large-scale SNNs, many training methods are borrowed from Artificial Neural Networks (ANNs), among which deep residual learning is the most commonly used. But the unique features of SNNs make prior intuition built upon ANNs not available for SNNs. Although there are a few studies that have made some pioneer attempts on the topology of Spiking ResNet, the advantages of different connections remain unclear. To tackle this issue, we analyze the merits and limitations of various residual connections and empirically demonstrate our ideas with extensive experiments. Then, based on our observations, we abstract the best-performing connections into densely additive (DA) connection, extend such a concept to other topologies, and propose four architectures for training large-scale SNNs, termed DANet, which brings up to 13.24 % accuracy gain on ImageNet. Besides, in order to present a detailed methodology for designing the topology of large-scale SNNs, we further conduct in-depth discussions on their applicable scenarios in terms of their performance on various scales of datasets and demonstrate their advantages over prior architectures. At a low training expense, our best-performing ResNet-50/101/152 obtain 73.71 % /76.13 % /77.22 % top-1 accuracy on ImageNet with 4 time steps. We believe that this work shall give more insights for future works to design the topology of their networks and promote the development of large-scale SNNs. The code will be publicly available.}, archive = {J_NEUCOM}, author = {Yudong Li and Yunlin Lei and Xu Yang}, doi = {10.1016/j.neucom.2024.128950}, journal = {Neurocomputing}, month = {2}, pages = {128950}, shortjournal = {Neurocomputing}, title = {Rethinking residual connection in training large-scale spiking neural networks}, volume = {616}, year = {2025}, }
• (2025). Graph-based context learning network for infrared small target detection. NEUCOM, 616, 128949. (www)

  @article{ , abstract = {Convolutional neural networks (CNNs) have shown remarkable performance in the field of infrared small target detection. However, due to the limitation of local receptive field, existing methods find it challenging to effectively model the contextual information associated with small targets. In this paper, we propose a Graph-based Context Learning Network (GCLNet) that addresses this issue by integrating global graph reasoning with local feature learning to perceive context information across multiple scales. Specifically, local feature learning blocks are embedded into the encoder to extract detailed textures that are crucial for small targets detection. At deep layers, the multiple graph reasoning module leverages a multi-graph interaction structure to promote significant information transfer, allowing for the optimization of global context learning. Moreover, the patch-based graph reasoning module divides the low-level features into multiple patches where the context information is explored to capture the saliency of small targets. The experimental results demonstrate that the proposed method outperforms the state-of-the-art methods, achieving the intersection over union (IoU) of 80.26% and 94.84% on the NUAA-SIRST and NUDT-SIRST datasets, respectively. Our code will be available at https://github.com/studymonster0/GCLNet .}, archive = {J_NEUCOM}, author = {Yiwei Shen and Qingwu Li and Chang Xu and Chenkai Chang and Qiyun Yin}, doi = {10.1016/j.neucom.2024.128949}, journal = {Neurocomputing}, month = {2}, pages = {128949}, shortjournal = {Neurocomputing}, title = {Graph-based context learning network for infrared small target detection}, volume = {616}, year = {2025}, }
• (2025). ICFNet: Interactive-complementary fusion network for monocular 3D human pose estimation. NEUCOM, 616, 128947. (www)

  @article{ , abstract = {Most existing methods for 3D human pose estimation from monocular images focus on learning the spatial correlation of either the global or local joints of the human body but fail to adequately capture the inherent dependencies between them. To address this limitation, we propose the Interactive Complementary Fusion Network (ICFNet), an algorithm designed to fully utilize the prior knowledge of both global and local joint relationships to enhance prediction performance. Specifically, we introduce two feature capturers: the Global Knowledge Prior Capturer (GKPC) and the Local Region Subject Capturer (LRSC), which respectively capture global body knowledge and local joint information. Additionally, we propose three joint constraint mechanisms to express the potential association dependencies between global and local joints, which are further modeled using two association capturers: the Refined-Regression Association Capture Module (RR-ACM) and the Generalized-Guidance Association Capture Module (GG-ACM). Moreover, we introduce a novel feature transformation module, the Link Conversion Module (LCM), to transform and augment pose features. The algorithm adopts a complementary process to enhance the interaction and fusion of global and local feature information by gradually imposing constraints on the physical topological features of the human body, thereby improving its modeling capabilities. Extensive experiments demonstrate that our proposed ICFNet achieves state-of-the-art results on two challenging benchmark datasets: Human 3.6M and MPI-INF-3DHP. The code and model are available at: https://github.com/PENG-LAU/ICFNet .}, archive = {J_NEUCOM}, author = {Yong Wang and Peng Liu and Hongbo Kang and Doudou Wu and Duoqian Miao}, doi = {10.1016/j.neucom.2024.128947}, journal = {Neurocomputing}, month = {2}, pages = {128947}, shortjournal = {Neurocomputing}, title = {ICFNet: Interactive-complementary fusion network for monocular 3D human pose estimation}, volume = {616}, year = {2025}, }
• (2025). The effect of the head number for multi-head self-attention in remaining useful life prediction of rolling bearing and interpretability. NEUCOM, 616, 128946. (www)

  @article{ , abstract = {As one of the machine learning (ML) models, the multi-head self-attention mechanism (MSM) is competent in encoding high-level feature representations, providing computing superiorities, and systematically processing sequences bypassing the recurrent neural networks (RNN) models. However, the model performance and computational results are affected by head number, and the lack of impact interpretability has become a primary obstacle due to the complex internal working mechanisms. Therefore, the effects of the head number of the MSM on the accuracy of the result, the robustness of the model, and computation efficiency are investigated in the remaining useful life (RUL) prediction of rolling bearings. The results show that the accuracy of prediction results will be reduced caused by large or few head numbers. In addition, the more heads are selected, the more robust and higher the predictive efficiency of the model is achieved. The above effects are explained relying on the visualization of the attention weight distribution and functional networks, which are constructed and solved by the equivalent fully connected layer and graph theory analysis, respectively. The model's attention coefficient distribution during training and prediction shows that the representative information will be captured inadequately if fewer heads are selected, which causes MSM to neglect to assign large attention coefficients to degraded information. On the contrary, representational degradation information and redundant information will be acquired by models with too many heads. MSM will be disturbed by this redundant information in the attention weight distribution, resulting in incorrect allocation of attention. Both of these cases will reduce the accuracy of the prediction results. In addition, the selection rules of the head number are established based on the feature complexity that is measured by the sample entropy (SamEn). The local range for head selection is also found based on the relationship between head number and feature complexity; The effects of the head number of the MSM on the robustness of the model and computation efficiency are explained by the changes in the three parameters (average of the clustering coefficients, global efficiency, and of the average shortest path length) of the graph, which is constructed after solving the function network. The research provides a reference for rolling bearing prediction with high computational accuracy, calculation efficiency, and strong robustness using MSM.}, archive = {J_NEUCOM}, author = {Qiwu Zhao and Xiaoli Zhang and Fangzhen Wang and Panfeng Fan and Erick Mbeka}, doi = {10.1016/j.neucom.2024.128946}, journal = {Neurocomputing}, month = {2}, pages = {128946}, shortjournal = {Neurocomputing}, title = {The effect of the head number for multi-head self-attention in remaining useful life prediction of rolling bearing and interpretability}, volume = {616}, year = {2025}, }
• (2025). A semantic consistent object detection model for domain adaptation based on mixed-class distribution metrics. NEUCOM, 616, 128944. (www)

  @article{ , abstract = {Unsupervised domain adaptation is crucial for mitigating the performance degradation caused by domain bias in object detection tasks. In previous studies, the focus has been on pixel-level and instance-level shift alignment to minimize domain discrepancy. However, it is important to note that this method may inadvertently align single-class instance features with mixed-class instance features that belong to multiple categories within the same image during image-level domain adaptation. This challenge arises because each image in object detection tasks contains objects of multiple categories. To achieve the same category feature alignment between single-class and mixed-class, our method considers features with different mixed categories as a new class and proposes a mixed-classes H -divergence to reduce domain bias for object detection. To enhance both single-class and mixed-class semantic information, and to achieve semantic separation for the mixed-classes in H -divergence, we employ Semantic Prediction Models (SPM) and Semantic Bridging Components (SBC). Furthermore, we reweigh the loss of the pixel domain discriminator based on the SPM results to reduce sample imbalance. Our extensive experiments on widely used datasets illustrate how our method can robustly improve object detection in domain bias settings.}, archive = {J_NEUCOM}, author = {Lijun Gou and Jinrong Yang and Hangcheng Yu and Pan Wang and Xiaoping Li and Tuo Shi}, doi = {10.1016/j.neucom.2024.128944}, journal = {Neurocomputing}, month = {2}, pages = {128944}, shortjournal = {Neurocomputing}, title = {A semantic consistent object detection model for domain adaptation based on mixed-class distribution metrics}, volume = {616}, year = {2025}, }
• (2025). Fixed-time lyapunov criteria of stochastic impulsive time-delay systems and its application to synchronization of chua’s circuit networks. NEUCOM, 616, 128943. (www)

  @article{ , abstract = {In this paper, we investigate the fixed-time stability and synchronization issues of stochastic impulsive delay complex networks (SICNs) with delayed impulses. Firstly, a new fixed-time stability criterion, which provides a direct connection between impulsive strength, impulsive instants, delays and the stability time, is established for stochastic delay systems with delayed impulsive. In this criterion, two types of delays: system delay and impulsive delay are considered, and the impulsive intensity at each impulse instant can be different, (facilitate or disrupt stability). It is the first time to consider the fixed-time stability issues of this case. Secondly, based on this criterion, a novel feedback controller and an adaptive controller are designed to study the fixed-time synchronization of SICNs with delayed impulses; Thirdly, by utilizing Lyapunov functional theory, stochastic analysis technique, matrix inequality analysis techniques and proposed stability criterion, the fixed-time synchronization conditions of SICNs, are derived in the form of linear matrix inequalities (LMIs). Finally, a numerical example and an application example, Chua’s circuit networks, are provided to illustrate the validity of the theoretical analysis.}, archive = {J_NEUCOM}, author = {Xiaofei Xing}, doi = {10.1016/j.neucom.2024.128943}, journal = {Neurocomputing}, month = {2}, pages = {128943}, shortjournal = {Neurocomputing}, title = {Fixed-time lyapunov criteria of stochastic impulsive time-delay systems and its application to synchronization of chua’s circuit networks}, volume = {616}, year = {2025}, }
• (2025). Clustering-based detection algorithm of remote state estimation under stealthy innovation-based attacks with historical data. NEUCOM, 616, 128942. (www)

  @article{ , abstract = {This paper investigates a security issue in cyber–physical systems (CPSs) concerning the performance of a multi-sensor remote state estimation under a novel attack called “Optimal Stealthy Innovation-Based Attacks with Historical Data”. The attacker is able to launch a linear attack to modify sensor measurements. The objective of the attacker is to maximize the deterioration of estimation performance while ensuring they remain undetected by the χ 2 detector. To counteract this new type of attack, a remote state estimator equipped with a detection mechanism that utilizes a Gaussian mixture model (GMM) is employed. We derive the error covariances for the remote state estimator with and without a GMM detection mechanism in a recursive manner under Optimal Stealthy Innovation-Based Attacks with Historical Data. The experimental results demonstrate the superiority of the GMM detection mechanism. However, it is observed that the estimation performance of the GMM-based system deteriorates as the system dimension increases. In order to address this issue, we propose two dimensionality reduction methods, namely kernel principal component analysis (KPCA) and variational autoencoder (VAE), to enhance the estimation performance. Finally, the results are illustrated via the simulation examples.}, archive = {J_NEUCOM}, author = {Shan Chen and Yuqing Ni and Lingying Huang and Xiaoli Luan and Fei Liu}, doi = {10.1016/j.neucom.2024.128942}, journal = {Neurocomputing}, month = {2}, pages = {128942}, shortjournal = {Neurocomputing}, title = {Clustering-based detection algorithm of remote state estimation under stealthy innovation-based attacks with historical data}, volume = {616}, year = {2025}, }
• (2025). A short text topic modeling method based on integrating gaussian and logistic coding networks with pre-trained word embeddings. NEUCOM, 616, 128941. (www)

  @article{ , abstract = {The development of neural networks has provided a flexible learning framework for topic modeling. Currently, topic modeling based on neural networks has garnered wide attention. Despite its widespread application, the implementation of neural topic modeling still needs to be improved due to the complexity of short texts. Short texts usually contains only a few words and a small amount of feature information, lacking sufficient word co-occurrence and context sharing information. This results in challenges such as sparse features and poor interpretability in topic modeling. To alleviate this issue, an innovative model called T opic M odeling of E nhanced N eural N etwork with word E mbedding (ENNETM) was proposed. Firstly, we introduced an enhanced network into the inference network part, which integrated the Gaussian and Logistic coding networks to improve the performance and the interpretability of topic extraction. Secondly, we introduced the pre-trained word embedding into the Gaussian decoding network part of the model to enrich the contextual semantic information. Comprehensive experiments were carried out on three public datasets, 20NewGroups, AG_news and TagMyNews, and the results showed that the proposed method outperformed several state-of-the-art models in topic extraction and text classification.}, archive = {J_NEUCOM}, author = {Si Zhang and Jiali Xu and Ning Hui and Peiyun Zhai}, doi = {10.1016/j.neucom.2024.128941}, journal = {Neurocomputing}, month = {2}, pages = {128941}, shortjournal = {Neurocomputing}, title = {A short text topic modeling method based on integrating gaussian and logistic coding networks with pre-trained word embeddings}, volume = {616}, year = {2025}, }
• (2025). Dynamic synchronous graph transformer network for region-level air-quality forecasting. NEUCOM, 616, 128924. (www)

  @article{ , abstract = {Accurate forecasting of air quality aids in mitigating air pollution, enhancing the well-being of residents, and supporting the city’s sustainable growth. Recent works have utilized graph neural network for spatial dependency modeling in air-quality forecasting task. However, many existing methods rely on separate components to individually capture temporal and spatial correlations, which makes it difficult to synchronously capture the multiscale spatiotemporal correlation (MSTCs) from the spatiotemporal graph. This paper proposed a dynamic synchronous graph transformer (DSGT) based on the Encoder-Decoder structure to forecast air quality of urban regions. It captures time-varying observed station readings through dynamic graph convolution operations and can learn the influence of auxiliary features. We designed a multiscale dynamic synchronous graph constructing way to construct graphs which can effectively encode the MSTCs. There is a multiscale spatiotemporal synchronous graph convolution component in DSGT for extracting multiscale spatiotemporal representation from the constructed graphs. The synchronous graph attention mechanism and temporal attention mechanism were designed to integrated into Encoder-Decoder structure to focus the long-term influence of auxiliary features and the short-term influence of multiscale spatiotemporal representation. Via extensive experiments on two real-world datasets, it is demonstrated that the proposed model outperforms existing methods in both short- and long-term forecasting.}, archive = {J_NEUCOM}, author = {Hanzhong Xia and Xiaoxia Chen and Binjie Chen and Yue Hu}, doi = {10.1016/j.neucom.2024.128924}, journal = {Neurocomputing}, month = {2}, pages = {128924}, shortjournal = {Neurocomputing}, title = {Dynamic synchronous graph transformer network for region-level air-quality forecasting}, volume = {616}, year = {2025}, }
• (2025). Potential knowledge extraction network for class-incremental learning. NEUCOM, 616, 128923. (www)

  @article{ , abstract = {Class-Incremental Learning (CIL) aims to dynamically learn new classes without forgetting the old ones, and it is typically achieved by extracting knowledge from old data and continuously transferring it to new tasks. In the replay-based approaches, selecting appropriate exemplars is of great importance since exemplars represent the most direct form of retaining old knowledge. In this paper, we propose a novel CIL framework: Potential Knowledge Extraction Network (PKENet), which addresses the issue of neglecting the knowledge of inter-sample relation in most existing works and suggests an innovative approach for exemplar selection. Specifically, to address the challenge of knowledge transfer, we design a relation consistency loss and a hybrid cross-entropy loss , where the former works by extracting structural knowledge from the old model while the latter captures graph-wise knowledge, enabling the new model to acquire more old knowledge. To enhance the anti-forgetting effect of exemplar set, we devise a maximum-forgetting-priority method for selecting samples most susceptible to interference from the model’s update. To overcome the prediction bias problem in CIL, we introduce the Total Direct Effect inference method into our model. Experimental results on CIFAR100, ImageNet-Full and ImageNet-Subset datasets show that multiple state-of-the-art CIL methods can be directly combined with our PKENet to reap significant performance improvement. Code: https://github.com/XXDyeah/PKENet .}, archive = {J_NEUCOM}, author = {Xidong Xi and Guitao Cao and Wenming Cao and Yong Liu and Yan Li and Hong Wang and He Ren}, doi = {10.1016/j.neucom.2024.128923}, journal = {Neurocomputing}, month = {2}, pages = {128923}, shortjournal = {Neurocomputing}, title = {Potential knowledge extraction network for class-incremental learning}, volume = {616}, year = {2025}, }
• (2025). A fast optimization approach for seeking nash equilibrium based on Nikaido–Isoda function, state transition algorithm and Gauss–Seidel technique. NEUCOM, 616, 128922. (www)

  @article{ , abstract = {This paper proposes a fast optimization approach for non-cooperative games with complicated payoff functions (non-smooth, non-concave, etc.). The Nikaido–Isoda function is employed to convert knotty Nash equilibrium problems (NEPs) into large-scale optimization problems with complex objective functions. To efficiently seek Nash equilibrium, the resulting optimization problems are decomposed into many subproblems where each player tries to maximize its payoff when observing others’ current strategies. All players’ strategies are updated iteratively until reaching Nash equilibrium. Specifically, a dynamic state transition algorithm (STA) is proposed to seek global optima of subproblems at each iteration, and the sequential quadratic programming (SQP) is embedded into dynamic STA for convergence acceleration. A Gauss–Seidel technique is utilized for players’ strategy updates to improve computational efficiency further. Numerical examples drawn from multidisciplinary contexts validate that the proposed approach could effectively seek out Nash equilibrium for simultaneously decreasing the time-consuming remarkably.}, archive = {J_NEUCOM}, author = {Xiaojun Zhou and Zheng Wang and Tingwen Huang}, doi = {10.1016/j.neucom.2024.128922}, journal = {Neurocomputing}, month = {2}, pages = {128922}, shortjournal = {Neurocomputing}, title = {A fast optimization approach for seeking nash equilibrium based on Nikaido–Isoda function, state transition algorithm and Gauss–Seidel technique}, volume = {616}, year = {2025}, }
• (2025). A survey of text classification based on pre-trained language model. NEUCOM, 616, 128921. (www)

  @article{ , abstract = {The utilization of text classification is widespread within the domain of Natural Language Processing (NLP). In recent years, pre-trained language models (PLMs) based on the Transformer architecture have made significant strides across various artificial intelligence tasks. Currently, text classification employing PLMs has emerged as a prominent research focus within NLP. While several review papers examine text classification and Transformer models, there is a notable lack of comprehensive surveys specifically addressing text classification grounded in PLMs. To address this gap, the present survey provides an extensive overview of text classification techniques that leverage PLMs. The primary components of this review include: (1) an introduction, (2) a systematic examination of PLMs, (3) deep learning-based text classification methodologies, (4) text classification approaches utilizing pre-trained models, (5) commonly used datasets and evaluation metrics in text classification, (6) prevalent challenges and emerging trends in the field, and (7) a conclusion.}, archive = {J_NEUCOM}, author = {Yujia Wu and Jun Wan}, doi = {10.1016/j.neucom.2024.128921}, journal = {Neurocomputing}, month = {2}, pages = {128921}, shortjournal = {Neurocomputing}, title = {A survey of text classification based on pre-trained language model}, volume = {616}, year = {2025}, }
• (2025). Precise occlusion-aware and feature-level reconstruction for occluded person re-identification. NEUCOM, 616, 128919. (www)

  @article{ , abstract = {Occluded person re-IDentification (re-ID) is a challenging task in surveillance scenarios that remains unresolved. To address it, existing methods primarily rely on auxiliary models, e.g. pose estimation, to explore visible parts by detecting human keypoints. However, these approaches inevitably encounter two issues: domain gap and information asymmetry. The former arises from pre-training auxiliary models on different domains, while the latter indicates that the occluded query has asymmetric valid cues compared to the holistic visible gallery. In this paper, we propose a novel Precise Occlusion-aware and Feature-level Reconstruction (POFR) network for occluded re-ID. POFR addresses the occlusion issue from two viewpoints: perceiving the occlusions other than visible human bodies and reconstructing the occluded parts at the feature level. The first perspective is achieved through occlusion-driven contrastive learning (OCL). OCL incorporates an occlusion generator capable of generating object and person-specific occlusions. Unlike previous coarse occlusions, our generator leverages segmented pedestrians and obstacles to generate realistic occlusions which are then used for contrastive learning. The second perspective is implemented through an occlusion-guided feature reconstruction (OFR) module. OFR initially learns an occlusion predictor to estimate the occlusion mask, which is subsequently utilized to recover features corresponding to the occluded regions. Benefiting from the occlusion generator, the occlusion predictor can be effectively supervised with the precise occlusion masks, thereby mitigating the domain gap problem. Additionally, the recovered features alleviate information asymmetry when matching an occluded query and a holistic gallery. Extensive experiments conducted on occluded, partial, and holistic datasets demonstrate the superior performance of our POFR over state-of-the-art methods. The source code will be made publicly available upon paper acceptance.}, archive = {J_NEUCOM}, author = {Xiujun Shu and Hanjun Li and Wei Wen and Ruizhi Qiao and Nannan Li and Weijian Ruan and Hanjing Su and Bo Wang and Shouzhi Chen and Jun Zhou}, doi = {10.1016/j.neucom.2024.128919}, journal = {Neurocomputing}, month = {2}, pages = {128919}, shortjournal = {Neurocomputing}, title = {Precise occlusion-aware and feature-level reconstruction for occluded person re-identification}, volume = {616}, year = {2025}, }
• (2025). HO-FMN: Hyperparameter optimization for fast minimum-norm attacks. NEUCOM, 616, 128918. (www)

  @article{ , abstract = {Gradient-based attacks are a primary tool to evaluate robustness of machine-learning models. However, many attacks tend to provide overly-optimistic evaluations as they use fixed loss functions, optimizers, step-size schedulers, and default hyperparameters. In this work, we tackle these limitations by proposing a parametric variation of the well-known fast minimum-norm attack algorithm, whose loss, optimizer, step-size scheduler, and hyperparameters can be dynamically adjusted. We re-evaluate 12 robust models, showing that our attack finds smaller adversarial perturbations without requiring any additional tuning. This also enables reporting adversarial robustness as a function of the perturbation budget, providing a more complete evaluation than that offered by fixed-budget attacks, while remaining efficient. We release our open-source code at https://github.com/pralab/HO-FMN .}, archive = {J_NEUCOM}, author = {Raffaele Mura and Giuseppe Floris and Luca Scionis and Giorgio Piras and Maura Pintor and Ambra Demontis and Giorgio Giacinto and Battista Biggio and Fabio Roli}, doi = {10.1016/j.neucom.2024.128918}, journal = {Neurocomputing}, month = {2}, pages = {128918}, shortjournal = {Neurocomputing}, title = {HO-FMN: Hyperparameter optimization for fast minimum-norm attacks}, volume = {616}, year = {2025}, }
• (2025). VIWHard: Text adversarial attacks based on important-word discriminator in the hard-label black-box setting. NEUCOM, 616, 128917. (www)

  @article{ , abstract = {In the hard-label black-box setting, the adversary only obtains the decision of the target model, which is more practical. Both the perturbed words and the sets of substitute words affect the performance of adversarial attack. We propose a hard-label black-box adversarial attack framework called VIWHard, which takes important words as perturbed words. In order to verify the words which highly impact on the classification of the target model, we design an important-word discriminator consisting of a binary classifier and a masked language model as an important component of VIWHard. Meanwhile, we use a masked language model to construct the context-preserving sets of substitute words for important words, which further improves the naturalness of the adversarial texts. We conduct experiments by attacking WordCNN, WordLSTM and BERT on seven datasets, which contain text classification, toxic information, and sensitive information datasets. Experimental results show that our method achieves powerful attacking performance and generates natural adversarial texts. The average attack success rate on the seven datasets reaches 98.556%, and the average naturalness of the adversarial texts reaches 7.894. Specially, on the four security datasets Jigsaw2018, HSOL, EDENCE, and FAS, our average attack success rate reaches 97.663%, and the average naturalness of the adversarial texts reaches 8.626. In addition, we evaluate the attack performance of VIWHard on large language models (LLMs), the generated adversarial examples are effective for LLMs.}, archive = {J_NEUCOM}, author = {Hua Zhang and Jiahui Wang and Haoran Gao and Xin Zhang and Huewei Wang and Wenmin Li}, doi = {10.1016/j.neucom.2024.128917}, journal = {Neurocomputing}, month = {2}, pages = {128917}, shortjournal = {Neurocomputing}, title = {VIWHard: Text adversarial attacks based on important-word discriminator in the hard-label black-box setting}, volume = {616}, year = {2025}, }
• (2025). Fully end-to-end EEG to speech translation using multi-scale optimized dual generative adversarial network with cycle-consistency loss. NEUCOM, 616, 128916. (www)

  @article{ , abstract = {Decoding auditory evoked electroencephalographic (EEG) signals to correlate them with speech acoustic features and construct transitional signals between different domain signals is a challenging and fascinating research topic. Brain–computer interface (BCI) technologies that incorporate auditory evoked potentials (AEPs) can not only leverage encoder–decoder architectures for signal decoding, but also employ generative adversarial networks (GANs) to translate from human neural activity to speech (T-HNAS). However, in previous research, the cascading ratio of transitional signals leads to varying degrees of information loss in the two-domain signals, and the optimal ratio of transitional signals differs across datasets, impacting the translation effectiveness. To address these issues, an improved dual generative adversarial network based on multi-scale optimization and cycle-consistency loss (MSCC-DualGAN) is proposed. We leverage the feature of cycle consistency loss, which facilitates cross-modal signal conversion, to replace transitional signals and maintain the integrity of signals in both domains during the loss computation process. Multi-scale optimization is utilized to refine the details of signals downsampled by the network, improving the similarity between features, thus enabling efficient, fully end-to-end EEG to speech translation. Furthermore, to validate the efficacy of this network, we construct a new EEG dataset and conduct studies using metrics such as mel cepstral distortion (MCD), pearson correlation coefficient (PCC), and structural similarity index measure (SSIM). Experimental results demonstrate that this new network significantly outperforms previous methods on auditory stimulus datasets.}, archive = {J_NEUCOM}, author = {Chen Ma and Yue Zhang and Yina Guo and Xin Liu and Hong Shangguan and Juan Wang and Luqing Zhao}, doi = {10.1016/j.neucom.2024.128916}, journal = {Neurocomputing}, month = {2}, pages = {128916}, shortjournal = {Neurocomputing}, title = {Fully end-to-end EEG to speech translation using multi-scale optimized dual generative adversarial network with cycle-consistency loss}, volume = {616}, year = {2025}, }
• (2025). Fooling human detectors via robust and visually natural adversarial patches. NEUCOM, 616, 128915. (www)

  @article{ , abstract = {DNNs are vulnerable to adversarial attacks. Physical attacks alter local regions of images by either physically equipping crafted objects or synthesizing adversarial patches. This design is applicable to real-world image capturing scenarios. Currently, adversarial patches are typically generated from random noise. Their textures are different from image textures. Also, these patches are developed without focusing on the relationship between human poses and adversarial robustness. The unnatural pose and texture make patches noticeable in practice. In this work, we propose to synthesize adversarial patches which are visually natural from the perspectives of both poses and textures. In order to adapt adversarial patches to human pose, we propose a patch adaption network PosePatch for patch synthesis, which is guided by perspective transform with estimated human poses. Meanwhile, we develop a network StylePatch to generate harmonized textures for adversarial patches. These networks are combined together for end-to-end training. As a result, our method can synthesize adversarial patches for arbitrary human images without knowing poses and localization in advance. Experiments on benchmark datasets and real-world scenarios show that our method is robust to human pose variations and synthesized adversarial patches are effective, and a user study is made to validate the naturalness.}, archive = {J_NEUCOM}, author = {Dawei Zhou and Hongbin Qu and Nannan Wang and Chunlei Peng and Zhuoqi Ma and Xi Yang and Xinbo Gao}, doi = {10.1016/j.neucom.2024.128915}, journal = {Neurocomputing}, month = {2}, pages = {128915}, shortjournal = {Neurocomputing}, title = {Fooling human detectors via robust and visually natural adversarial patches}, volume = {616}, year = {2025}, }
• (2025). Hyperspectral image denoising via cooperated self-supervised CNN transform and nonconvex regularization. NEUCOM, 616, 128912. (www)

  @article{ , abstract = {Methods that leverage the sparsity and the low-rankness in the transformed domain have gained growing interest for hyperspectral image (HSI) denoising. Recently, many researches simultaneously utilizing low-rankness and local smoothness have emerged. Although these approaches achieve great denoising performance, they exhibit several limitations. First, the widely adopted l 1 norm is a biased function, potentially leading to blurring edges. Second, employing tensor singular value decomposition (T-SVD) to ensure low-rankness brings a heavy computational burden. Additionally, the manually designed regularization norm is fixed for all testing data, which may cause a generalization problem. To address these challenges, this work proposes a novel optimization model for HSI denoising that incorporates the self-supervised CNN transform and TV regularization (CTTV) with the nonconvex function induced norm. The CNN-based transform could implicitly ensure the low-rankness of the tensor and learn the potential information in the noisy data. Furthermore, we exploit the unbiased nonconvex minimax concave penalty (MCP) to enforce the local smoothness of the extracted features while preserving sharp edges. We design an algorithm to solve the proposed model built on the hybrid of the half-quadratic splitting (HQS) and the alternating direction method of multipliers (ADMM), in which the network parameter and the denoised image are separately optimized. Extensive experiments on various datasets indicate that our proposed method can achieve state-of-the-art performance in HSI denoising.}, archive = {J_NEUCOM}, author = {Ruizhi Hou and Fang Li}, doi = {10.1016/j.neucom.2024.128912}, journal = {Neurocomputing}, month = {2}, pages = {128912}, shortjournal = {Neurocomputing}, title = {Hyperspectral image denoising via cooperated self-supervised CNN transform and nonconvex regularization}, volume = {616}, year = {2025}, }
• (2025). An optimal transport-guided diffusion framework with mitigating mode mixture. NEUCOM, 616, 128910. (www)

  @article{ , abstract = {Diffusion probability models (DPMs) have achieved excellent results in image generation; however, their inference process is slow and tends to produce more mixed images. The autoencoder optimal transport (OT) model addresses the mode collapse/mixture problem from the OT perspective but produces low-quality images. Therefore, to generate high-quality images and mitigate mode mixture, we propose an innovative OT-guided diffusion framework. The key is to find the optimal truncation step M to ensure that the class boundaries of the original data do not intersect during the forward process, ensuring that the generated image belongs to the same class as the initial point in the reverse process. The value of M is determined by evaluating the Peak Signal-to-Noise Ratio, enabling us to mitigate the generation of mixed images. Specifically, our approach first involves embedding the images’ manifold into the latent space through an encoder. The images are subsequently decoded using latent codes, which are generated through an OT map from the Gaussian distribution to the empirical latent distribution. Finally, the trained M -step DPM is utilized to refine the image generated by the decoder. Experimental results demonstrate that our method not only improves image quality but also alleviates mode mixture in diffusion models. Additionally, it enhances sampling efficiency and reduces training cost compared to classical diffusion models.}, archive = {J_NEUCOM}, author = {Shenghao Li and Zhanpeng Wang and Zhongxuan Luo and Na Lei}, doi = {10.1016/j.neucom.2024.128910}, journal = {Neurocomputing}, month = {2}, pages = {128910}, shortjournal = {Neurocomputing}, title = {An optimal transport-guided diffusion framework with mitigating mode mixture}, volume = {616}, year = {2025}, }
• (2025). RGBT tracking via frequency-aware feature enhancement and unidirectional mixed attention. NEUCOM, 616, 128908. (www)

  @article{ , abstract = {RGBT object tracking is widely used due to the complementary nature of RGB and TIR modalities. However, RGBT trackers based on Transformer or CNN face significant challenges in effectively enhancing and extracting features from one modality and fusing them into another modality. To achieve effective regional feature representation and adequate information fusion, we propose a novel tracking method that employs frequency-aware feature enhancement and bidirectional multistage feature fusion. Firstly, we propose an Early Region Feature Enhancement (ERFE) module, which is comprised of the Frequency-aware Self-region Feature Enhancement (FSFE) block and the Cross-attention Cross-region Feature Enhancement (CCFE) block. The FFT-based FSFE block can enhance the feature of the template or search region separately, while the CCFE block can improve feature representation by considering the template and search region jointly. Secondly, we propose a Bidirectional Multistage Feature Fusion (BMFF) module, with the Complementary Feature Extraction Attention (CFEA) module as its core component. The CFEA module including the Unidirectional Mixed Attention (UMA) block and the Context Focused Attention (CFA) block, can extract information from one modality. When RGB is the primary modality, TIR is the auxiliary modality, and vice versa. The auxiliary modal features processed by CFEA are added to the primary modal features. This information fusion process is bidirectional and multistage. Thirdly, extensive experiments on three benchmark datasets — RGBT234, LaSHeR, and GTOT — demonstrate that our tracker outperforms the advanced RGBT tracking methods.}, archive = {J_NEUCOM}, author = {Jianming Zhang and Jing Yang and Zikang Liu and Jin Wang}, doi = {10.1016/j.neucom.2024.128908}, journal = {Neurocomputing}, month = {2}, pages = {128908}, shortjournal = {Neurocomputing}, title = {RGBT tracking via frequency-aware feature enhancement and unidirectional mixed attention}, volume = {616}, year = {2025}, }
• (2025). CRMSP: A semi-supervised approach for key information extraction with class-rebalancing and merged semantic pseudo-labeling. NEUCOM, 616, 128907. (www)

  @article{ , abstract = {There is a growing demand in the field of Key Information Extraction (KIE) to apply semi-supervised learning (SSL) to save manpower and costs, as training document data using fully-supervised methods requires labor-intensive manual annotation. The main challenges of applying SSL in the KIE are (1) underestimation of the confidence of tail classes in the long-tailed distribution and (2) difficulty in achieving intra-class compactness and inter-class separability of tail features. To address these challenges, we propose a novel semi-supervised approach for KIE with Class-Rebalancing and Merged Semantic Pseudo-Labeling (CRMSP). Firstly, the Class-Rebalancing Pseudo-Labeling (CRP) module introduces a reweighting factor to rebalance pseudo-labels, increasing attention to tail classes. Secondly, we propose the Merged Semantic Pseudo-Labeling (MSP) module to cluster tail features of unlabeled data by assigning samples to Merged Prototypes (MP). Additionally, we designed a new contrastive loss specifically for MSP. Extensive experimental results on three well-known benchmarks demonstrate that CRMSP achieves state-of-the-art performance. Remarkably, CRMSP achieves 3.24% f1-score improvement over state-of-the-art on the CORD.}, archive = {J_NEUCOM}, author = {Qi Zhang and Yonghong Song and Pengcheng Guo and Yangyang Hui}, doi = {10.1016/j.neucom.2024.128907}, journal = {Neurocomputing}, month = {2}, pages = {128907}, shortjournal = {Neurocomputing}, title = {CRMSP: A semi-supervised approach for key information extraction with class-rebalancing and merged semantic pseudo-labeling}, volume = {616}, year = {2025}, }
• (2025). DETrack: Depth information is predictable for tracking. NEUCOM, 616, 128906. (www)

  @article{ , abstract = {The purpose of multi-object tracking lies in the estimation of both the bounding boxes of targets and their identities. Nonetheless, occlusion brought by the object interactions often cause identity switches and trajectory loss. Inspired by the human vision of three-dimensional tracking properties, we propose a tracking framework based on depth estimation called DETrack to address this issue. This framework features a Depth Information Module (DIM) under monocular conditions, which can produce depth features as an association cue for multi-object tracking. In addition, to actively retrieves information lost in trajectories, we have also put forward a ”refind” component, which echoes how human vision compensates for objects out of sight. Our framework can seamlessly integrate with most trackers, and introduce introducing an entirely new data dimension to the tracking task. We have tested DETrack using the MOT17 and DanceTrack benchmark datasets and compared it with alternative methods. The test results demonstrate that our technique works effectively with current MOT trackers, and it significantly enhances tracking results based on HOTA, IDF1, and MOTA metrics on both datasets.}, archive = {J_NEUCOM}, author = {Weiyu Zhao and Yizhuo Jiang and Yan Gao and Jie Li and Xinbo Gao}, doi = {10.1016/j.neucom.2024.128906}, journal = {Neurocomputing}, month = {2}, pages = {128906}, shortjournal = {Neurocomputing}, title = {DETrack: Depth information is predictable for tracking}, volume = {616}, year = {2025}, }
• (2025). Diversified recommendation with weighted hypergraph embedding: Case study in music. NEUCOM, 616, 128905. (www)

  @article{ , abstract = {Recommender systems serve a dual purpose for users: sifting out inappropriate or mismatched information while accurately identifying items that align with their preferences. Numerous recommendation algorithms rely on rich feature data to deliver personalized suggestions. However, in scenarios without explicit features, balancing accuracy and diversity in recommendations is a pressing concern. To address this challenge, exemplified by music recommendation, we introduce the Diversified Weighted Hypergraph Recommendation algorithm (DWHRec). In DWHRec, the initial connections between users and items are modeled using a weighted hypergraph, where additional entities linked to users and items, such as artists, albums, and tags, are simultaneously integrated into the hypergraph structure. To capture users’ latent preferences, a random-walk embedding method is applied to the hypergraph. Accuracy is measured by the match between users and items, and diversity is gauged by the variety of recommended item types. Extensive experiments conducted on two real-world music datasets show that DWHRec substantially outperforms eight state-of-the-art algorithms in terms of accuracy and diversity. Beyond music recommendation, DWHRec is a versatile framework that can be applied to other domains with similar data structures. The algorithm code is available on GitHub. 1}, archive = {J_NEUCOM}, author = {Chaoguang Luo and Liuying Wen and Yong Qin and Philip S. Yu and Liangwei Yang and Zhineng Hu}, doi = {10.1016/j.neucom.2024.128905}, journal = {Neurocomputing}, month = {2}, pages = {128905}, shortjournal = {Neurocomputing}, title = {Diversified recommendation with weighted hypergraph embedding: Case study in music}, volume = {616}, year = {2025}, }
• (2025). Self-training: A survey. NEUCOM, 616, 128904. (www)

  @article{ , abstract = {Self-training methods have gained significant attention in recent years due to their effectiveness in leveraging small labeled datasets and large unlabeled observations for prediction tasks. These models identify decision boundaries in low-density regions without additional assumptions about data distribution, using the confidence scores of a learned classifier. The core principle of self-training involves iteratively assigning pseudo-labels to unlabeled samples with confidence scores above a certain threshold, enriching the labeled dataset and retraining the classifier. This paper presents self-training methods for binary and multi-class classification, along with variants and related approaches such as consistency-based methods and transductive learning. We also briefly describe self-supervised learning and reinforced self-training. Furthermore, we highlight popular applications of self-training and discuss the importance of dynamic thresholding and reducing pseudo-label noise for performance improvement. To the best of our knowledge, this is the first thorough and complete survey on self-training.}, archive = {J_NEUCOM}, author = {Massih-Reza Amini and Vasilii Feofanov and Loïc Pauletto and Liès Hadjadj and Émilie Devijver and Yury Maximov}, doi = {10.1016/j.neucom.2024.128904}, journal = {Neurocomputing}, month = {2}, pages = {128904}, shortjournal = {Neurocomputing}, title = {Self-training: A survey}, volume = {616}, year = {2025}, }
• (2025). Spatialspectral-backdoor: Realizing backdoor attack for deep neural networks in brain–computer interface via EEG characteristics. NEUCOM, 616, 128902. (www)

  @article{ , abstract = {In recent years, electroencephalogram (EEG) based on the brain–computer interface (BCI) systems have become increasingly advanced, with researcher using deep neural networks as tools to enhance performance. BCI systems heavily rely on EEG signals for effective human–computer interactions, and deep neural networks show excellent performance in processing and classifying these signals. Nevertheless, the vulnerability to backdoor attack is still a major problem. Backdoor attack is the injection of specially designed triggers into the model training process, which can lead to significant security issues. Therefore, in order to simulate the negative impact of backdoor attack and bridge the research gap in the field of BCI, this paper proposes a new backdoor attack method to call researcher attention to the security issues of BCI. In this paper, Spatialspectral-Backdoor is proposed to effectively attack the BCI system. The method is carefully designed to target the spectral active backdoor attack of the BCI system and includes a multi-channel preference method to select the electrode channels sensitive to the target task. Ultimately, the effectiveness of the comparison and ablation experiments is validated on the publicly available BCI competition datasets. The results show that the average attack success rate and clean sample accuracy of Spatialspectral-Backdoor in the BCI scenario are 97.12% and 85.16%, respectively, compared with other backdoor attack methods. Furthermore, by observing the infection ratio of backdoor triggers and visualization of the feature space, the proposed Spatialspectral-Backdoor outperforms other backdoor attack methods.}, archive = {J_NEUCOM}, author = {Fumin Li and Mengjie Huang and Wenlong You and Longsheng Zhu and Hanjing Cheng and Rui Yang}, doi = {10.1016/j.neucom.2024.128902}, journal = {Neurocomputing}, month = {2}, pages = {128902}, shortjournal = {Neurocomputing}, title = {Spatialspectral-backdoor: Realizing backdoor attack for deep neural networks in brain–computer interface via EEG characteristics}, volume = {616}, year = {2025}, }
• (2025). Shared hybrid attention transformer network for colon polyp segmentation. NEUCOM, 616, 128901. (www)

  @article{ , abstract = {In the field of medical imaging, the automatic detection and segmentation of colon polyps is crucial for the early diagnosis of colorectal cancer. Currently, Transformer methods are commonly employed for colon polyp segmentation tasks, often utilizing dual attention mechanisms. However, these attention mechanisms typically utilize channel attention and spatial attention in a serial or parallel manner, which increases computational costs and model complexity. To address these issues, we propose a Shared Hybrid Attention Transformer (SHAT) framework, which shares queries and keys, thereby avoiding redundant computations and reducing computational complexity. Additionally, we introduce differential subtraction attention module to enhance feature fusion capability and significantly improve the delineation of polyp boundaries, effectively capture complex image details and edge information involved in the colon polyp images comparing with existing techniques. Our approach overcomes the limitations of existing colon polyp segmentation techniques. Experimental results on a large-scale annotated colon polyp image dataset demonstrate that our method excels in localizing and segmenting polyps of various sizes, shapes, and textures with high robustness. The source code for the SHAT framework is available at https://github.com/peanutHao/SHAT .}, archive = {J_NEUCOM}, author = {Zexuan Ji and Hao Qian and Xiao Ma}, doi = {10.1016/j.neucom.2024.128901}, journal = {Neurocomputing}, month = {2}, pages = {128901}, shortjournal = {Neurocomputing}, title = {Shared hybrid attention transformer network for colon polyp segmentation}, volume = {616}, year = {2025}, }
• (2025). Prompt-guided bidirectional deep fusion network for referring image segmentation. NEUCOM, 616, 128899. (www)

  @article{ , abstract = {Referring image segmentation involves accurately segmenting objects based on natural language descriptions. This poses challenges due to the intricate and varied nature of language expressions, as well as the requirement to identify relevant image regions among multiple objects. Current models predominantly employ language-aware early fusion techniques, which may lead to misinterpretations of language expressions due to the lack of explicit visual guidance of the language encoder. Additionally, early fusion methods are unable to adequately leverage high-level contexts. To address these limitations, this paper introduces the Prompt-guided Bidirectional Deep Fusion Network (PBDF-Net) to enhance the fusion of language and vision modalities. In contrast to traditional unidirectional early fusion approaches, our approach employs a prompt-guided bidirectional encoder fusion (PBEF) module to promote mutual cross-modal fusion across multiple stages of the vision and language encoders. Furthermore, PBDF-Net incorporates a prompt-guided cross-modal interaction (PCI) module during the late fusion stage, facilitating a more profound integration of contextual information from both modalities, resulting in more accurate target segmentation. Comprehensive experiments conducted on the RefCOCO, RefCOCO+, G-Ref and ReferIt datasets substantiate the efficacy of our proposed method, demonstrating significant advancements in performance compared to existing approaches.}, archive = {J_NEUCOM}, author = {Junxian Wu and Yujia Zhang and Michael Kampffmeyer and Xiaoguang Zhao}, doi = {10.1016/j.neucom.2024.128899}, journal = {Neurocomputing}, month = {2}, pages = {128899}, shortjournal = {Neurocomputing}, title = {Prompt-guided bidirectional deep fusion network for referring image segmentation}, volume = {616}, year = {2025}, }
• (2025). RockNet: Deep progressive lithology recognition model based on feature saliency and fusion. NEUCOM, 616, 128898. (www)

  @article{ , abstract = {Accurate lithology recognition is pivotal for comprehending subsurface structures and forecasting resource reservoirs in geological exploration. Most existing approaches rarely utilize multi-view heterogeneous rock microscopic images, limiting recognition performance in modern geological practices. To tackle the challenges above, we propose a deep progressive lithology recognition model named RockNet for rock microscopic images, based on local feature saliency and feature fusion. RockNet includes Multi-channel Feature Fusion (MFF) blocks and Local Feature Salience (LFS) blocks. The MFF block captures and fuses hierarchical features from each view of rock images, while the LFS block extracts subtle information across different views. In addition, we further design a novel loss function and a multi-scale prediction fusion strategy to optimize the training and inference process. Finally, RockNet adopts a deep progressive learning strategy to enhance its ability to recognize complex lithological patterns. Experimental results show that RockNet outperforms 12 comparative methods regarding accuracy, precision, recall, F1 score, and specificity. Our work will assist oil and gas exploration and groundwater resources assessment, contributing significantly to resource development and sustainable environmental stewardship.}, archive = {J_NEUCOM}, author = {Xiangyuan Zhu and Mincan Li and Zhiming Lan and Jianguo Chen and Zerui Li and Keqin Li}, doi = {10.1016/j.neucom.2024.128898}, journal = {Neurocomputing}, month = {2}, pages = {128898}, shortjournal = {Neurocomputing}, title = {RockNet: Deep progressive lithology recognition model based on feature saliency and fusion}, volume = {616}, year = {2025}, }
• (2025). ALDII: Adaptive learning-based document image inpainting to enhance the handwritten chinese character legibility of human and machine. NEUCOM, 616, 128897. (www)

  @article{ , abstract = {Document Image Inpainting (DII) has been applied to degraded documents, including financial and historical documents, to enhance the legibility of images for: (1) human readers by providing high visual quality images; and (2) machine recognizers such as Optical Character Recognition (OCR), thereby reducing recognition errors. With the advent of Deep Learning (DL), DL-based DII methods have achieved remarkable enhancements in terms of either human or machine legibility. However, focusing on improving machine legibility causes visual image degradation, affecting human readability. To address this contradiction, we propose an adaptive learning-based DII method, namely ALDII, that applies domain adaptation strategy, our approach acts like a plug-in module that is capable of constraining a total feature space before optimizing legibility of human and machine, respectively. We evaluate our ALDII on a Chinese handwritten character dataset, which includes single-character and text-line images. Compared to other state-of-the-art approaches, experimental results demonstrated superior performance of our ALDII with metrics of both human and machine legibility.}, archive = {J_NEUCOM}, author = {Qinglin Mao and Jingjin Li and Hang Zhou and Pushpendu Kar and Anthony Graham Bellotti}, doi = {10.1016/j.neucom.2024.128897}, journal = {Neurocomputing}, month = {2}, pages = {128897}, shortjournal = {Neurocomputing}, title = {ALDII: Adaptive learning-based document image inpainting to enhance the handwritten chinese character legibility of human and machine}, volume = {616}, year = {2025}, }
• (2025). Human key point detection method based on enhanced receptive field and transformer. NEUCOM, 616, 128894. (www)

  @article{ , abstract = {The existing key point detection network models have complex structures, so it is difficult to deploy on edge devices. Meanwhile, the convolution is localized and limited by the size of convolution kernels, which cannot effectively capture long-range dependencies. To address this problem, this paper introduces a lightweight Convolutional Transformer network (LHFormer Net) for human pose estimation. Considering that the sampling area of convolution kernels for different dimensional feature maps is fixed and the contextual information is singular, the enhanced receptive field block is designed to extract richer feature information and reduce information loss in feature maps. Based on the global modeling features of Transformer encoder, convolutional position encoding and multi-head self-attention are used to capture the spatial constraint relationship between key points in the deep feature extraction. Finally, a lightweight deconvolution module is used to generate higher resolution features to achieve multi-resolution supervision, which can effectively solve the problem of scale variation in pose estimation, more accurately locate key points of small and medium-sized people, ad further improve the detection accuracy of the network. Compared with other networks, the experimental results on the open access datasets COCO2017 and MPII show that the proposed network achieves a good balance between model complexity and detection performance.}, archive = {J_NEUCOM}, author = {Hongyu Liang and Jianfeng Yang and Wenjuan Xie and Jinsheng Xiao}, doi = {10.1016/j.neucom.2024.128894}, journal = {Neurocomputing}, month = {2}, pages = {128894}, shortjournal = {Neurocomputing}, title = {Human key point detection method based on enhanced receptive field and transformer}, volume = {616}, year = {2025}, }
• (2025). High-order rotor hopfield neural networks for associative memory. NEUCOM, 616, 128893. (www)

  @article{ , abstract = {Multistate associative memory models have shown a remarkable ability to remember non-binary data in recent years, including the complex-valued Hopfield neural networks (CHNNs) and their advanced counterpart of rotor Hopfield neural networks (RHNNs). However, the noise robustness of these models deteriorates significantly as the number of stored patterns and the resolution increase. To address this issue, inspired by the complex connections observed in biological neural systems, high-order connections are incorporated into CHNNs and RHNNs, resulting in the high-order complex-valued Hopfield neural networks (HCHNNs) and the high-order rotor Hopfield neural networks (HRHNNs). By abstracting virtual neurons, high-order connection-based update equations and projection rules are simultaneously modified as complex versions. The maximum storage capacity of the network is increased from N to nearly ( N + M ) , where N and M represent the number of neurons and the number of high-order connections. The associative memory capabilities of HRHNNs were validated on the CIFAR-10, MNIST, and CelebA datasets, demonstrating superior robustness to noise compared to RHNNs as the number of memory patterns increased.}, archive = {J_NEUCOM}, author = {Bingxuan Chen and Hao Zhang}, doi = {10.1016/j.neucom.2024.128893}, journal = {Neurocomputing}, month = {2}, pages = {128893}, shortjournal = {Neurocomputing}, title = {High-order rotor hopfield neural networks for associative memory}, volume = {616}, year = {2025}, }
• (2025). TSSD-based quasi-synchronization of stochastic delayed reaction–diffusion neural networks under deceptional attacks. NEUCOM, 616, 128890. (www)

  @article{ , abstract = {The quasi-synchronization (QS) issue of stochastic delayed reaction–diffusion neural networks (SDRDNNs) under deceptive attacks is investigated in this paper. A control strategy in terms of time-space sampled-data (TSSD) is presented for the QS issue of SDRDNNs under deceptive attacks, which can not only enhance the cybersecurity of communications but also reduce network bandwidth consumption. By looped Lyapunov–Krasovskii functional (LKF), free-weighting matrix (FWM) technique, second-order B-L integral inequality, Itô’s formula, generalized Itô’s isometry, Dynkin’s formula and the extended Halanay’s inequality, a newly less conservative QS criterion is established for SDRDNNs under norm-bounded deceptive attacks. Furthermore, to increase the feasibility of the QS criterion, by the total expectation formula and independent increment of Brownian motion, an innovative and effective method for estimating the mathematical expectation of the d w ( t ) d t -dependent cross-term in the corresponding FWM-based zero equation is proposed. Additionally, this study also develops the TSSD exponential synchronization (ES) criteria for SDRDNNs without deceptive attacks and for SDRDNNs under deceptive attacks with Bernoulli distribution. Finally, a simulated example consisting of three cases is provided to validate the feasibility and effectiveness of the theoretical results.}, archive = {J_NEUCOM}, author = {Wenpin Luo and Jun Yang and Yaqian Wang and Xingwen Liu and Kaibo Shi}, doi = {10.1016/j.neucom.2024.128890}, journal = {Neurocomputing}, month = {2}, pages = {128890}, shortjournal = {Neurocomputing}, title = {TSSD-based quasi-synchronization of stochastic delayed reaction–diffusion neural networks under deceptional attacks}, volume = {616}, year = {2025}, }
• (2025). Edge feature empowered graph attention network for sum rate maximization in heterogeneous D2D communication system. NEUCOM, 616, 128883. (www)

  @article{ , abstract = {Wireless mobile technology has advanced to a stage where data rates have significantly increased, and the capacity for device connectivity is poised for exponential growth. Achieving high data rate transmission in complex communication systems has become critical. Traditional iterative methods face challenges due to their high computational complexity and time-consuming processes. Graph neural networks (GNNs) have shown great potential in optimizing network resource allocation problems. GNNs exhibit strong learning capabilities in handling graph-structured data and effectively utilizing the topology of communication networks. However, current approaches primarily focus on node features, often neglecting the valuable information in edge features. We propose the Edge Feature Empowered Jumping Knowledge Graph Attention Network (EJGAT) architecture to break this limitation. This model employs a hierarchical attention mechanism and jumping knowledge connection to enhance learning capability and mitigate the over-smoothing problem. Simulation experiments illustrate that the proposed algorithm outperforms benchmarks in terms of both average sum rate and learning efficiency. Comprehensive experimental validation further highlights the algorithm’s potential applicability to various system parameters and its robustness to corrupted input features.}, archive = {J_NEUCOM}, author = {Qiushi Sun and Yang He and Yin Li and Ovanes Petrosian}, doi = {10.1016/j.neucom.2024.128883}, journal = {Neurocomputing}, month = {2}, pages = {128883}, shortjournal = {Neurocomputing}, title = {Edge feature empowered graph attention network for sum rate maximization in heterogeneous D2D communication system}, volume = {616}, year = {2025}, }
• (2025). HifiDiff: High-fidelity diffusion model for face hallucination from tiny non-frontal faces. NEUCOM, 616, 128882. (www)

  @article{ , abstract = {Obtaining a high-quality frontal facial image from a low-resolution (LR) non-frontal facial image is crucial for many facial analysis tasks. Recently, diffusion models (DMs) have made impressive progress in near-frontal face super-resolution. However, when faced with non-frontal LR faces, the existing DMs exhibit poor identity preservation and facial detail fidelity. In this paper, we present a novel high-fidelity DM named HifiDiff for simultaneously super-resolving and frontalizing tiny non-frontal facial images. It consists of a two-stage pipeline: facial preview and facial refinement. In the first stage, we pretrain a coarse restoration module to obtain a coarse high-resolution (HR) frontal face, which serves as a superior constraint condition to enhance the ability to solve complex inverse transform issues. In the second stage, we leverage the strong generation capabilities of the latent DM to refine the facial details. Specifically, we design a two-pathway control structure that consists of a facial prior guidance (FPG) module and an identity consistency (IDC) module to control the denoising process. FPG encodes multilevel features derived from latent coarse HR frontal faces and employs hybrid cross-attention to capture their intrinsic correlations with the denoiser features, thereby improving the fidelity of the facial details. IDC utilizes contrastive learning to extract high-level semantic identity-representing features to constrain the denoiser, thereby maintaining the fidelity of facial identities. Extensive experiments demonstrate that our HifiDiff produces both high-fidelity and realistic HR frontal facial images, surpassing other state-of-the-art methods in qualitative and quantitative analyses, as well as in downstream facial recognition tasks.}, archive = {J_NEUCOM}, author = {Wei Wang and Xing Wang and Yuguang Shi and Xiaobo Lu}, doi = {10.1016/j.neucom.2024.128882}, journal = {Neurocomputing}, month = {2}, pages = {128882}, shortjournal = {Neurocomputing}, title = {HifiDiff: High-fidelity diffusion model for face hallucination from tiny non-frontal faces}, volume = {616}, year = {2025}, }
• (2025). Position-aware representation learning with anatomical priors for enhanced pancreas tumor segmentation. NEUCOM, 616, 128881. (www)

  @article{ , abstract = {Accurate pancreatic tumor segmentation in CT images is crucial but challenging due to the complex anatomy and varied tumor appearance. Previous methods predominantly adopt two-stage segmentation approaches to identify and localize tumors and rely heavily on CNN-extracted texture features. In this study, we propose a tumor position-aware branch to learn pancreatic anatomical priors and integrate them into a standard 3D U-Net segmentation network. The tumor position-aware branch consists of three innovative components. Firstly, the proposed method utilizes discrete information bottleneck theory to extract compact and informative segmentation features with pancreatic anatomical priors. Secondly, we propose a coordinate position encoding transformer that encodes the spatial coordinates of each patch within the CT volume. This encoding provides the model with a global positional context, allowing it to effectively model the spatial relationships between anatomical structures. Thirdly, a probability margin regularization loss is proposed to further eliminate the interference of background patches on the learning of pancreatic anatomical positions. Our model is trained and validated our model on the public Medical Segmentation Decathlon (MSD) dataset and a private clinical dataset. Experimental results demonstrate that our approach achieves competitive performance compared to state-of-the-art (SOTA) methods in both pancreas and tumor segmentation, with Dice scores of 82.11% for the pancreas and 55.56% for the tumor on the MSD dataset. The proposed framework offers an effective solution to leverage anatomical priors and enhance representation learning for improved pancreatic tumor segmentation.}, archive = {J_NEUCOM}, author = {Kaiqi Dong and Peijun Hu and Yu Tian and Yan Zhu and Xiang Li and Tianshu Zhou and Xueli Bai and Tingbo Liang and Jingsong Li}, doi = {10.1016/j.neucom.2024.128881}, journal = {Neurocomputing}, month = {2}, pages = {128881}, shortjournal = {Neurocomputing}, title = {Position-aware representation learning with anatomical priors for enhanced pancreas tumor segmentation}, volume = {616}, year = {2025}, }
• (2025). Learning dual-pixel alignment for defocus deblurring. NEUCOM, 616, 128880. (www)

  @article{ , abstract = {It is a challenging task to recover sharp image from a single defocus blurry image in real-world applications. On many modern cameras, dual-pixel (DP) sensors create two-image views, based on which stereo information can be exploited to benefit defocus deblurring. Despite the impressive results achieved by existing DP defocus deblurring methods, the misalignment between DP image views is still not studied, leaving room for improving DP defocus deblurring. In this work, we propose a Dual-Pixel Alignment Network (DPANet) for defocus deblurring. Generally, DPANet is an encoder–decoder with skip-connections, where two branches with shared parameters in the encoder are employed to extract and align deep features from left and right views, and one decoder is adopted to fuse aligned features for predicting the sharp image. Due to that DP views suffer from different blur amounts, it is not trivial to align left and right views. To this end, we propose novel encoder alignment module (EAM) and decoder alignment module (DAM). In particular, a correlation layer is suggested in EAM to measure the disparity between DP views, whose deep features can then be accordingly aligned using deformable convolutions. DAM can further enhance the alignment of skip-connected features from encoder and deep features in decoder. By introducing several EAMs and DAMs, DP views in DPANet can be well aligned for better predicting latent sharp image. Experimental results on real-world datasets show that our DPANet is notably superior to state-of-the-art deblurring methods in reducing defocus blur while recovering visually plausible sharp structures and textures.}, archive = {J_NEUCOM}, author = {Yu Li and Yaling Yi and Xinya Shu and Dongwei Ren and Qince Li and Wangmeng Zuo}, doi = {10.1016/j.neucom.2024.128880}, journal = {Neurocomputing}, month = {2}, pages = {128880}, shortjournal = {Neurocomputing}, title = {Learning dual-pixel alignment for defocus deblurring}, volume = {616}, year = {2025}, }
• (2025). PSscheduler: A parameter synchronization scheduling algorithm for distributed machine learning in reconfigurable optical networks. NEUCOM, 616, 128876. (www)

  @article{ , abstract = {With the increasing size of training datasets and models, parameter synchronization stage puts a heavy burden on the network, and communication has become one of the main performance bottlenecks of distributed machine learning (DML). Concurrently, optical circuit switch (OCS) with high bandwidth and reconfigurable features has increasingly introduced into the construction of network topology, obtaining the reconfigurable optical networks. Actually, OCS is conducive to accelerating the parameter synchronization stage, and thus improves training performance. However, unreasonable circuit scheduling algorithm has a great impact on parameter synchronization time because of non-negligible OCS switching delay. Besides, most of the existing circuit scheduling algorithms do not effectively use the training characteristics of DML, and the performance gains are limited. Therefore, in this paper, we study the parameter synchronization scheduling algorithm in reconfigurable optical networks, and propose PSscheduler by jointly optimizing the circuit scheduling and deployment of parameter servers in parameter server (PS) architecture. Specifically, a mathematical optimization model is established first, which takes into account the deployment of parameter servers, the allocation of parameter blocks and circuit scheduling. Subsequently, the mathematical model is solved by relaxed variables and deterministic rounding approach. The results of simulation based on real DML workloads demonstrate that compared to Sunflow and HLF , PSscheduler is more stable and can reduce parameter synchronization time (PST) by up to 46.61% and 25%, respectively.}, archive = {J_NEUCOM}, author = {Ling Liu and Xiaoqiong Xu and Pan Zhou and Xi Chen and Daji Ergu and Hongfang Yu and Gang Sun and Mohsen Guizani}, doi = {10.1016/j.neucom.2024.128876}, journal = {Neurocomputing}, month = {2}, pages = {128876}, shortjournal = {Neurocomputing}, title = {PSscheduler: A parameter synchronization scheduling algorithm for distributed machine learning in reconfigurable optical networks}, volume = {616}, year = {2025}, }
• (2025). Modeling inter-modal incongruous sentiment expressions for multi-modal sarcasm detection. NEUCOM, 616, 128874. (www)

  @article{ , abstract = {Multi-modal sarcasm detection (MSD) presents a formidable and intricate endeavor. Despite strides made by extant models, two principal hurdles persist: Firstly, prevailing methodologies merely address superficial disparities between textual inputs and associated images, neglecting nuanced inter-modal combinations. Secondly, satirical instances frequently involve intricate emotional expressions, highlighting the imperative of leveraging emotional cues across modalities to discern satirical nuances. Accordingly, this research proposes the utilization of a deep graph convolutional network that integrates cross-modal mapping information to effectively identify significant incongruent sentiment expressions across various modalities for the purpose of multi-modal sarcasm detection. Specifically, we first design a cross-modal mapping network, which obtains the interaction information between these two modalities by mapping text feature vectors and image feature vectors two by two to compensate for the lack of multi-modal data in the fusion process. Additionally, we employ external knowledge of ANPS as a bridge to construct cross-correlation graphs through highly correlated satirical cues and their connection weights between image and text modalities. Afterward, the GCN architecture with retrieval-based attentional mechanisms will effectively capture satirical cues. The experiments conducted on two publicly available datasets demonstrate a significant enhancement in the performance of our method when compared to numerous contemporary models.}, archive = {J_NEUCOM}, author = {Lisong Ou and Zhixin Li}, doi = {10.1016/j.neucom.2024.128874}, journal = {Neurocomputing}, month = {2}, pages = {128874}, shortjournal = {Neurocomputing}, title = {Modeling inter-modal incongruous sentiment expressions for multi-modal sarcasm detection}, volume = {616}, year = {2025}, }
• (2025). Event-triggered robust hierarchical control for uncertain multiplayer stackelberg games via adaptive dynamic programming. NEUCOM, 616, 128873. (www)

  @article{ , abstract = {This paper investigates the event-triggered robust hierarchical control (ETRHC) problem of uncertain multi-player nonlinear systems subject to actuator faults by using adaptive dynamic programming and integral sliding mode technique. Different from existing results where the control policies of all players are updated simultaneously, a hierarchical decision-making problem is considered as a Stackelberg game. The Stackelberg game consists of a single leader and multiple followers, the leader acts a control policy in advance by considering the responses of all the followers, and each follower responds optimally to the leader’s policy. The proposed control structure comprises of two components, namely integral sliding mode control and ETRHC. In the first step, the integral sliding mode control policy is developed to cope with actuator faults and matched uncertainties, and then, the fault-free multi-player nonlinear systems with mismatched uncertainties is obtained. In the second step, by designing an appropriate performance index function for each player, the ETRHC of the fault-free multi-player nonlinear system with mismatched uncertainties is converted to an event-triggered approximate optimal control of its nominal form, and the hierarchical decision-making problem is addressed. Subsequently, the ETRHC laws are derived by solving event-triggered Hamilton–Jacobi equations with the critic-only learning. Theoretical analysis demonstrates that the integral sliding mode-based ETRHC scheme guarantees the multi-player uncertain nonlinear systems with actuator faults to be asymptotically stable. Finally, the quadrotor attitude system is adopted to verify the effectiveness of the present scheme.}, archive = {J_NEUCOM}, author = {Yongwei Zhang and Bo Zhao and Derong Liu and Marios M. Polycarpou and Shiguo Peng and Shunchao Zhang}, doi = {10.1016/j.neucom.2024.128873}, journal = {Neurocomputing}, month = {2}, pages = {128873}, shortjournal = {Neurocomputing}, title = {Event-triggered robust hierarchical control for uncertain multiplayer stackelberg games via adaptive dynamic programming}, volume = {616}, year = {2025}, }
• (2025). Efficient motion feature aggregation for optical flow via locality-sensitive hashing. NEUCOM, 616, 128870. (www)

  @article{ , abstract = {Optical flow estimation aims to find the 2D motion field by identifying the corresponding pixels between two images. With the tremendous progress of deep learning, previous works rely on CNNs to regress the pixelwise correspondence. However, it is still challenging to estimate displacements at occlusions, where one point is imaged in the current frame but not in the next frame. To solve this problem, recent work proposes to aggregate motion features at a global scale. The aggregation allows motion information to be passed from non-occluded pixels to occluded pixels, which helps resolve ambiguities caused by occlusions. Though the global motion aggregation (GMA) works well, its computational complexity is quadratic to the input resolution. The complexity can be a limitation of GMA when applied to higher resolution. In this paper, we propose an efficient motion feature aggregation (EMFA) module, to approximate the original GMA at a much lower computational cost. The core insight of EMFA originates from a clustering scheme, where the motion information of one point is aggregated from the points in the same cluster. The clustering is implemented by a traditional hashing algorithm, called locality-sensitive hashing (LSH). Finally, we apply the EMFA module to the optical flow estimation at high resolution. Experimental results on publicly available datasets show that our method outperforms the state-of-the-art ones.}, archive = {J_NEUCOM}, author = {Weihai Chen and Ziyang Liu and Xingming Wu and Zhong Liu and Zhengguo Li}, doi = {10.1016/j.neucom.2024.128870}, journal = {Neurocomputing}, month = {2}, pages = {128870}, shortjournal = {Neurocomputing}, title = {Efficient motion feature aggregation for optical flow via locality-sensitive hashing}, volume = {616}, year = {2025}, }
• (2025). Enhancing open N-ary information extraction using relation embedding and multihead relation attention mechanism. NEUCOM, 616, 128867. (www)

  @article{ , abstract = {Open Information Extraction (Open IE) is the task of identifying structured and machine-readable information from natural language text within an open domain context. This research area has gained significant importance in the field of natural language processing (NLP), attracting considerable attention for its potential to extract valuable information from unstructured textual data. Previous investigations heavily relied on manual extraction patterns and various NLP tools. While these methods often produce errors that accumulate and propagate throughout the systems, ultimately affecting the accuracy of the results. Moreover, recent Open IE studies have focused on extracting binary relations involving two entities. However, these binary approaches occasionally lead to the omission of essential information in the text, preventing a deeper comprehension of the content. This limitation arises from the fact that real-world relations often involve multiple entities, but binary approaches may oversimplify these relations and miss additional details crucial for a thorough understanding of text. To address these challenges, our study introduces an innovative system called “Open N-ary Information EXtraction (ONIEX).” This system incorporates two novel techniques: multihead relation attention mechanism and relation embedding. Multihead relation attention, in combination with relation embedding, enables the system to focus on relations extracted through the SpanBERT model and accurately identify associated entities for each relation. The ONIEX system's superior performance is substantiated through extensive experiments conducted on the OpenIE4 and LSOIE datasets, benchmark datasets for Open n-ary Information Extraction (Open n-ary IE). The results demonstrate the superiority of the ONIEX system over the existing state-of-the-art systems.}, archive = {J_NEUCOM}, author = {Mitra Isaee and Afsaneh Fatemi and Mohammadali Nematbakhsh}, doi = {10.1016/j.neucom.2024.128867}, journal = {Neurocomputing}, month = {2}, pages = {128867}, shortjournal = {Neurocomputing}, title = {Enhancing open N-ary information extraction using relation embedding and multihead relation attention mechanism}, volume = {616}, year = {2025}, }
• (2025). MDTNet: Partial transformer with degradation-aware module for restoring old photos with multiple degradations. NEUCOM, 616, 128866. (www)

  @article{ , abstract = {Old photos preserve the memories of the old days and have special meanings in life. However, they often suffer from severe and complex degradations, such as cracks, dirt, and noise. Old photo restoration is thus a meaningful but challenging task. Most of the current deep learning methods to repair corrupted old photos focus on one single type of degradation or limited number of degradations. The lack of proper old photo datasets makes the digital restoration problem even more challenging. In this paper, we propose MDTNet, which is a novel transformer-based method for restoring old photos with multiple degradations. MDTNet is an efficient end-to-end transformer architecture with only a single encoder and a decoder to complete the restoration. Specifically, a novel partial transformer encoder is designed with partial multi-head self-attention to extract local-global context information from the valid masked area, which represents the unbroken area. In addition, a degradation-aware module is proposed in the decoder to automatically learn different degradation information, to improve the effect of multiple degradation restoration. Moreover, a synthetic old photo dataset — the SynOld dataset — was built, which simulates the multiple degradations of real old photos. Experimental results show that the proposed method outperforms the state-of-the-art methods on both public datasets and the custom-developed dataset. The code is available at https://github.com/YuanZhaoc/MDTNet .}, archive = {J_NEUCOM}, author = {Liqin Cao and Yuan Zhao and Fan Zhang and Xuan Zhang and Ju Hua Liu and Yanfei Zhong}, doi = {10.1016/j.neucom.2024.128866}, journal = {Neurocomputing}, month = {2}, pages = {128866}, shortjournal = {Neurocomputing}, title = {MDTNet: Partial transformer with degradation-aware module for restoring old photos with multiple degradations}, volume = {616}, year = {2025}, }
• (2025). Multimodal invariant feature prompt network for brain tumor segmentation with missing modalities. NEUCOM, 616, 128847. (www)

  @article{ , abstract = {Accurate brain tumor image segmentation can significantly assist doctors in clinical diagnosis and treatment. The complementary information from multimodal magnetic resonance imaging provides more comprehensive information on brain tumors for lesion segmentation. However, the complexity of clinical scenarios may lead to modal loss or damage, thereby reducing segmentation performance. Therefore, brain tumor image segmentation becomes a challenging task in the absence of modalities. In recent years, leveraging interactive mechanisms to learn shared feature representations across multiple modalities has become a dominant approach, especially for enhancing segmentation accuracy in scenarios with incomplete modalities. However, a significant challenge in most existing methods is their inability to effectively capture complete multimodal information due to the loss of modality-specific details during interaction processes. To overcome this challenge, we introduce a novel network, referred to as the Multimodal Invariant Feature Prompt Network (MIFPN), which enhances learning by incorporating modality prompts within multimodal interactions, thus ensuring a more comprehensive acquisition of modality information. This network aims to introduce modality-invariant feature prompts in modal information interaction, guiding the model to learn the missing modality information and facilitating integration. To fully leverage modal information, we learned both modality-invariant and modality-specific information, and fused them during the prompt learning process to obtain a more comprehensive understanding of modal information. Considering the different contributions of various modalities to the segmentation results, we designed modality-aware masks and modality selection strategies to merge shallow features of the encoder. MIFPN effectively mitigates the brain tumor image segmentation problem under missing modalities by learning from missing modality prompts and modality-aware fusion. Extensive experiments conducted on the BraTS2020 and BraTS2018 datasets show that our method outperforms state-of-the-art methods in most scenarios across 15 missing modality configurations. The code is available at https://github.com/diaoyq121/MIFPN .}, archive = {J_NEUCOM}, author = {Yueqin Diao and Huihui Fang and Hanyi Yu and Fan Li and Yanwu Xu}, doi = {10.1016/j.neucom.2024.128847}, journal = {Neurocomputing}, month = {2}, pages = {128847}, shortjournal = {Neurocomputing}, title = {Multimodal invariant feature prompt network for brain tumor segmentation with missing modalities}, volume = {616}, year = {2025}, }
• (2025). Observer-based fully distributed bipartite consensus of multiagent systems with disturbance rejection. NEUCOM, 616, 128842. (www)

  @article{ , abstract = {Observer-based output feedback control method is utilized to deal with the bipartite consensus problem of multiagent systems (MASs) suffering deterministic disturbances. Based on the leaderless and leader-follower methods, two fully distributed observer-based output feedback controllers are devised to guarantee the bipartite consensus of MASs. Moreover, due to the limited bandwidth of communication channels in practical systems, an event-triggered output feedback controller for the bipartite consensus of MASs is also developed and can guarantee that Zeno behavior does not occur. Finally, the effectiveness and advantages of the control protocols are verified via illustrate examples.}, archive = {J_NEUCOM}, author = {Jian-Qiao Wang and Jin-Liang Wang and Xueming Dong}, doi = {10.1016/j.neucom.2024.128842}, journal = {Neurocomputing}, month = {2}, pages = {128842}, shortjournal = {Neurocomputing}, title = {Observer-based fully distributed bipartite consensus of multiagent systems with disturbance rejection}, volume = {616}, year = {2025}, }
• (2025). Deep multi-similarity hashing via label-guided network for cross-modal retrieval. NEUCOM, 616, 128830. (www)

  @article{ , abstract = {Due to low storage cost and efficient retrieval advantages, hashing technologies have gained broad attention in the field of cross-modal retrieval in recent years. However, most current cross-modal hashing usually employs random sampling or semi-hard negative mining to construct training batches for model optimization, which ignores the distribution relationships between raw samples, generating redundant and unbalanced pairs, and resulting in sub-optimal embedding spaces. In this work, we address this dilemma with a novel deep cross-modal hashing framework, called Deep Multi-similarity Hashing via Label-Guided Networks (DMsH-LN), to learn a high separability public embedding space and generate discriminative binary descriptors. Specifically, by utilizing pair mining and weighting to jointly calculate self-similarity and relative similarity between pairs, the multi-similarity loss is extended to cross-modal hashing to alleviate the negative impacts caused by redundant and imbalanced samples on hash learning, enhancing the distinguishing ability of the obtained discrete codes. Besides, to capture fine-grained semantic supervised signals, the Label-guided Network is proposed to learn class-specific semantic signals, which could effectively guide the parameter optimization of the Image Network and Text Network. Extensive experiments are conducted on four benchmark datasets, which demonstrate that the DMsH-LN framework achieves excellent retrieval performance. The source codes of DMsH-LN are downloaded from https://github.com/QinLab-WFU/DMsH-LN .}, archive = {J_NEUCOM}, author = {Lei Wu and Qibing Qin and Jinkui Hou and Jiangyan Dai and Lei Huang and Wenfeng Zhang}, doi = {10.1016/j.neucom.2024.128830}, journal = {Neurocomputing}, month = {2}, pages = {128830}, shortjournal = {Neurocomputing}, title = {Deep multi-similarity hashing via label-guided network for cross-modal retrieval}, volume = {616}, year = {2025}, }
• (2025). Generalization and risk bounds for recurrent neural networks. NEUCOM, 616, 128825. (www)

  @article{ , abstract = {Recurrent Neural Networks (RNNs) have achieved great success in the prediction of sequential data. However, their theoretical studies are still lagging behind because of their complex interconnected structures. In this paper, we establish a new generalization error bound for vanilla RNNs, and provide a unified framework to calculate the Rademacher complexity that can be applied to a variety of loss functions. When the ramp loss is used, we show that our bound is tighter than the existing bounds based on the same assumptions on the Frobenius and spectral norms of the weight matrices and a few mild conditions. Our numerical results show that our new generalization bound is the tightest among all existing bounds in three public datasets. Our bound improves the second tightest one by an average percentage of 13.80% and 3.01% when the tanh and ReLU activation functions are used, respectively. Moreover, we derive a sharp estimation error bound for RNN-based estimators obtained through empirical risk minimization (ERM) in multi-class classification problems when the loss function satisfies a Bernstein condition.}, archive = {J_NEUCOM}, author = {Xuewei Cheng and Ke Huang and Shujie Ma}, doi = {10.1016/j.neucom.2024.128825}, journal = {Neurocomputing}, month = {2}, pages = {128825}, shortjournal = {Neurocomputing}, title = {Generalization and risk bounds for recurrent neural networks}, volume = {616}, year = {2025}, }
• (2025). Selective-hop graph neural networks for semi-supervised node classification. NEUCOM, 616, 128812. (www)

  @article{ , abstract = {Message-passing Graph Neural Networks (GNNs) are remarkably successful in graph representation learning. Most of them propagate and aggregate information within fixed-size receptive-fields (RFs) for all nodes. In this paper, our empirical studies discover that different-sized RFs are often desired for distinct graphs and target nodes. To this end, we propose a dynamic selection mechanism by allowing each node to adaptively adjust the size of its RFs according to the graph properties. In particular, we design an architectural unit called Selective-Hop (SH) block, in which information from neighborhoods within different hops are merged using dual attention mechanisms: (i) node-wise attention on each network branch to capture mutual information between the subgraphs and the entire graph; and (ii) channel-wise attention on different network branches to detect the significance of different hops at the node-level granularity. As such, the SH block determines neighbors flexibly without the constraints of hops. By stacking several SH blocks in GNN style, we instantiate two new GNN variants, SHGCN and SHSGC. Empirical results on eight benchmark datasets for semi-supervised node classification reveal that our new models achieve a considerable improvement over the state-of-the-art methods under low-homophily or heterophily and maintain competitive performance under homophily.}, archive = {J_NEUCOM}, author = {Hui Yan and Guoguo Ai and Xin Li}, doi = {10.1016/j.neucom.2024.128812}, journal = {Neurocomputing}, month = {2}, pages = {128812}, shortjournal = {Neurocomputing}, title = {Selective-hop graph neural networks for semi-supervised node classification}, volume = {616}, year = {2025}, }
• (2025). Virtual sample generation for small sample learning: A survey, recent developments and future prospects. NEUCOM, 615, 128934. (www)

  @article{ , abstract = {Virtual sample generation (VSG) technology aims to generate virtual samples based on real samples, in order to expand the size of the datasets and improve model performance. However, there is limited research summarizing VSG technology, which motivates this paper. In recent years, VSG technology has grown as a crucial tool for augmenting datasets and enhancing model performance, particularly in the fields like image recognition, medicine, and quality control where small datasets are common issues. This paper aims to provide an updated review of VSG technology, focusing on three key techniques which are important for small sample analysis studies, including sampling-based, information diffusion-based, and Generative Adversarial Networks (GANs)-based technology. In this review, we seek to identify the key trends in this field and to provide insights regarding the opportunities and challenges.}, archive = {J_NEUCOM}, author = {Jianming Wen and Ao Su and Xiaolin Wang and Hao Xu and Jijie Ma and Kang Chen and Xinyang Ge and Zisheng Xu and Zhong Lv}, doi = {10.1016/j.neucom.2024.128934}, journal = {Neurocomputing}, month = {1}, pages = {128934}, shortjournal = {Neurocomputing}, title = {Virtual sample generation for small sample learning: A survey, recent developments and future prospects}, volume = {615}, year = {2025}, }
• (2025). A novel random fast multi-label deep forest classification algorithm. NEUCOM, 615, 128903. (www)

  @article{ , abstract = {With the development of big data, a large amount of multi-label data has emerged in various fields. The use of deep classification models to solve multi-label classification problems is the current research hotspot. However, the existing deep classification algorithms still face two great challenges in solving the multi-label classification problem: the poor multi-label classification accuracy and the high time cost. In order to solve the above challenges, this paper proposes a novel random fast multi-label deep forest classification algorithm (RFDF). Firstly, the label-by-label approach is adopted for the training and prediction of base classifiers, while the multi-label cascade model is constructed, so as to obtain better multi-label classification results. Then, an adaptive method is proposed to determine the number of base classifiers, and a random features partitioning strategy is provided to generate the initial base classifier, achieving parallel training of the base classifier. Finally, the proposed RFDF is compared with the existing advanced algorithms, and the experimental results show that the RFDF achieves better results on most of the metrics of the ten datasets, and obtains lower time cost compared to VDSDF and better stability.}, archive = {J_NEUCOM}, author = {Tao Li and Jie-Xue Jia and Jian-Yu Li and Xian-Wei Xin and Jiu-Cheng Xu}, doi = {10.1016/j.neucom.2024.128903}, journal = {Neurocomputing}, month = {1}, pages = {128903}, shortjournal = {Neurocomputing}, title = {A novel random fast multi-label deep forest classification algorithm}, volume = {615}, year = {2025}, }
• (2025). DoA-ViT: Dual-objective affine vision transformer for data insufficiency. NEUCOM, 615, 128896. (www)

  @article{ , abstract = {Vision Transformers (ViTs) excel in large-scale image recognition tasks but struggle with limited data due to ineffective patch-level local information utilization. Existing methods focus on enhancing local representations at the model level but often treat all features equally, leading to noise from irrelevant information. Effectively distinguishing between discriminative features and irrelevant information helps minimize the interference of noise at the model level. To tackle this, we introduce Dual-objective Affine Vision Transformer (DoA-ViT), which enhances ViTs for data-limited tasks by improving feature discrimination. DoA-ViT incorporates a learnable affine transformation that associates transformed features with class-specific ones while preserving their intrinsic features. Additionally, an adaptive patch-based enhancement mechanism is designed to assign importance scores to patches, minimizing the impact of irrelevant information. These enhancements can be seamlessly integrated into existing ViTs as plug-and-play components. Extensive experiments on small-scale datasets show that DoA-ViT consistently outperforms existing methods, with visualization results highlighting its ability to identify critical image regions effectively.}, archive = {J_NEUCOM}, author = {Qiang Ren and Junli Wang}, doi = {10.1016/j.neucom.2024.128896}, journal = {Neurocomputing}, month = {1}, pages = {128896}, shortjournal = {Neurocomputing}, title = {DoA-ViT: Dual-objective affine vision transformer for data insufficiency}, volume = {615}, year = {2025}, }
• (2025). Superpixel semantics representation and pre-training for vision–language tasks. NEUCOM, 615, 128895. (www)

  @article{ , abstract = {The key to integrating visual language tasks is to establish a good alignment strategy. Recently, visual semantic representation has achieved fine-grained visual understanding by dividing grids or image patches. However, the coarse-grained semantic interactions in image space should not be ignored, which hinders the extraction of complex contextual semantic relations at the scene boundaries. This paper proposes superpixels as comprehensive and robust visual primitives, which mine coarse-grained semantic interactions by clustering perceptually similar pixels, speeding up the subsequent processing of primitives. To capture superpixel-level semantic features, we propose a Multiscale Difference Graph Convolutional Network (MDGCN). It allows parsing the entire image as a fine-to-coarse visual hierarchy. To reason actual semantic relations, we reduce potential noise interference by aggregating difference information between adjacent graph nodes. Finally, we propose a multi-level fusion rule in a bottom-up manner to avoid understanding deviation by mining complementary spatial information at different levels. Experiments show that the proposed method can effectively promote the learning of multiple downstream tasks. Encouragingly, our method outperforms previous methods on all metrics.}, archive = {J_NEUCOM}, author = {Siyu Zhang and Yeming Chen and Yaoru Sun and Fang Wang and Jun Yang and Lizhi Bai and Shangce Gao}, doi = {10.1016/j.neucom.2024.128895}, journal = {Neurocomputing}, month = {1}, pages = {128895}, shortjournal = {Neurocomputing}, title = {Superpixel semantics representation and pre-training for vision–language tasks}, volume = {615}, year = {2025}, }
• (2025). Dual-dimensional contrastive learning for incomplete multi-view clustering. NEUCOM, 615, 128892. (www)

  @article{ , abstract = {Incomplete multi-view clustering (IMVC) is a critical task in real-world applications, where missing data in some views can severely limit the ability to leverage complementary information across views. This issue leads to incomplete sample representations, hindering model performance. Current contrastive learning methods for IMVC exacerbate the problem by directly constructing data pairs from incomplete samples, ignoring essential information and resulting in class collisions, where samples from different classes are incorrectly grouped together due to a lack of label guidance. These challenges are particularly detrimental in fields like recommendation systems and bioinformatics, where accurate clustering of high-dimensional and incomplete data is essential for decision-making. To address these issues, we propose Dual-dimensional Contrastive Learning (DCL), an online IMVC model that fills missing values through multi-view consistency transfer, enabling simultaneous clustering and representation learning via instance-level and cluster-level contrastive learning in both row and column spaces. DCL mitigates class collision issues by generating high-confidence pseudo-labels and using an optimal transport matrix, significantly improving clustering accuracy. Extensive experiments demonstrate that DCL achieves state-of-the-art results across five datasets. The code is available at https://github.com/2251821381/DCL .}, archive = {J_NEUCOM}, author = {Zhengzhong Zhu and Chujun Pu and Xuejie Zhang and Jin Wang and Xiaobing Zhou}, doi = {10.1016/j.neucom.2024.128892}, journal = {Neurocomputing}, month = {1}, pages = {128892}, shortjournal = {Neurocomputing}, title = {Dual-dimensional contrastive learning for incomplete multi-view clustering}, volume = {615}, year = {2025}, }
• (2025). Hierarchical convolutional neural networks with post-attention for speech emotion recognition. NEUCOM, 615, 128879. (www)

  @article{ , abstract = {Speech emotion recognition (SER) is a key prerequisite for natural human–computer interaction. However, existing SER systems still face great challenges, particularly in the extraction of discriminative and high-quality emotional features. To address this challenge, this study proposes hc-former, a hierarchical convolutional neural network (CNN) with post-attention. Unlike traditional CNNs and recurrent neural networks (RNNs), our model adeptly extracts potent class-discriminative features that integrate spatiotemporal information and long-term dependence. The class-discriminative features extracted by hc-former, which emphasize both interclass separation and intraclass compactness, can more effectively represent different class emotions often confused with one another, leading to superior classification results. Our experimental results further indicate the exceptional performance of hc-former for SER on benchmark datasets, surpassing other peer models in terms of performance while utilizing fewer parameters.}, archive = {J_NEUCOM}, author = {Yonghong Fan and Heming Huang and Henry Han}, doi = {10.1016/j.neucom.2024.128879}, journal = {Neurocomputing}, month = {1}, pages = {128879}, shortjournal = {Neurocomputing}, title = {Hierarchical convolutional neural networks with post-attention for speech emotion recognition}, volume = {615}, year = {2025}, }
• (2025). CNN explanation methods for ordinal regression tasks. NEUCOM, 615, 128878. (www)

  @article{ , abstract = {The use of Convolutional Neural Network (CNN) models for image classification tasks has gained significant popularity. However, the lack of interpretability in CNN models poses challenges for debugging and validation. To address this issue, various explanation methods have been developed to provide insights into CNN models. This paper focuses on the validity of these explanation methods for ordinal regression tasks, where the classes have a predefined order relationship. Different modifications are proposed for two explanation methods to exploit the ordinal relationships between classes: Grad-CAM based on Ordinal Binary Decomposition (GradOBD-CAM) and Ordinal Information Bottleneck Analysis (OIBA). The performance of these modified methods is compared to existing popular alternatives. Experimental results demonstrate that GradOBD-CAM outperforms other methods in terms of interpretability for three out of four datasets, while OIBA achieves superior performance compared to IBA.}, archive = {J_NEUCOM}, author = {Javier Barbero-Gómez and Ricardo P.M. Cruz and Jaime S. Cardoso and Pedro A. Gutiérrez and César Hervás-Martínez}, doi = {10.1016/j.neucom.2024.128878}, journal = {Neurocomputing}, month = {1}, pages = {128878}, shortjournal = {Neurocomputing}, title = {CNN explanation methods for ordinal regression tasks}, volume = {615}, year = {2025}, }
• (2025). Adaptive selection of spectral–spatial features for hyperspectral image classification using a modified-CBAM-based network. NEUCOM, 615, 128877. (www)

  @article{ , abstract = {Convolutional neural networks (CNNs) have demonstrated strong capabilities in hyperspectral image (HSI) classification. However, it is still a challenge to adaptively adjust the size of the receptive fields (RFs) of CNNs base on the information of different scales in HSI to achieve adaptive selection of spectral–spatial features. In the paper, we modify the convolutional block attention module (CBAM) and propose a modified-CBAM-based network (MCNet) to adaptively select spectral–spatial features for HSI classification. In particular, the modified CBAM not only enables the model to adjust its RF size according to the information of different scales in HSI, but also enables the model to achieve a joint focus on important spectral and spatial features. This is very important to adaptively select more descriptive and discriminative spectral–spatial features. The proposed MCNet is compared with currently popular methods on Indian Pines, Kennedy Space Center, University of Pavia, and Botswana HSI datasets. The results show that MCNet has better classification results than other methods on overall accuracy, average accuracy, and Kappa.}, archive = {J_NEUCOM}, author = {He Fu and Cailing Wang and Zhanlong Chen}, doi = {10.1016/j.neucom.2024.128877}, journal = {Neurocomputing}, month = {1}, pages = {128877}, shortjournal = {Neurocomputing}, title = {Adaptive selection of spectral–spatial features for hyperspectral image classification using a modified-CBAM-based network}, volume = {615}, year = {2025}, }
• (2025). Instance-dependent cost-sensitive parametric learning. NEUCOM, 615, 128875. (www)

  @article{ , abstract = {Instance-dependent cost-sensitive learning addresses classification problems where each observation has a different misclassification cost. In this paper, we propose cost-sensitive parametric models to minimize the expectation of losses. A loss function incorporating the misclassification costs is defined, which serves as the objective function for obtaining cost-sensitive parameter estimators. The consistency and asymptotic normality of these estimators are established under general conditions, theoretically demonstrating their good performance. Additionally, we derive bounds for the bias introduced when regularizing the optimization problem, which is generally necessary in practice. To conclude, the effectiveness of the proposed estimators is evaluated through an extensive novel simulation study and the analysis of five real data sets, further demonstrating their proficiency in practical settings.}, archive = {J_NEUCOM}, author = {Jorge C-Rella and Gerda Claeskens and Ricardo Cao and Juan M. Vilar}, doi = {10.1016/j.neucom.2024.128875}, journal = {Neurocomputing}, month = {1}, pages = {128875}, shortjournal = {Neurocomputing}, title = {Instance-dependent cost-sensitive parametric learning}, volume = {615}, year = {2025}, }
• (2025). FPGA-based component-wise LSTM training accelerator for neural granger causality analysis. NEUCOM, 615, 128871. (www)

  @article{ , abstract = {Component-wise LSTM (cLSTM) constitutes multiple LSTM cells of distinct parameters, which has particular benefits of functional Magnetic Resonance Imaging (fMRI)-based neural Granger causality (NGC) analysis for the human brain. Back-propagation through time training on CPU and GPU suffers from low utilization due to inherent data dependencies within the LSTM cell. Moreover, batch 1 cLSTM training and few weight reuses across input feature maps worsen such a utilization problem. To this end, this study provides an FPGA-based training solution for cLSTM-based NGC analysis. The proposed cLSTM training accelerator identifies different data dependencies in forward and backward paths, and features two key components: (1) a fine-grained pipeline within the LSTM cell that achieves the lowest initial interval, and (2) a coarse-grained pipeline that trains input feature sequences across different LSTM cells in parallel. Experiments on the DAN sub-brain network from the COBRE dataset demonstrate the efficacy of FPGA-based cLSTM training, which achieves microseconds iteration latency compared with milliseconds on general-purpose platforms, e.g., 465 × and 216 × faster than Intel Core 13900K CPU and Nvidia RTX 2080Ti respectively. To the best of our knowledge, this work is the first to demonstrate LSTM training on FPGA, significantly accelerating the analysis and modeling of complex brain networks, and offering valuable advancements for neuroscience research at the edge.}, archive = {J_NEUCOM}, author = {Chuliang Guo and Yufei Chen and Yu Fu}, doi = {10.1016/j.neucom.2024.128871}, journal = {Neurocomputing}, month = {1}, pages = {128871}, shortjournal = {Neurocomputing}, title = {FPGA-based component-wise LSTM training accelerator for neural granger causality analysis}, volume = {615}, year = {2025}, }
• (2025). Extending the learning using privileged information paradigm to logistic regression. NEUCOM, 615, 128869. (www)

  @article{ , abstract = {Learning using privileged information paradigm is a learning scenario that exploits privileged features, available at training time, but not at prediction, as additional information for training models. This paper delves into the learning of logistic regression models using privileged information. We provide two new algorithms. For its development, the parameters of a conventional logistic regression trained with all available features, privileged and regular, are projected onto the parameter space associated to regular features (available at training and prediction time). The projection to obtain the model parameters is performed by the minimization of two different loss functions governed by logit terms and posterior probabilities. In addition, a metric is proposed to determine whether the use of privileged information can enhance performance. Experimental results report improvements of our proposals over the performance of conventional logistic regression learned without privileged information.}, archive = {J_NEUCOM}, author = {Mario Martínez-García and Susana García-Gutierrez and Lasai Barreñada and Iñaki Inza and Jose A. Lozano}, doi = {10.1016/j.neucom.2024.128869}, journal = {Neurocomputing}, month = {1}, pages = {128869}, shortjournal = {Neurocomputing}, title = {Extending the learning using privileged information paradigm to logistic regression}, volume = {615}, year = {2025}, }
• (2025). Data-driven distributed output consensus control for multi-agent systems with unknown internal state. NEUCOM, 615, 128868. (www)

  @article{ , abstract = {This paper discusses the topic of optimal output consensus control of linear time-invariant discrete-time multi-agent systems (MASs). The attainment of optimal output consensus control for MASs hinges upon the resolution of the interconnected Hamilton–Jacobi-Bellman equation, a task typically precluded by the inherent intractability of analytical solutions. Furthermore, most real-world systems are too complex to obtain the internal states of systems. To address these issues, a modified deep Q-learning network is constructed using current and historical system data rather than a precise model of the system. First, reconstructing the internal state of each agent using an adaptive distributed observer based on output feedback prevents the system instability brought on by the augmented systems. Then the local error system of the agent can be redefined. Based on the redefined error system, a data-driven adaptive dynamic programming (ADP) method is introduced, realized by using the actor–critic neural network structure. In addition, an experience replay strategy is proposed to reduce the propagation of estimation bias and improve the learning speed. Finally, the comparative numerical simulations substantiate the efficacy of the proposed algorithm in a quantifiable manner.}, archive = {J_NEUCOM}, author = {Cuijuan Zhang and Lianghao Ji and Shasha Yang and Xing Guo and Huaqing Li}, doi = {10.1016/j.neucom.2024.128868}, journal = {Neurocomputing}, month = {1}, pages = {128868}, shortjournal = {Neurocomputing}, title = {Data-driven distributed output consensus control for multi-agent systems with unknown internal state}, volume = {615}, year = {2025}, }
• (2025). Token embeddings augmentation benefits parameter-efficient fine-tuning under long-tailed distribution. NEUCOM, 615, 128853. (www)

  @article{ , abstract = {Pre-trained vision-language models, particularly those utilizing CLIP, have advanced various visual tasks. Parameter-Efficient Fine-Tuning (PEFT) on such models is a mainstream trend for downstream tasks. Despite advancements, long-tailed distribution still hampers image recognition performance in current PEFT schemes. Therefore, this paper proposes Token Embeddings Augmentation (TEA) to tackle long-tailed learning under PEFT paradigm. Based on patch token semantic mining, TEA uncovers category-specific semantic details within patch tokens to enhance token embeddings, named Patch-based Embeddings Augmentation (PEA). Then, a Probability Gate (PG) strategy is designed to effectively enrich semantic information of tail categories using enhanced embeddings. A Token Embeddings Consistency (TEC) loss is further introduced to prioritize category semantic information within tokens. Extensive experiments on multiple long-tailed distribution datasets show that our method improves the performance of various PEFT methods with different classification loss functions, especially for tail categories. Our optimal approach achieves the state-of-the-art results on multiple datasets with negligible parameters or inference latency, thus enhancing the practicality of PEFT in long-tailed distributions.}, archive = {J_NEUCOM}, author = {Weiqiu Wang and Zining Chen and Zhicheng Zhao and Fei Su}, doi = {10.1016/j.neucom.2024.128853}, journal = {Neurocomputing}, month = {1}, pages = {128853}, shortjournal = {Neurocomputing}, title = {Token embeddings augmentation benefits parameter-efficient fine-tuning under long-tailed distribution}, volume = {615}, year = {2025}, }
• (2025). Deep feature response discriminative calibration. NEUCOM, 615, 128848. (www)

  @article{ , abstract = {Deep neural networks (DNNs) have numerous applications across various domains. Several optimization techniques, such as ResNet and SENet, have been proposed to improve model accuracy. These techniques improve the model performance by adjusting or calibrating feature responses according to a uniform standard. However, they lack the discriminative calibration for different features, thereby introducing limitations in the model output. Therefore, we propose a method that discriminatively calibrates feature responses. The preliminary experimental results indicate that the neural feature response follows a Gaussian distribution. Consequently, we compute confidence values by employing the Gaussian probability density function, and then integrate these values with the original response values. The objective of this integration is to improve the feature discriminability of the neural feature response. Based on the calibration values, we propose a plugin-based calibration module incorporated into a modified ResNet architecture, termed Response Calibration Networks (ResCNet). Extensive experiments on datasets like CIFAR-10, CIFAR-100, SVHN, and ImageNet demonstrate the effectiveness of the proposed approach.}, archive = {J_NEUCOM}, author = {Wenxiang Xu and Tian Qiu and Linyun Zhou and Zunlei Feng and Mingli Song and Huiqiong Wang}, doi = {10.1016/j.neucom.2024.128848}, journal = {Neurocomputing}, month = {1}, pages = {128848}, shortjournal = {Neurocomputing}, title = {Deep feature response discriminative calibration}, volume = {615}, year = {2025}, }
• (2025). NN2ViT: Neural networks and vision transformers based approach for visual anomaly detection in industrial images. NEUCOM, 615, 128845. (www)

  @article{ , abstract = {Ensuring product quality through automated anomaly detection is crucial in manufacturing. Traditional methods often struggle to capture both local and global features effectively, relying heavily on predefined templates that limit their adaptability and accuracy. To address these challenges, this study propose NN2ViT, a novel approach that integrates a Single Shot Detector (SSD) for local feature detection and the Segment Anything Model (SAM) for global feature segmentation. This integration allows for a comprehensive analysis of anomalies in industrial images. Our method improves anomaly segmentation performance by fine-tuning SAM for precise segmentation in industrial product images. Experiments on the MVTec benchmark dataset demonstrate that NN2ViT outperforms traditional models and achieved the highest 95.54% and 96.23% Image AUROC and AP scores, respectively thus enhancing interpretability and adaptability to various anomaly patterns. This research presents a significant advancement in manufacturing quality control, contributing to improved product quality and operational efficiency.}, archive = {J_NEUCOM}, author = {Junaid Abdul Wahid and Muhammad Ayoub and Mingliang Xu and Xiaoheng Jiang and Lei Shi and Shabir Hussain}, doi = {10.1016/j.neucom.2024.128845}, journal = {Neurocomputing}, month = {1}, pages = {128845}, shortjournal = {Neurocomputing}, title = {NN2ViT: Neural networks and vision transformers based approach for visual anomaly detection in industrial images}, volume = {615}, year = {2025}, }
• (2025). Aircraft trajectory prediction in terminal airspace with intentions derived from local history. NEUCOM, 615, 128843. (www)

  @article{ , abstract = {Aircraft trajectory prediction aims to estimate the future movements of aircraft in a scene, which is a crucial step for intelligent air traffic management such as capacity estimation and conflict detection. Current approaches primarily rely on inputting absolute locations, which improves the prediction accuracy but limits the model’s generalization ability to unseen environments. To bridge the gap, we propose to alternatively learn aircraft’s intentions from a repository of historical trajectories. Based on the observation that aircraft traveling through the same airspace may exhibit comparable behaviors, we utilize a location-adaptive threshold to identify nearby neighbors for a given query aircraft within the repository. The retrieved candidates are next filtered based on contextual information, such as landing time and landing direction, to eliminate less relevant components. The resulting set of nearby candidates are referred to as the local history, which emphasizes the modeling of aircraft’s local behavior. Moreover, an attention-based local history encoder is presented to aggregate information from all nearby candidates to generate a latent feature for capturing the aircraft’s intention. This latent feature is robust to normalized input trajectories, relative to the current location of the target aircraft, thus improving the model’s generalization capability to unseen areas. Our proposed intention modeling method is model-agnostic, which can be leveraged as an additional condition by any trajectory prediction model for improved robustness and accuracy. For evaluation, we integrate the intention modeling component into our previous diffusion-based aircraft trajectory prediction framework. We conduct experiments on two real-world aircraft trajectory datasets in both towered and non-towered terminal airspace. The experimental results show that our method captures various maneuvering patterns effectively, outperforming existing methods by a large margin in terms of both ADE and FDE.}, archive = {J_NEUCOM}, author = {Yifang Yin and Sheng Zhang and Yicheng Zhang and Yi Zhang and Shili Xiang}, doi = {10.1016/j.neucom.2024.128843}, journal = {Neurocomputing}, month = {1}, pages = {128843}, shortjournal = {Neurocomputing}, title = {Aircraft trajectory prediction in terminal airspace with intentions derived from local history}, volume = {615}, year = {2025}, }
• (2025). Event-triggered explorized IRL-based decentralized fault-tolerant guaranteed cost control for interconnected systems against actuator failures. NEUCOM, 615, 128837. (www)

  @article{ , abstract = {This paper presents a novel data-based decentralized guaranteed cost (DGC) fault tolerant control (FTC) scheme for the large-scale systems subject to actuator faults and mismatched interconnection. First, the FTC issues of interconnected systems are converted into a series of near optimal event-triggered control (ETC) methods for isolated subsystems via constructing a modified performance index function of each subsystem. By means of adaptive dynamic programming (ADP) algorithm, the upper bound of performance index function of large-scale systems can be obtained by solving the Hamilton-Jacobi-Bellman (HJB) equation of each auxiliary subsystem. Second, according to the proposed ADP-based decentralized approach and utilizing the event-based synchronous integral reinforcement learning (IRL) algorithm, a model-free guaranteed cost (GC) FTC approach is developed for interconnected large-scale system which can relax the restriction on the condition that system functions must be known. Further, the ultimate uniformly bounded (UUB) stability of auxiliary subsystems can be proved according to the Lyapunov principle. Finally, the effectiveness of the proposed control method is verified by presenting the simulation results.}, archive = {J_NEUCOM}, author = {Yuling Liang and Yanhong Luo and Hanguang Su and Xiaoling Zhang and Hongbin Chang and Jun Zhang}, doi = {10.1016/j.neucom.2024.128837}, journal = {Neurocomputing}, month = {1}, pages = {128837}, shortjournal = {Neurocomputing}, title = {Event-triggered explorized IRL-based decentralized fault-tolerant guaranteed cost control for interconnected systems against actuator failures}, volume = {615}, year = {2025}, }
• (2025). Memory-enhanced hierarchical transformer for video paragraph captioning. NEUCOM, 615, 128835. (www)

  @article{ , abstract = {Video paragraph captioning aims to describe a video that contains multiple events with a paragraph of generated coherent sentences. Such a captioning task is full of challenges since the high requirements for visual–textual relevance and semantic coherence across the captioning paragraph of a video. In this work, we introduce a memory-enhanced hierarchical transformer for video paragraph captioning. Our model adopts a hierarchical structure, where the outer layer transformer extracts visual information from a global perspective and captures the relevancy between event segments throughout the entire video, while the inner layer transformer further mines local details within each event segment. By thoroughly exploring both global and local visual information at the video and event levels, our model can provide comprehensive visual feature cues for promising paragraph caption generation. Additionally, we design a memory module to capture similar patterns among event segments within a video, which preserves contextual information across event segments and updates its memory state accordingly. Experimental results on two popular datasets, ActivityNet Captions and YouCook2, demonstrate that our proposed model can achieve superior performance, generating higher quality caption while maintaining consistency in the content of video.}, archive = {J_NEUCOM}, author = {Benhui Zhang and Junyu Gao and Yuan Yuan}, doi = {10.1016/j.neucom.2024.128835}, journal = {Neurocomputing}, month = {1}, pages = {128835}, shortjournal = {Neurocomputing}, title = {Memory-enhanced hierarchical transformer for video paragraph captioning}, volume = {615}, year = {2025}, }
• (2025). On the shortcut learning in multilingual neural machine translation. NEUCOM, 615, 128833. (www)

  @article{ , abstract = {In this study, we revisit the commonly-cited off-target issue in multilingual neural machine translation (MNMT). By carefully designing experiments on different MNMT scenarios and models, we attribute the off-target issue to the overfitting of the shortcuts of (non-centric, centric) language mappings. Specifically, the learned shortcuts biases MNMT to mistakenly translate non-centric languages into the centric language instead of the expected non-centric language for zero-shot translation. Analyses on learning dynamics show that the shortcut learning generally occurs in the later stage of model training, and multilingual pretraining accelerates and aggravates the shortcut learning. Based on these observations, we propose a simple and effective training strategy to eliminate the shortcuts in MNMT models by leveraging the forgetting nature of model training. The only difference from the standard training is that we remove the training instances that may induce the shortcut learning in the later stage of model training. Without introducing any additional data and computational costs, our approach can consistently and significantly improve the zero-shot translation performance by alleviating the shortcut learning for different MNMT models and benchmarks.}, archive = {J_NEUCOM}, author = {Wenxuan Wang and Wenxiang Jiao and Jen-tse Huang and Zhaopeng Tu and Michael Lyu}, doi = {10.1016/j.neucom.2024.128833}, journal = {Neurocomputing}, month = {1}, pages = {128833}, shortjournal = {Neurocomputing}, title = {On the shortcut learning in multilingual neural machine translation}, volume = {615}, year = {2025}, }
• (2025). CSSANet: A channel shuffle slice-aware network for pulmonary nodule detection. NEUCOM, 615, 128827. (www)

  @article{ , abstract = {Lung cancer stands as the leading cause of cancer related mortality worldwide. Precise and automated identification of lung nodules through 3D Computed Tomography (CT) scans is an essential part of screening for lung cancer effectively. Due to the small size of pulmonary nodules and the close correlation between neighboring slices of 3D CT images, most of the existing methods only consider the characteristics of a single slice, thus easily lead to insufficient detection accuracy of pulmonary nodules. To solve this problem, this paper proposes a Channel Shuffle Slice-Aware Network (CSSANet), which aims to fully exploit the spatial correlation between slices and effectively utilize the intra-slice features and inter-slice contextual information to achieve accurate detection of lung nodules. Specifically, we design a Group Shuffle Attention module (GSA module) to fuse the inter-slice feature in order to enhance the discrimination and extraction of corresponding shape information of distinct nodules in the same group of slices. Experiments and ablation study on a publicly available LUNA16 dataset demonstrate that the proposed method can enhance the detection sensitivity effectively. The Competition Performance Metric (CPM) score of 89.8 % is superior over other representative detection models.}, archive = {J_NEUCOM}, author = {Muwei Jian and Huihui Huang and Haoran Zhang and Rui Wang and Xiaoguang Li and Hui Yu}, doi = {10.1016/j.neucom.2024.128827}, journal = {Neurocomputing}, month = {1}, pages = {128827}, shortjournal = {Neurocomputing}, title = {CSSANet: A channel shuffle slice-aware network for pulmonary nodule detection}, volume = {615}, year = {2025}, }
• (2025). Attribute enhanced random walk for community detection in attributed networks. NEUCOM, 615, 128826. (www)

  @article{ , abstract = {Traditional community detection methods often rely solely on network topology, potentially overlooking the influence of attributes. However, community detection in attributed networks presents a unique challenge due to the complex interplay between network topology and node attributes. In this paper, we present AERW, a novel approach that integrates both network topology and node attributes for effective community detection. AERW leverages the principle of homophily by aggregating neighbor attributes within triangles to capture higher-order structural information. Then constructing a bipartite graph to effectively integrate topological and attribute information, ensuring that both aspects are equally represented in the community detection process. Next AERW utilizes hierarchical clustering applied to the probability transition matrix derived from the random walk, enabling the identification of community structures without needing prior knowledge of the number of communities. Extensive experimentation across synthetic and real-world networks underscores AERW’s efficacy, establishing it as a superior approach in community detection within attributed networks.}, archive = {J_NEUCOM}, author = {Zhili Qin and Haoran Chen and Zhongjing Yu and Qinli Yang and Junming Shao}, doi = {10.1016/j.neucom.2024.128826}, journal = {Neurocomputing}, month = {1}, pages = {128826}, shortjournal = {Neurocomputing}, title = {Attribute enhanced random walk for community detection in attributed networks}, volume = {615}, year = {2025}, }
• (2025). L2,1-norm regularized quaternion matrix completion using sparse representation and approximate QSVD. NEUCOM, 615, 128823. (www)

  @article{ , abstract = {Color image completion is a challenging problem in computer vision. Despite the effectiveness of low-rank quaternion matrix completion (LRQMC) methods, their reliance on quaternion singular value decomposition (QSVD) makes them computationally expensive. We propose a novel method based on quaternion Qatar Riyal decomposition (QQR) and quaternion L 2 , 1 -norm, named QLNM-QQR, which reduces computational complexity by avoiding the need for calculating the QSVD of large quaternion matrices. Furthermore, we introduce two improvements: IRQLNM-QQR, which utilizes iteratively reweighted quaternion L 2 , 1 -norm minimization, and QLNM-QQR-SR, which integrates sparse regularization. Additionally, an analysis of the convergence of the proposed algorithms is conducted. Our experiments show that IRQLNM-QQR outperforms QLNM-QQR on both natural and medical color images. The QLNM-QQR-SR method achieves higher PSNR and SSIM values than state-of-the-art methods, with PSNR averaging 4–8 dB higher than non-sparse methods and 0.4 dB higher than the sparse-based TNN-SR method.}, archive = {J_NEUCOM}, author = {Juan Han and Kit Ian Kou and Jifei Miao and Lizhi Liu and Haojiang Li}, doi = {10.1016/j.neucom.2024.128823}, journal = {Neurocomputing}, month = {1}, pages = {128823}, shortjournal = {Neurocomputing}, title = {L2,1-norm regularized quaternion matrix completion using sparse representation and approximate QSVD}, volume = {615}, year = {2025}, }
• (2025). Domain generalization through latent distribution exploration for motor imagery EEG classification. NEUCOM, 614, 128889. (www)

  @article{ , abstract = {Electroencephalography (EEG)-based Motor Imagery (MI) brain-computer interface (BCI) systems play essential roles in motor function rehabilitation for patients with post-stroke. Existing neural networks for decoding MI EEG face challenges due to nonstationary characteristics and subject-specific variations of EEG data. To address these challenges and improve generalization performance, this study proposes a domain generalization (DG) model that eliminates the need for user-specific calibration in real-life applications. Specifically, the proposed model comprises two branches: the first branch applies several independent decision-making networks to decode and classify subjects’ motor intentions, while the second branch adaptively assigns weights to classification results and fuses them into a comprehensive decision. Both branches utilize EEGNet and ShallowConvNet to extract time-frequency-spatial features. By implementing multiple classification networks, the model can learn a broad range of data distributions from source subjects, which contributes to improved generalization performance on target subjects. The proposed EEG-DG framework was evaluated on BCI Competition IV Dataset 2a, 2b and PhysioNet. Results show that the proposed framework significantly enhances the classification performance of MI EEG, outperforming several state-of-the-art models on all three datasets, underlining its superior efficacy in real-world scenarios and exceptional generalization performance. The source code can be accessed at https://github.com/DrugLover/Multibranch-DG-EEG .}, archive = {J_NEUCOM}, author = {Hao Song and Qingshan She and Feng Fang and Su Liu and Yun Chen and Yingchun Zhang}, doi = {10.1016/j.neucom.2024.128889}, journal = {Neurocomputing}, month = {1}, pages = {128889}, shortjournal = {Neurocomputing}, title = {Domain generalization through latent distribution exploration for motor imagery EEG classification}, volume = {614}, year = {2025}, }
• (2025). Multi-sensor information fusion in internet of vehicles based on deep learning: A review. NEUCOM, 614, 128886. (www)

  @article{ , abstract = {Environmental perception is a crucial component of intelligent driving technology, providing the informational foundation for intelligent decision-making and collaborative control. Due to the limitations of single sensors and the continuous advancements in deep learning and sensor technologies, multi-sensor information fusion in the Internet of Vehicles (IoV) has emerged as a major research hotspot. This approach is also a primary solution for achieving full self-driving. However, given the complexity of the technology, there are still many challenges in achieving accurate and reliable real-time multi-source information perception. Current discussions often focus on specific aspects of multi-sensor fusion in intelligent driving, while detailed discussions on sensor fusion in the context of the IoV are relatively scarce. To provide a comprehensive discussion and analysis of multi-sensor information fusion in IoV, this paper first provides a detailed introduction to its developmental background and the commonly involved sensors. Subsequently, a detailed analysis of the strategies, deep learning architectures, and methods for multi-sensor information fusion in the IoV is presented. Finally, the specific applications and key issues related to multi-sensor information fusion in IoV are discussed from multiple perspectives, along with an analysis of future development trends. This paper aims to serve as a valuable reference for advancing multi-sensor information fusion technology in IoV environments and supporting the realization of full self-driving.}, archive = {J_NEUCOM}, author = {Di Tian and Jiabo Li and Jingyuan Lei}, doi = {10.1016/j.neucom.2024.128886}, journal = {Neurocomputing}, month = {1}, pages = {128886}, shortjournal = {Neurocomputing}, title = {Multi-sensor information fusion in internet of vehicles based on deep learning: A review}, volume = {614}, year = {2025}, }
• (2025). HyV-summ: Social media video summarization on custom dataset using hybrid techniques. NEUCOM, 614, 128852. (www)

  @article{ , abstract = {The proliferation of social networking platforms such as YouTube, Facebook, Instagram, and X has led to an exponential growth in multimedia content, with billions of videos uploaded every hour. Efficient management of such vast amount of data necessitates advanced summarization techniques in order to eliminate irrelevant and redundant information. A summarized video, containing the most distinct frames or key frames, provides a concise representation of the original content. Existing deep learning and non-deep learning techniques for video summarization have certain limitations. Deep learning methods are complex and resource-intensive, while non-deep learning algorithms often fail to extract informative features from vast social media videos. This paper addresses the issue by proposing a novel hybrid technique, named Hybrid Video Summarization ( HyV-Summ ), which integrates deep and non-deep learning techniques to leverage their respective strengths by focusing only on social media content. We developed a custom dataset, SocialSum to train our proposed model HyV-Summ , since existing benchmark datasets like TVSum and SumMe contain diverse types of content not specific to social media videos. We provide a comparative analysis of existing techniques and datasets with our proposed techniques and dataset. The results demonstrate that HyV-Summ outperforms existing techniques, such as Long Short Term Memory (LSTM)-based and Generative Adversarial Network (GAN)-based summarization by achieving higher F1-scores while applied on both the SocialSum dataset and available datasets.}, archive = {J_NEUCOM}, author = {Jayanta Paul and Anuska Roy and Abhijit Mitra and Jaya Sil}, doi = {10.1016/j.neucom.2024.128852}, journal = {Neurocomputing}, month = {1}, pages = {128852}, shortjournal = {Neurocomputing}, title = {HyV-summ: Social media video summarization on custom dataset using hybrid techniques}, volume = {614}, year = {2025}, }
• (2025). S2CA: Shared concept prototypes and concept-level alignment for text–video retrieval. NEUCOM, 614, 128851. (www)

  @article{ , abstract = {Text–video retrieval, as a fundamental task of cross-modal learning, relies on effectively establishing the semantic association between text and video. At present, mainstream semantic alignment methods for text–video adopt instance-level alignment strategies, ignoring the fine-grained concept association and the “concept-level alignment” characteristics of text–video. In this regard, we propose S hared C oncept Prototypes and C oncept-level A lignment ( S2CA ) to achieve concept-level alignment. Specifically, we utilize the text–video Shared Concept Prototypes mechanism to bridge the correspondence between text and video. On this basis, we use cross-attention and Gumbel-softmax to obtain Discrete Concept Allocation Matrices and then assign text and video tokens to corresponding concept prototypes. In this way, texts and videos are decoupled into multiple Conceptual Aggregated Features , thereby achieving Concept-level Alignment . In addition, we use CLIP as the teacher model and adopt the Align-Transform-Reconstruct distillation framework to strengthen the multimodal semantic learning ability. The extensive experiments on MSR-VTT, DiDeMo, ActivityNet and MSVD prove the effectiveness of our method.}, archive = {J_NEUCOM}, author = {Yuxiao Li and Yu Xin and Jiangbo Qian and Yihong Dong}, doi = {10.1016/j.neucom.2024.128851}, journal = {Neurocomputing}, month = {1}, pages = {128851}, shortjournal = {Neurocomputing}, title = {S2CA: Shared concept prototypes and concept-level alignment for text–video retrieval}, volume = {614}, year = {2025}, }
• (2025). Graph-enhanced visual representations and question-guided dual attention for visual question answering. NEUCOM, 614, 128850. (www)

  @article{ , abstract = {Visual Question Answering (VQA) has witnessed significant advancements recently, due to the application of deep learning in the field of vision-language research. Most current VQA models focus on merging visual and text features, but it is essential for these models to also consider the relationships between different parts of an image and use question information to highlight important features. This study proposes a method to enhance neighboring image region features and learn question-aware visual representations. First, we construct a region graph to represent spatial relationships between objects in the image. Then, graph convolutional network (GCN) is used to propagate information across neighboring regions, enriching each region’s feature representation by integrating contextual information. To capture long-range dependencies, the graph is enhanced with random walk with restart (RWR), enabling multi-hop reasoning across distant regions. Furthermore, a question-aware dual attention mechanism is introduced to further refine region features at both region and feature levels, ensuring that the model emphasizes key regions that are critical for answering the question. The enhanced region representations are then combined with the encoded question to predict an answer. Through extensive experiments on VQA benchmarks, the study demonstrates state-of-the-art performance by leveraging regional dependencies and question guidance. The integration of GCNs and random walks in the graph helps capture contextual information to focus visual attention selectively, resulting in significant improvements over existing methods on VQA 1.0 and VQA 2.0 benchmark datasets.}, archive = {J_NEUCOM}, author = {Abdulganiyu Abdu Yusuf and Chong Feng and Xianling Mao and Yunusa Haruna and Xinyan Li and Ramadhani Ally Duma}, doi = {10.1016/j.neucom.2024.128850}, journal = {Neurocomputing}, month = {1}, pages = {128850}, shortjournal = {Neurocomputing}, title = {Graph-enhanced visual representations and question-guided dual attention for visual question answering}, volume = {614}, year = {2025}, }
• (2025). Adversarial contrastive representation training with external knowledge injection for zero-shot stance detection. NEUCOM, 614, 128849. (www)

  @article{ , abstract = {Zero-shot stance detection (ZSSD) is a task that involves identifying the author’s perspective on specific issues in text, particularly when the target topic has not been encountered during the model training process, to address rapidly evolving topics on social media. This paper introduces a ZSSD framework named KEL-CA. To enable the model to more effectively utilize transferable stance features for representing unseen targets, the framework incorporates a multi-layer contrastive learning and adversarial domain transfer module. Unlike traditional contrastive or adversarial learning, our framework captures both correlations and distinctions between invariant and specific features, as well as between different stance labels, and enhances the generalization ability and robustness of the features. Subsequently, to address the problem of insufficient information about the target context, we designed a dual external knowledge injection module that uses a large language model (LLM) to extract external knowledge from a Wikipedia-based local knowledge base and a Chain-of-Thought (COT) process to ensure the timeliness and relevance of the knowledge to infer the stances of unseen targets. Experimental results demonstrate that our approach outperforms existing models on two benchmark datasets, thereby validating its efficacy in ZSSD tasks.}, archive = {J_NEUCOM}, author = {Yifan Ding and Ying Lei and Anqi Wang and Xiangrun Liu and Tuanfei Zhu and Yizhou Li}, doi = {10.1016/j.neucom.2024.128849}, journal = {Neurocomputing}, month = {1}, pages = {128849}, shortjournal = {Neurocomputing}, title = {Adversarial contrastive representation training with external knowledge injection for zero-shot stance detection}, volume = {614}, year = {2025}, }
• (2025). An analysis of pre-trained stable diffusion models through a semantic lens. NEUCOM, 614, 128846. (www)

  @article{ , abstract = {Recently, generative models for images have garnered remarkable attention, due to their effective generalization ability and their capability to generate highly detailed and realistic content. Indeed, the success of generative networks ( e.g. , BigGAN, StyleGAN, Diffusion Models) has driven researchers to develop increasingly powerful models. As a result, we have observed an unprecedented improvement in terms of both image resolution and realism, making generated images indistinguishable from real ones. In this work, we focus on a family of generative models known as Stable Diffusion Models (SDMs), which have recently emerged due to their ability to generate images in a multimodal setup ( i.e. , from a textual prompt) and have outperformed adversarial networks by learning to reverse a diffusion process. Given the complexity of these models that makes it hard to retrain them, researchers started to exploit pre-trained SDMs to perform downstream tasks ( e.g. , classification and segmentation), where semantics plays a fundamental role. In this context, understanding how well the model preserves semantic information may be crucial to improve its performance. This paper presents an approach aimed at providing insights into the properties of a pre-trained SDM through the semantic lens. In particular, we analyze the features extracted by the U-Net within a SDM to explore whether and how the semantic information of an image is preserved in its internal representation. For this purpose, different distance measures are compared, and an ablation study is performed to select the layer (or combination of layers) of the U-Net that best preserves the semantic information. We also seek to understand whether semantics are preserved when the image undergoes simple transformations ( e.g. , rotation, flip, scale, padding, crop, and shift) and for a different number of diffusion denoising steps. To evaluate these properties, we consider popular benchmarks for semantic segmentation tasks ( e.g. , COCO, and Pascal-VOC). Our experiments suggest that the first encoder layer at 16 × 16 resolution effectively preserves semantic information. However, increasing inference steps (even for a minimal amount of noise) and applying various image transformations can affect the diffusion U-Net’s internal feature representation. Additionally, we propose some examples taken from a video benchmark (DAVIS dataset), where we investigate if an object instance within a video preserves its internal representation even after several frames. Our findings suggest that the internal object representation remains consistent across multiple frames in a video, as long as the configuration changes are not excessive.}, archive = {J_NEUCOM}, author = {Simone Bonechi and Paolo Andreini and Barbara Toniella Corradini and Franco Scarselli}, doi = {10.1016/j.neucom.2024.128846}, journal = {Neurocomputing}, month = {1}, pages = {128846}, shortjournal = {Neurocomputing}, title = {An analysis of pre-trained stable diffusion models through a semantic lens}, volume = {614}, year = {2025}, }
• (2025). Revising similarity relationship hashing for unsupervised cross-modal retrieval. NEUCOM, 614, 128844. (www)

  @article{ , abstract = {Previous methods have made promising progress, but there are still some limitations in narrowing the gap between modalities and exploring and preserving intrinsic multimodal semantics. Furthermore, there has been a failure to effectively incorporate the hash codes to correct poorly trained instance pairs during the training process. To overcome the above-mentioned issues, we propose a novel unsupervised hash learning framework, Revising Similarity Relationship Hashing (RSRH). Firstly, we constructed a feature cross-reconstruction module to narrow the gap between modalities. In addition, we build a multimodal fusion similarity map that nonlinearly combines intra- and inter-modal similarity maps to generate multimodal representations with complementary relationships. Finally, we propose a multimodal fusion graph update module for updating poorly trained instance pairs, improving retrieval performance. Experimental data show that our method outperforms many current mainstream hashing methods in performance, and its effectiveness and superiority have been fully validated.}, archive = {J_NEUCOM}, author = {You Wu and Bo Li and Zhixin Li}, doi = {10.1016/j.neucom.2024.128844}, journal = {Neurocomputing}, month = {1}, pages = {128844}, shortjournal = {Neurocomputing}, title = {Revising similarity relationship hashing for unsupervised cross-modal retrieval}, volume = {614}, year = {2025}, }
• (2025). Minimum control of cluster synchronization effort in diffusion coupled nonlinear networks. NEUCOM, 614, 128841. (www)

  @article{ , abstract = {This paper studies the design of minimal control effort for cluster synchronization (CS) in a diffusion-coupled nonlinear network under directed graph. Under the conditions that the directed graph satisfies the cluster input equivalence condition and the system possesses a bounded Jacobian matrix, we obtain CS of diffusion-coupled nonlinear network with non-diagonal coupling matrix. Based on matrix measure and balancing theorem, we obtain the local minimization controllers for the minimal control effort of CS. Finally, the theoretical results are validated through a numerical example involving a network of coupled FitzHugh–Nagumo neurons with a general topology of interactions.}, archive = {J_NEUCOM}, author = {Jinkui Zhang and Shidong Zhai and Wei Zhu}, doi = {10.1016/j.neucom.2024.128841}, journal = {Neurocomputing}, month = {1}, pages = {128841}, shortjournal = {Neurocomputing}, title = {Minimum control of cluster synchronization effort in diffusion coupled nonlinear networks}, volume = {614}, year = {2025}, }
• (2025). Interpretable neural network classification model using first-order logic rules. NEUCOM, 614, 128840. (www)

  @article{ , abstract = {Over the past decade, the field of neural networks has made significant strides, particularly in deep learning. However, their limited interpretability has constrained their application in certain critical domains, drawing widespread criticism. Researchers have proposed various methods for explaining neural networks to address this challenge. This paper focuses on rule-based explanations for neural network classification problems. We propose IRCnet, a scalable classification model based on first-order logic rules. IRCnet consists of layers for learning conjunction and disjunction rules, utilizing binary logic activation functions to enhance interpretability. The model is initially trained using a continuous-weight version, which is later binarized to produce a discrete-weight version. During training, we innovatively employed gradient approximation method to handle the non-differentiable weight binarization function, thereby enabling the training of split matrices used for binarization. Finally, Conjunctive Normal Form (CNF) or Disjunctive Normal Form (DNF) rules are extracted from the model’s discrete-weight version. Experimental results indicate that our model achieves the highest or near-highest performance across various classification metrics in multiple structured datasets while demonstrating significant scalability. It effectively balances classification accuracy with the complexity of the generated rules.}, archive = {J_NEUCOM}, author = {Haiming Tuo and Zuqiang Meng and Zihao Shi and Daosheng Zhang}, doi = {10.1016/j.neucom.2024.128840}, journal = {Neurocomputing}, month = {1}, pages = {128840}, shortjournal = {Neurocomputing}, title = {Interpretable neural network classification model using first-order logic rules}, volume = {614}, year = {2025}, }
• (2025). Deep neural networks for knowledge-enhanced molecular modeling. NEUCOM, 614, 128838. (www)

  @article{ , abstract = {Designing neural networks for molecular modeling is a crucial task in the field of artificial intelligence. The goal is to utilize neural networks to understand and design molecules, which has significant implications for drug development and other real-world applications. Recently, with the advancement of deep learning, molecular modeling has made considerable progress. However, current methods are primarily data-driven, overlooking the role of domain knowledge, such as molecular shapes, in the modeling process. In this paper, we systematically investigate how incorporating molecular shape knowledge can enhance molecular modeling. Specifically, we design two deep neural networks, ShapePred and ShapeGen , to utilize molecular shapes in molecule prediction and generation. Experimental results demonstrate that integrating shape knowledge can significantly improve model performance. Notably, ShapePred exhibits strong performance across 11 molecule prediction datasets, while ShapeGen can more efficiently generate high-quality drug molecules based on given target proteins.}, archive = {J_NEUCOM}, author = {Siyu Long and Jianyu Wu and Yi Zhou and Fan Sha and Xinyu Dai}, doi = {10.1016/j.neucom.2024.128838}, journal = {Neurocomputing}, month = {1}, pages = {128838}, shortjournal = {Neurocomputing}, title = {Deep neural networks for knowledge-enhanced molecular modeling}, volume = {614}, year = {2025}, }
• (2025). Improved exploration–exploitation trade-off through adaptive prioritized experience replay. NEUCOM, 614, 128836. (www)

  @article{ , abstract = {Experience replay is an indispensable part of deep reinforcement learning algorithms that enables the agent to revisit and reuse its past and recent experiences to update the network parameters. In many baseline off-policy algorithms, such as deep Q-networks (DQN), transitions in the replay buffer are typically sampled uniformly. This uniform sampling is not optimal for accelerating the agent’s training towards learning the optimal policy. A more selective and prioritized approach to experience sampling can yield improved learning efficiency and performance. In this regard, this work is devoted to the design of a novel prioritizing strategy to adaptively adjust the sampling probabilities of stored transitions in the replay buffer. Unlike existing sampling methods, the proposed algorithm takes into consideration the exploration–exploitation trade-off (EET) to rank transitions, which is of utmost importance in learning an optimal policy. Specifically, this approach utilizes temporal difference and Bellman errors as criteria for sampling priorities. To maintain balance in EET throughout training, the weights associated with both criteria are dynamically adjusted when constructing the sampling priorities. Additionally, any bias introduced by this sample prioritization is mitigated through assigning importance-sampling weight to each transition in the buffer. The efficacy of this prioritization scheme is assessed through training the DQN algorithm across various OpenAI Gym environments. The results obtained underscore the significance and superiority of our proposed algorithm over state-of-the-art methods. This is evidenced by its accelerated learning pace, greater cumulative reward, and higher success rate.}, archive = {J_NEUCOM}, author = {Hossein Hassani and Soodeh Nikan and Abdallah Shami}, doi = {10.1016/j.neucom.2024.128836}, journal = {Neurocomputing}, month = {1}, pages = {128836}, shortjournal = {Neurocomputing}, title = {Improved exploration–exploitation trade-off through adaptive prioritized experience replay}, volume = {614}, year = {2025}, }
• (2025). Invisible and robust watermarking model based on hierarchical residual fusion multi-scale convolution. NEUCOM, 614, 128834. (www)

  @article{ , abstract = {In current deep learning based watermarking technologies, it remains challenging to fully integrate the features of watermark and cover image. Most watermarking models with fixed-size kernel convolution exhibit restricted feature extraction ability, leading to incomplete feature fusion. To address this issue, a hierarchical residual fusion multi-scale convolution (HRFMS) module is designed. The method extracts image features from various receptive fields and implements feature interaction by residual connection. To produce watermarked image with high visual quality and attack resistance, a watermarking model based on the HRFMS is devised to achieve multi-scale feature fusion. Moreover, to minimize image distortion caused by watermark information, an attention mask layer is designed to guide the distribution of watermark information. The experimental results demonstrate that the invisibility and the robustness of the HRFMSNet are excellent. The watermarked images generated by the HRFMSNet are nearly visually indistinguishable from the cover images. The average peak signal-to-noise ratio of the watermarked images is 37.13 dB, and most of the bit error rates of the decoded messages are below 0.02.}, archive = {J_NEUCOM}, author = {Jun-Zhuo Zou and Ming-Xuan Chen and Li-Hua Gong}, doi = {10.1016/j.neucom.2024.128834}, journal = {Neurocomputing}, month = {1}, pages = {128834}, shortjournal = {Neurocomputing}, title = {Invisible and robust watermarking model based on hierarchical residual fusion multi-scale convolution}, volume = {614}, year = {2025}, }
• (2025). Mitigating gradient conflicts via expert squads in multi-task learning. NEUCOM, 614, 128832. (www)

  @article{ , abstract = {The foundation of multi-task learning lies in the collaboration and interaction among tasks. However, in numerous real-world scenarios, certain tasks usually necessitate distinct, specialized knowledge. The mixing of these different task-specific knowledge often results in gradient conflicts during the optimization process, posing a significant challenge in the design of effective multi-task learning systems. This study proposes a straightforward yet effective multi-task learning framework that employs groups of expert networks to decouple the learning of task-specific knowledge and mitigate such gradient conflicts. Specifically, this approach partitions the feature channels into task-specific and shared components. The task-specific subsets are processed by dedicated experts to distill specialized knowledge. The shared features are captured by a point-wise aggregation layer from the whole outputs of all experts, demonstrating superior performance in capturing inter-task interactions. By considering both task-specific knowledge and shared features, the proposed approach exhibits superior performance in multi-task learning. Extensive experiments conducted on the PASCAL-Context and NYUD-v2 datasets have demonstrated the superiority of the proposed approach compared to other state-of-the-art methods. Furthermore, a benchmark dataset for multi-task learning in underwater scenarios has been developed, encompassing object detection and underwater image enhancement tasks. Comprehensive experiments on this dataset consistently validate the effectiveness of the proposed multi-task learning strategy. The source code is available at https://github.com/chenjie04/Multi-Task-Learning-PyTorch .}, archive = {J_NEUCOM}, author = {Jie Chen and Meng Joo Er}, doi = {10.1016/j.neucom.2024.128832}, journal = {Neurocomputing}, month = {1}, pages = {128832}, shortjournal = {Neurocomputing}, title = {Mitigating gradient conflicts via expert squads in multi-task learning}, volume = {614}, year = {2025}, }
• (2025). Multi-modal soft prompt-tuning for chinese clickbait detection. NEUCOM, 614, 128829. (www)

  @article{ , abstract = {With the rapid growth of Chinese online services, clickbait has proliferated at an unprecedented rate, designed to manipulate users into clicking for increased traffic or advertising promotion. Such clickbait not only facilitates the spread of fake news and misinformation but also enables click-jacking attacks, redirecting users to deceptive websites that steal personal information. These harmful activities can result in significant losses and serious repercussions. The widespread presence of clickbait underscores both the importance and the challenges of developing effective detection methods. To date, the research paradigm of clickbait detection evolved from deep neural networks to fine-tuned Pre-trained Language Models (PLMs) and, more recently, into prompt-tuning models. However, these methods may suffer two main limitations: (1) they fail to utilize the multi-modal context information in news or posts and explore the higher-level feature representations to enhance the performance of clickbait detection; (2) they largely ignore the diverse range of Chinese expressive forms and neglect the complex semantics and syntactic structures of textual content to assist in learning a better news representation. To overcome these limitations, we proposed a Multi-modal Soft Prompt-tuning Method (MSP) for Chinese Clickbait Detection, which jointly models the textual and image information into a continuous prompt embedding as the input of PLMs. Specifically, firstly, the soft prompt-tuning model including Graph Attention Network and Contrastive Language-Image Pre-training are employed to learn the feature representations of texts and images in news or posts, respectively. Then the obtained text and image representations are re-input into the soft prompt-tuning model with automatic template generation. The extensive experiments on three Chinese clickbait detection datasets demonstrate that our MSP achieved state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Ye Wang and Yi Zhu and Yun Li and Liting Wei and Yunhao Yuan and Jipeng Qiang}, doi = {10.1016/j.neucom.2024.128829}, journal = {Neurocomputing}, month = {1}, pages = {128829}, shortjournal = {Neurocomputing}, title = {Multi-modal soft prompt-tuning for chinese clickbait detection}, volume = {614}, year = {2025}, }
• (2025). Improving semi-autoregressive machine translation with the guidance of syntactic dependency parsing structure. NEUCOM, 614, 128828. (www)

  @article{ , abstract = {The advent of non-autoregressive machine translation (NAT) accelerates the decoding superior to autoregressive machine translation (AT) significantly, while bringing about a performance decrease. Semi-autoregressive neural machine translation (SAT), as a compromise, enjoys the merits of both autoregressive and non-autoregressive decoding. However, current SAT methods face the challenges of information-limited initialization and rigorous termination. This paper develops a layer-and-length-based syntactic labeling method and introduces a syntactic dependency parsing structure-guided two-stage semi-autoregressive translation (SDPSAT) structure, which addresses the above challenges with a syntax-based initialization and termination. Additionally, we also present a Mixed Training strategy to shrink exposure bias. Experiments on seven widely-used datasets reveal that our SDPSAT surpasses traditional SAT models with reduced word repetition and achieves competitive results with the AT baseline at a 2 × ∼ 3 × speedup.}, archive = {J_NEUCOM}, author = {Xinran Chen and Sufeng Duan and Gongshen Liu}, doi = {10.1016/j.neucom.2024.128828}, journal = {Neurocomputing}, month = {1}, pages = {128828}, shortjournal = {Neurocomputing}, title = {Improving semi-autoregressive machine translation with the guidance of syntactic dependency parsing structure}, volume = {614}, year = {2025}, }
• (2025). Long–Short observation-driven prediction network for pedestrian crossing intention prediction with momentary observation. NEUCOM, 614, 128824. (www)

  @article{ , abstract = {Pedestrian crossing intention prediction aims to predict whether the pedestrian will cross the road, which is crucial for the decision-making of intelligent vehicles and ensuring traffic safety. Existing methods just rely on long-term observation and rarely consider it challenging to obtain sufficiently long and precise observation in real-world scenarios. Focus on momentary observation, which only contains two frames of the preceding and current time, we propose a novel Long–Short Observation-driven Prediction Network (LSOP-Net). LSOP-Net comprises two critical components, the Momentary Observation feature Extraction Module (MOE-Module) and the Multimodal Long–Short-term feature Fusion Module (MLSFusion). Utilizing a hybrid training strategy and an external long-term feature pool, the MOE-Module is proposed to extract features with long-term patterns from momentary observations, which effectively mitigates feature deficiency arising from momentary observations. Based on a feature selection fusion mechanism, the MLSFusion is proposed to explicitly model the importance relationship between various modalities’ long–short-term features and the output, which adaptively fuses the long–short-term features from various modalities. Experimental results on the JAAD and PIE datasets demonstrate that our approach achieves superior performance in pedestrian crossing intention prediction with momentary observation.}, archive = {J_NEUCOM}, author = {Hui Liu and Chunsheng Liu and Faliang Chang and Yansha Lu and Minhang Liu}, doi = {10.1016/j.neucom.2024.128824}, journal = {Neurocomputing}, month = {1}, pages = {128824}, shortjournal = {Neurocomputing}, title = {Long–Short observation-driven prediction network for pedestrian crossing intention prediction with momentary observation}, volume = {614}, year = {2025}, }
• (2025). A transformer-based network with adaptive spatial prior for visual tracking. NEUCOM, 614, 128821. (www)

  @article{ , abstract = {Single object tracking (SOT) in complex scenes presents significant challenges in computer vision. In recent years, transformer has shown its demonstrated efficacy in visual object tracking tasks, due to its capacity to capture the long-range dependencies between image pixels. However, two limitations hinder the performance improvement of transformer-based trackers. Firstly, transformer splits and partitions the image into a sequence of patches, which disrupts the internal structural information of the object. Secondly, transformer-based trackers encode the target template and search region together, potentially leading to confusion between the target and background during feature interaction. To address the above issues, we propose a fully transformer-based tracking framework via learning structural prior information, called SPformer. In other words, a self-attention spatial-prior generative network is established for simulating the spatial associations between features. Moreover, the cross-attention structural prior extractors based on Gaussian and arbitrary distributions are developed to seek the semantic interaction features between the object template and the search region, effectively mitigating feature confusion. Extensive experiments on eight prevailing benchmarks demonstrate that SPformer outperforms existing state-of-art (SOAT) trackers. We further analyze the effectiveness of the two proposed prior modules and validate their application in target tracking models.}, archive = {J_NEUCOM}, author = {Feng Cheng and Gaoliang Peng and Junbao Li and Benqi Zhao and Jeng-Shyang Pan and Hang Li}, doi = {10.1016/j.neucom.2024.128821}, journal = {Neurocomputing}, month = {1}, pages = {128821}, shortjournal = {Neurocomputing}, title = {A transformer-based network with adaptive spatial prior for visual tracking}, volume = {614}, year = {2025}, }
• (2025). Reinforcement learning method based on sample regularization and adaptive learning rate for AGV path planning. NEUCOM, 614, 128820. (www)

  @article{ , abstract = {This paper proposes the proximal policy optimization (PPO) method based on sample regularization (SR) and adaptive learning rate (ALR) to address the issues of limited exploration ability and slow convergence speed in Autonomous Guided Vehicle (AGV) path planning using reinforcement learning algorithms in dynamic environments. Firstly, the regularization term based on empirical samples is designed to solve the bias and imbalance issues of training samples, and the sample regularization is added to the objective function to improve the policy selectivity of the PPO algorithm, thereby increasing the AGV’s exploration ability during the training process in the working environment. Secondly, the Fisher information matrix of the Kullback-Leibler (KL) divergence approximation and the KL divergence constraint term are exploited to design the policy update mechanism based on the dynamically adjustable adaptive learning rate throughout training. The method considers the geometric structure of the parameter space and the change of the policy gradient, aiming to optimize parameter update direction and enhance convergence speed and stability of the algorithm. Finally, the AGV path planning scheme based on reinforcement learning is established for simulation verification and comparations in two-dimensional raster map and Gazebo 3D simulation environment. Simulation results verify the feasibility and superiority of the proposed method applied to the AGV path planning problem.}, archive = {J_NEUCOM}, author = {Jun Nie and Guihua Zhang and Xiao Lu and Haixia Wang and Chunyang Sheng and Lijie Sun}, doi = {10.1016/j.neucom.2024.128820}, journal = {Neurocomputing}, month = {1}, pages = {128820}, shortjournal = {Neurocomputing}, title = {Reinforcement learning method based on sample regularization and adaptive learning rate for AGV path planning}, volume = {614}, year = {2025}, }
• (2025). Multi-attribute dynamic attenuation learning improved spiking actor network. NEUCOM, 614, 128819. (www)

  @article{ , abstract = {Deep reinforcement learning (DRL) has shown promising results in solving robotic control and decision tasks, which can learn the high-dimensional state and action information well. Despite their successes, conventional neural-based DRL models are criticized for low energy efficiency, making them laborious to be widely applied in low-power electronics. With more biologically plausible plasticity principles, spiking neural networks (SNNs) are now considered an energy-efficient and robust alternative. The most existing dynamics and learning paradigms for spiking neurons with a common Leaky Integrate-and-Fire (LIF) neuron model often result in relatively low efficiency and poor robustness. To address these limitations, we propose a multi-attribute dynamic attenuation learning improved spiking actor network (MADA-SAN) for reinforcement learning to achieve effective decision-making. The resistance, membrane voltage and membrane current of spiking neurons are updated from a fixed value into dynamic attenuation. By enhancing the temporal relation dependencies in neurons, this model can learn the spatio-temporal relevance of complex continuous information well. Extensive experimental results show MADA-SAN performs better than its counterpart deep actor network on six continuous control tasks from OpenAI gym. Besides, we further validated the proposed MADA-LIF can achieve comparable performance with other state-of-the-art algorithms on MNIST and DVS-gesture recognition tasks.}, archive = {J_NEUCOM}, author = {Rong Xiao and Zhiyuan Hu and Jie Zhang and Chenwei Tang and Jiancheng Lv}, doi = {10.1016/j.neucom.2024.128819}, journal = {Neurocomputing}, month = {1}, pages = {128819}, shortjournal = {Neurocomputing}, title = {Multi-attribute dynamic attenuation learning improved spiking actor network}, volume = {614}, year = {2025}, }
• (2025). Fixed-time event-triggered pinning synchronization of complex network via aperiodically intermittent control. NEUCOM, 614, 128818. (www)

  @article{ , abstract = {This paper studies the fixed-time event-triggered pinning synchronization problem for complex network using aperiodic intermittent control. A novel fixed-time stability lemma with aperiodic intermittent characteristic is first proposed. By designing appropriate event-triggered aperiodic intermittent pinning controller (ETAIPC) based on the average control rate, several conditions are derived to ensure the fixed-time synchronization. The upper bound of setting-time is independent of any initial values and only concerns with design parameters, network size and node dimension. A simple to execute selection algorithm is adopted to renew the pinning node set. The Zeno behavior is also excluded through a rigorous theoretical analysis. Simulation examples are employed to demonstrate the efficacy of the obtained method.}, archive = {J_NEUCOM}, author = {Junru Zhang and Jian-An Wang and Jie Zhang and Mingjie Li and Zhicheng Zhao and Xinyu Wen}, doi = {10.1016/j.neucom.2024.128818}, journal = {Neurocomputing}, month = {1}, pages = {128818}, shortjournal = {Neurocomputing}, title = {Fixed-time event-triggered pinning synchronization of complex network via aperiodically intermittent control}, volume = {614}, year = {2025}, }
• (2025). Secure bipartite consensus of leader–follower multi-agent systems under denial-of-service attacks via observer-based dynamic event-triggered control. NEUCOM, 614, 128817. (www)

  @article{ , abstract = {This paper investigates secure bipartite consensus (SBC) control problems of leader–follower multi-agent systems (MASs) subjected to denial-of-service (DoS) attacks via observer-based dynamic event-triggered control (DETC). An observer-based DETC protocol with two combining measurements (follower–follower and follower–leader) is first proposed corresponding to valid DoS attack intervals and valid safe intervals. It is concluded that the SBC of MASs without input saturation can be achieved via the observer-based DETC protocol under a signed directed graph if some sufficient inequality conditions hold. And the Zeno behavior can be excluded. Then, a resembling result corresponding to a signed undirected graph is obtained. Furthermore, by using low-gain feedback technic, the semi-global SBC issue under a saturated controller is also considered. Finally, a numerical simulation and two application simulations are displayed to illustrate the effectiveness of the proposed control protocol.}, archive = {J_NEUCOM}, author = {Haichuan Xu and Fanglai Zhu and Xufeng Ling}, doi = {10.1016/j.neucom.2024.128817}, journal = {Neurocomputing}, month = {1}, pages = {128817}, shortjournal = {Neurocomputing}, title = {Secure bipartite consensus of leader–follower multi-agent systems under denial-of-service attacks via observer-based dynamic event-triggered control}, volume = {614}, year = {2025}, }
• (2025). Quantifying neural network uncertainty under volatility clustering. NEUCOM, 614, 128816. (www)

  @article{ , abstract = {Time-series with volatility clustering pose a unique challenge to uncertainty quantification (UQ) for returns forecasts. Methods for UQ such as Deep Evidential regression offer a simple way of quantifying return forecast uncertainty without the costs of a full Bayesian treatment. However, the Normal-Inverse-Gamma (NIG) prior adopted by Deep Evidential regression is prone to miscalibration as the NIG prior is assigned to latent mean and variance parameters in a hierarchical structure. Moreover, it also overparameterizes the marginal data distribution. These limitations may affect the accurate delineation of epistemic (model) and aleatoric (data) uncertainties. We propose a Scale Mixture Distribution as a simpler alternative which can provide favourable complexity-accuracy trade-off and assign separate subnetworks to each model parameter. To illustrate the performance of our proposed method, we apply it to two sets of financial time-series exhibiting volatility clustering: cryptocurrencies and U.S. equities and test the performance in some ablation studies.}, archive = {J_NEUCOM}, author = {Steven Y.K. Wong and Jennifer S.K. Chan and Lamiae Azizi}, doi = {10.1016/j.neucom.2024.128816}, journal = {Neurocomputing}, month = {1}, pages = {128816}, shortjournal = {Neurocomputing}, title = {Quantifying neural network uncertainty under volatility clustering}, volume = {614}, year = {2025}, }
• (2025). Unveiling the potential of graph coloring in feature set partitioning: A study on high-dimensional datasets. NEUCOM, 614, 128814. (www)

  @article{ , abstract = {The branch of machine learning known as multi-view ensemble learning (MEL) is young and evolving quickly. The learning procedure in this case makes use of subsets of different features from the same dataset, and the prediction produced is then combined. The vertical partition of the dataset in regard to the portion of the feature set in a single source dataset is referred to as the view. View construction is a crucial job in MEL because an adequate number of good-quality views improves MEL’s performance. A well-known method of dividing up the nodes of a graph is called “graph coloring”, which involves giving each vertex a unique color so that no two neighboring vertex pairs share the same color. This approach can be utilized in a number of diverse fields including clustering. In this study, high-dimension features are partitioned using graph coloring, which is used to perform heterogeneous feature grouping. In order to automatically create views in MEL over high-dimensional datasets, the Graph coloring-based feature set partitioning ( GC -FSP) technique is used. A support vector machine and artificial neural network have been used with 15 high-dimensional data sets to demonstrate the efficacy of the GC -FSP based MEL framework. Compared to single-view learning and other cutting-edge FSP-based MEL techniques, the results show that it is successful in enhancing classification performance. The outcomes have undergone non-parametric statistical study and the intended MEL framework has produced improved classification accuracy that is both acceptable and accurate.}, archive = {J_NEUCOM}, author = {Aditya Kumar and Jainath Yadav}, doi = {10.1016/j.neucom.2024.128814}, journal = {Neurocomputing}, month = {1}, pages = {128814}, shortjournal = {Neurocomputing}, title = {Unveiling the potential of graph coloring in feature set partitioning: A study on high-dimensional datasets}, volume = {614}, year = {2025}, }
• (2025). Totipotent neural controllers for modular soft robots: Achieving specialization in body–brain co-evolution through hebbian learning. NEUCOM, 614, 128811. (www)

  @article{ , abstract = {Multi-cellular organisms typically originate from a single cell, the zygote, that then develops into a multitude of structurally and functionally specialized cells. The potential of generating all the specialized cells that make up an organism is referred to as cellular “totipotency”, a concept introduced by the German plant physiologist Haberlandt in the early 1900s. In an attempt to reproduce this mechanism in synthetic organisms, we present a model based on a kind of modular robot called Voxel-based Soft Robot (VSR), where both the body, i.e ., the arrangement of voxels, and the brain, i.e ., the Artificial Neural Network (ANN) controlling each module, are subject to an evolutionary process aimed at optimizing the locomotion capabilities of the robot. In an analogy between totipotent cells and totipotent ANN-controlled modules, we then include in our model an additional level of adaptation provided by Hebbian learning, which allows the ANNs to adapt their weights during the execution of the locomotion task. Our in silico experiments reveal two main findings. Firstly, we confirm the common intuition that Hebbian plasticity effectively allows better performance and adaptation. Secondly and more importantly, we verify for the first time that the performance improvements yielded by plasticity are in essence due to a form of specialization at the level of single modules (and their associated ANNs): thanks to plasticity, modules specialize to react in different ways to the same set of stimuli, i.e ., they become functionally and behaviorally different even though their ANNs are initialized in the same way. This mechanism, which can be seen as a form of totipotency at the level of ANNs, can have, in our view, profound implications in various areas of Artificial Intelligence (AI) and applications thereof, such as modular robotics and multi-agent systems.}, archive = {J_NEUCOM}, author = {Andrea Ferigo and Giovanni Iacca and Eric Medvet and Giorgia Nadizar}, doi = {10.1016/j.neucom.2024.128811}, journal = {Neurocomputing}, month = {1}, pages = {128811}, shortjournal = {Neurocomputing}, title = {Totipotent neural controllers for modular soft robots: Achieving specialization in body–brain co-evolution through hebbian learning}, volume = {614}, year = {2025}, }
• (2025). Learning from different perspectives for regret reduction in reinforcement learning: A free energy approach. NEUCOM, 614, 128797. (www)

  @article{ , abstract = {Reinforcement learning (RL) is the core method for interactive learning in living and artificial creatures. Nevertheless, in contrast to humans and animals, artificial RL agents are very slow in learning and suffer from the curse of dimensionality. This is partially due to using RL in isolation; i.e. lack of social learning and social diversity. We introduce a free energy-based social RL for learning novel tasks. Society is formed by the learning agent and some diverse virtual ones. That diversity is in their perception while all agents use the same interaction samples for learning and share the same action set. Individual difference in perception is mostly the cause of perceptual aliasing however, it can result in virtual agents’ faster learning in early trials. Our free energy method provides a knowledge integration method for the main agent to benefit from that diversity to reduce its regret. It rests upon Thompson sampling policy and behavioral policy of main and virtual agents. Therefore, it is applicable to a variety of tasks, discrete or continuous state space, model-free, and model-based tasks as well as to different reinforcement learning methods. Through a set of experiments, we show that this general framework highly improves learning speed and is clearly superior to previous existing methods. We also provide convergence proof.}, archive = {J_NEUCOM}, author = {Milad Ghorbani and Reshad Hosseini and Seyed Pooya Shariatpanahi and Majid Nili Ahmadabadi}, doi = {10.1016/j.neucom.2024.128797}, journal = {Neurocomputing}, month = {1}, pages = {128797}, shortjournal = {Neurocomputing}, title = {Learning from different perspectives for regret reduction in reinforcement learning: A free energy approach}, volume = {614}, year = {2025}, }
• (2025). Adversarial diffusion for few-shot scene adaptive video anomaly detection. NEUCOM, 614, 128796. (www)

  @article{ , abstract = {Few-shot anomaly detection for video surveillance is challenging due to the diverse nature of target domains. Existing methodologies treat it as a one-class classification problem, training on a reduced sample of nominal scenes. The focus is on either reconstructive or predictive frame methodologies to learn a manifold against which outliers can be detected during inference. We posit that the quality of image reconstruction or future frame prediction is inherently important in identifying anomalous pixels in video frames. In this paper, we enhance the image synthesis and mode coverage for video anomaly detection (VAD) by integrating a Denoising Diffusion model with a future frame prediction model. Our novel VAD pipeline includes a Generative Adversarial Network combined with denoising diffusion to learn the underlying non-anomalous data distribution and generate in one-step high fidelity future-frame samples. We further regularize the image reconstruction with perceptual quality metrics such as Multi-scale Structural Similarity Index Measure and Peak Signal-to-Noise Ratio , ensuring high-quality output under few episodic training iterations. Extensive experiments demonstrate that our method outperforms state-of-the-art techniques across multiple benchmarks, validating that high-quality image synthesis in frame prediction leads to robust anomaly detection in videos.}, archive = {J_NEUCOM}, author = {Yumna Zahid and Christine Zarges and Bernie Tiddeman and Jungong Han}, doi = {10.1016/j.neucom.2024.128796}, journal = {Neurocomputing}, month = {1}, pages = {128796}, shortjournal = {Neurocomputing}, title = {Adversarial diffusion for few-shot scene adaptive video anomaly detection}, volume = {614}, year = {2025}, }
• (2025). Augmented ELBO regularization for enhanced clustering in variational autoencoders. NEUCOM, 614, 128795. (www)

  @article{ , abstract = {With significant advances in deep neural networks, various new algorithms have emerged that effectively model latent structures within data, surpassing traditional clustering methods. Each data point is expected to belong to a single cluster in a typical clustering algorithm. However, clustering based on variational autoencoders (VAEs) represents the expectation of the overall clusters, denoted as c = 1 , … , K in the KL divergence term. Consequently, the latent embedding z can be learned to exist across multiple clusters with relatively balanced probabilities, rather than being strongly associated with a specific cluster. This study introduces an additional regularizer to encourage the latent embedding z to have a strong affiliation with specific clusters. We introduce optimization methods to maximize the ELBO that includes the newly added regularization term and explore methods to eliminate computationally challenging terms. The positive impact of this regularization on clustering accuracy was verified by examining the variance of the final cluster probabilities. Furthermore, an enhancement in the clustering performance was observed when regularization was introduced.}, archive = {J_NEUCOM}, author = {Kwangtek Na and Ju-Hong Lee and Eunchan Kim}, doi = {10.1016/j.neucom.2024.128795}, journal = {Neurocomputing}, month = {1}, pages = {128795}, shortjournal = {Neurocomputing}, title = {Augmented ELBO regularization for enhanced clustering in variational autoencoders}, volume = {614}, year = {2025}, }
• (2025). SDDA: A progressive self-distillation with decoupled alignment for multimodal image–text classification. NEUCOM, 614, 128794. (www)

  @article{ , abstract = {Multimodal image–text classification endeavors to deduce the correct category based on the information encapsulated in image–text pairs. Despite the commendable performance achieved by current image–text methodologies, the intrinsic multimodal heterogeneity persists as a challenge, with the contributions from diverse modalities exhibiting considerable variance. In this study, we address this issue by introducing a novel decoupled multimodal Self-Distillation (SDDA) approach, aimed at facilitating fine-grained alignment of shared and private features of image–text features in a low-dimensional space, thereby reducing information redundancy. Specifically, each modality representation is decoupled in an autoregressive manner into two segments within a modality-irrelevant/exclusive space. SDDA imparts additional knowledge transfer to each decoupled segment via self-distillation, while also offering flexible, richer multimodal knowledge supervision for unimodal features. Multimodal classification experiments conducted on two publicly available benchmark datasets verified the efficacy of the algorithm, demonstrating that SDDA surpasses the state-of-the-art baselines.}, archive = {J_NEUCOM}, author = {Xiaohao Chen and Qianjun Shuai and Feng Hu and Yongqiang Cheng}, doi = {10.1016/j.neucom.2024.128794}, journal = {Neurocomputing}, month = {1}, pages = {128794}, shortjournal = {Neurocomputing}, title = {SDDA: A progressive self-distillation with decoupled alignment for multimodal image–text classification}, volume = {614}, year = {2025}, }
• (2025). Mixed-scale cross-modal fusion network for referring image segmentation. NEUCOM, 614, 128793. (www)

  @article{ , abstract = {Referring image segmentation aims to segment the target by a given language expression. Recently, the bottom-up fusion network utilizes language features to highlight the most relevant regions during the visual encoder stage. However, it is not comprehensive that establish only the relationship between pixels and words. To alleviate this problem, we propose a mixed-scale cross-modal fusion method that widens the interaction between vision and language. Specially, at each stage, pyramid pooling is used to augment visual perception and improve the interaction between visual and linguistic features, thereby highlighting relevant regions in the visual data. Additionally, we employ a simple multi-scale feature fusion module to effectively combine multi-scale aligned features. Experiments conducted on Standard RIS benchmarks demonstrate that the proposed method achieves favorable performance against state-of-the- art approaches. Moreover, we conducted experiments on different visual backbones respectively, and the proposed method yielded better and significantly improved performance results.}, archive = {J_NEUCOM}, author = {Xiong Pan and Xuemei Xie and Jianxiu Yang}, doi = {10.1016/j.neucom.2024.128793}, journal = {Neurocomputing}, month = {1}, pages = {128793}, shortjournal = {Neurocomputing}, title = {Mixed-scale cross-modal fusion network for referring image segmentation}, volume = {614}, year = {2025}, }
• (2025). CRISP: A cross-modal integration framework based on the surprisingly popular algorithm for multimodal named entity recognition. NEUCOM, 614, 128792. (www)

  @article{ , abstract = {The multimodal named entity recognition task on social media involves recognizing named entities with textual and visual information, which is of great significance for information processing. Nevertheless, many existing models still face the following challenges. First, in the process of cross-modal interaction, the attention mechanism sometimes focuses on trivial parts in the images that are not relevant to entities, which not only neglects valuable information but also inevitably introduces visual noise. Second, the gate mechanism is widely used for filtering out visual information to reduce the influence of noise on text understanding. However, the gate mechanism neglects capturing fine-grained semantic relevance between modalities, which easily affects the filtration process. To address these issues, we propose a cross-modal integration framework based on the surprisingly popular algorithm, aiming at enhancing the integration of effective visual guidance and reducing the interference of irrelevant visual noise. Specifically, we design a dual-branch interaction module that includes the attention mechanism and the surprisingly popular algorithm, allowing the model to focus on valuable but overlooked parts in the images. Furthermore, we compute the matching degree between modalities at the multi-granularity level, using the Choquet integral to establish a more reasonable basis for filtering out visual noise. We have conducted extensive experiments on public datasets, and the experimental results demonstrate the advantages of our model.}, archive = {J_NEUCOM}, author = {Haitao Liu and Xianwei Xin and Jihua Song and Weiming Peng}, doi = {10.1016/j.neucom.2024.128792}, journal = {Neurocomputing}, month = {1}, pages = {128792}, shortjournal = {Neurocomputing}, title = {CRISP: A cross-modal integration framework based on the surprisingly popular algorithm for multimodal named entity recognition}, volume = {614}, year = {2025}, }
• (2025). A review of AI edge devices and lightweight CNN and LLM deployment. NEUCOM, 614, 128791. (www)

  @article{ , abstract = {Artificial Intelligence of Things (AIoT) which integrates artificial intelligence (AI) and the Internet of Things (IoT), has attracted increasing attention recently. With the remarkable development of AI, convolutional neural networks (CNN) have achieved great success from research to deployment in many applications. However, deploying complex and state-of-the-art (SOTA) AI models on edge applications is increasingly a big challenge. This paper investigates literature that deploys lightweight CNNs on AI edge devices in practice. We provide a comprehensive analysis of them and many practical suggestions for researchers: how to obtain/design lightweight CNNs, select suitable AI edge devices, and compress and deploy them in practice. Finally, future trends and opportunities are presented, including the deployment of large language models, trustworthy AI and robust deployment.}, archive = {J_NEUCOM}, author = {Kailai Sun and Xinwei Wang and Xi Miao and Qianchuan Zhao}, doi = {10.1016/j.neucom.2024.128791}, journal = {Neurocomputing}, month = {1}, pages = {128791}, shortjournal = {Neurocomputing}, title = {A review of AI edge devices and lightweight CNN and LLM deployment}, volume = {614}, year = {2025}, }
• (2025). HADT: Image super-resolution restoration using hybrid attention-dense connected transformer networks. NEUCOM, 614, 128790. (www)

  @article{ , abstract = {Image super-resolution (SR) plays a vital role in vision tasks, in which Transformer-based methods outperform conventional convolutional neural networks. Existing work usually uses residual linking to improve the performance, but this type of linking provides limited information transfer within the block. Also, existing work usually restricts the self-attention computation to a single window to improve feature extraction. This means transformer-based networks can only use feature information within a limited spatial range. To handle the challenge, this paper proposes a novel Hybrid Attention-Dense Connected Transformer Network (HADT) to utilize the potential feature information better. HADT is constructed by stacking an attentional transformer block (ATB), which contains an Effective Dense Transformer Block (EDTB) and a Hybrid Attention Block (HAB). EDTB combines dense connectivity and swin-transformer to enhance feature transfer and improve model representation, and meanwhile, HAB is used for cross-window information interaction and joint modeling of features for better visualization. Based on the experiments, our method is effective on SR tasks with magnification factors of 2, 3, and 4. For example, using the Urban100 dataset in an experiment with an amplification factor of 4 our method has a PSNR value that is 0.15 dB higher than the previous method and reconstructs a more detailed texture.}, archive = {J_NEUCOM}, author = {Ying Guo and Chang Tian and Jie Liu and Chong Di and Keqing Ning}, doi = {10.1016/j.neucom.2024.128790}, journal = {Neurocomputing}, month = {1}, pages = {128790}, shortjournal = {Neurocomputing}, title = {HADT: Image super-resolution restoration using hybrid attention-dense connected transformer networks}, volume = {614}, year = {2025}, }
• (2025). Adaptive feature alignment network with noise suppression for cross-domain object detection. NEUCOM, 614, 128789. (www)

  @article{ , abstract = {Recently, unsupervised domain adaptive object detection methods have been proposed to address the challenge of detecting objects across different domains without labeled data in the target domain. These methods focus on aligning features either at the image level or the instance level. However, due to the absence of annotations in the target domain, existing approaches encounter challenges such as background noise at the image level and prototype aggregation noise at the instance level. To tackle these issues, we introduce a novel adaptive feature alignment network for cross-domain object detection, comprising two key modules. Firstly, we present an adaptive foreground-aware attention module equipped with a set of learnable part prototypes for image-level alignment. This module dynamically generates foreground attention maps, enabling the model to prioritize foreground features, thus reducing the impact of background noise. Secondly, we propose a class-aware prototype alignment module incorporating an optimal transport algorithm for instance-level alignment. This module mitigates the adverse effects of region–prototype aggregation noise by aligning prototypes with instances based on their semantic similarities. By integrating these two modules, our approach achieves better image-level and instance-level feature alignment. Extensive experiments across three challenging scenarios demonstrate the effectiveness of our method, outperforming state-of-the-art approaches in terms of object detection performance.}, archive = {J_NEUCOM}, author = {Wei Jiang and Yujie Luan and Kewei Tang and Lijun Wang and Nan Zhang and Huiling Chen and Heng Qi}, doi = {10.1016/j.neucom.2024.128789}, journal = {Neurocomputing}, month = {1}, pages = {128789}, shortjournal = {Neurocomputing}, title = {Adaptive feature alignment network with noise suppression for cross-domain object detection}, volume = {614}, year = {2025}, }
• (2025). Physically-guided open vocabulary segmentation with weighted patched alignment loss. NEUCOM, 614, 128788. (www)

  @article{ , abstract = {Open vocabulary segmentation is a challenging task that aims to segment out the thousands of unseen categories. Directly applying CLIP to open-vocabulary semantic segmentation is challenging due to the granularity gap between its image-level contrastive learning and the pixel-level recognition required for segmentation. To address these challenges, we propose a unified pipeline that leverages physical structure regularization to enhance the generalizability and robustness of open vocabulary segmentation. By incorporating physical structure information, which is independent of the training data, we aim to reduce bias and improve the model’s performance on unseen classes. We utilize low-level structures such as edges and keypoints as regularization terms, as they are easier to obtain and strongly correlated with segmentation boundary information. These structures are used as pseudo-ground truth to supervise the model. Furthermore, inspired by the effectiveness of comparative learning in human cognition, we introduce the weighted patched alignment loss. This loss function contrasts similar and dissimilar samples to acquire low-dimensional representations that capture the distinctions between different object classes. By incorporating physical knowledge and leveraging weighted patched alignment loss, we aim to improve the model’s generalizability, robustness, and capability to recognize diverse object classes. The experiments on the COCO Stuff, Pascal VOC, Pascal Context-59, Pascal Context-459, ADE20K-150, and ADE20K-847 datasets demonstrate that our proposed method consistently improves baselines and achieves new state-of-the-art in the open vocabulary segmentation task.}, archive = {J_NEUCOM}, author = {Weide Liu and Jieming Lou and Xingxing Wang and Wei Zhou and Jun Cheng and Xulei Yang}, doi = {10.1016/j.neucom.2024.128788}, journal = {Neurocomputing}, month = {1}, pages = {128788}, shortjournal = {Neurocomputing}, title = {Physically-guided open vocabulary segmentation with weighted patched alignment loss}, volume = {614}, year = {2025}, }
• (2025). Dual-referenced assistive network for action quality assessment. NEUCOM, 614, 128786. (www)

  @article{ , abstract = {Action quality assessment (AQA) aims to evaluate the performing quality of a specific action. It is a challenging task as it requires to identify the subtle differences between the videos containing the same action. Most of existing AQA methods directly adopt a pretrained network designed for other tasks to extract video features, which are too coarse to describe fine-grained details of action quality. In this paper, we propose a novel Dual-Referenced Assistive (DuRA) network to polish original coarse-grained features into fine-grained quality-oriented representations. Specifically, we introduce two levels of referenced assistants to highlight the discriminative quality-related contents by comparing a target video and the referenced objects, instead of obtrusively estimating the quality score from an individual video. Firstly, we design a Rating-guided Attention module, which takes advantage of a series of semantic-level referenced assistants to acquire implicit hierarchical semantic knowledge and progressively emphasize quality-focused features embedded in original inherent information. Subsequently, we further design a couple of Consistency Preserving constraints, which introduce a set of individual-level referenced assistants to further eliminate score-unrelated information through more detailed comparisons of differences between actions. The experiments show that our proposed method achieves promising performance on the AQA-7 and MTL-AQA datasets.}, archive = {J_NEUCOM}, author = {Keyi Huang and Yi Tian and Chen Yu and Yaping Huang}, doi = {10.1016/j.neucom.2024.128786}, journal = {Neurocomputing}, month = {1}, pages = {128786}, shortjournal = {Neurocomputing}, title = {Dual-referenced assistive network for action quality assessment}, volume = {614}, year = {2025}, }
• (2025). Multi-attention associate prediction network for visual tracking. NEUCOM, 614, 128785. (www)

  @article{ , abstract = {Classification-regression prediction networks have realized impressive success in several modern deep trackers. However, there is an inherent difference between classification and regression tasks, so they have diverse even opposite demands for feature matching. Existed models always ignore the key issue and only employ a unified matching block in two task branches, decaying the decision quality. Besides, these models also struggle with decision misalignment situation. In this paper, we propose a multi-attention associate prediction network (MAPNet) to tackle the above problems. Concretely, two novel matchers, i.e., category-aware matcher and spatial-aware matcher, are first designed for feature comparison by integrating self, cross, channel or spatial attentions organically. They are capable of fully capturing the category-related semantics for classification and the local spatial contexts for regression, respectively. Then, we present a dual alignment module to enhance the correspondences between two branches, which is useful to find the optimal tracking solution. Finally, we describe a Siamese tracker built upon the proposed prediction network, which achieves the leading performance on five tracking benchmarks, consisting of LaSOT, TrackingNet, GOT-10k, TNL2k and UAV123, and surpasses other state-of-the-art approaches.}, archive = {J_NEUCOM}, author = {Xinglong Sun and Haijiang Sun and Shan Jiang and Jiacheng Wang and Xilai Wei and Zhonghe Hu}, doi = {10.1016/j.neucom.2024.128785}, journal = {Neurocomputing}, month = {1}, pages = {128785}, shortjournal = {Neurocomputing}, title = {Multi-attention associate prediction network for visual tracking}, volume = {614}, year = {2025}, }
• (2025). A three-stage model for camouflaged object detection. NEUCOM, 614, 128784. (www)

  @article{ , abstract = {Camouflaged objects are typically assimilated into their backgrounds and exhibit fuzzy boundaries. The complex environmental conditions and the high intrinsic similarity between camouflaged targets and their surroundings pose significant challenges in accurately locating and segmenting these objects in their entirety. While existing methods have demonstrated remarkable performance in various real-world scenarios, they still face limitations when confronted with difficult cases, such as small targets, thin structures, and indistinct boundaries. Drawing inspiration from human visual perception when observing images containing camouflaged objects, we propose a three-stage model that enables coarse-to-fine segmentation in a single iteration. Specifically, our model employs three decoders to sequentially process subsampled features, cropped features, and high-resolution original features. This proposed approach not only reduces computational overhead but also mitigates interference caused by background noise. Furthermore, considering the significance of multi-scale information, we have designed a multi-scale feature enhancement module that enlarges the receptive field while preserving detailed structural cues. Additionally, a boundary enhancement module has been developed to enhance performance by leveraging boundary information. Subsequently, a mask-guided fusion module is proposed to generate fine-grained results by integrating coarse prediction maps with high-resolution feature maps. Our network shows superior performance without introducing unnecessary complexities. Upon acceptance of the paper, the source code will be made publicly available at https://github.com/clelouch/TSNet .}, archive = {J_NEUCOM}, author = {Tianyou Chen and Hui Ruan and Shaojie Wang and Jin Xiao and Xiaoguang Hu}, doi = {10.1016/j.neucom.2024.128784}, journal = {Neurocomputing}, month = {1}, pages = {128784}, shortjournal = {Neurocomputing}, title = {A three-stage model for camouflaged object detection}, volume = {614}, year = {2025}, }
• (2025). Bipartite containment tracking for nonlinear MASs under FDI attack based on model-free adaptive iterative learning control. NEUCOM, 614, 128783. (www)

  @article{ , abstract = {The bipartite containment control problem for a type of heterogeneous multi-agent systems (MASs) under false data injection (FDI) attack is handled in this work by using the distributed model-free adaptive iterative learning control scheme with attack compensation. The unknown non-affine nonlinear dynamics of each agent is first transformed into an equivalent attack-related data model along the iteration axis using a compact form dynamic linearization method. Then, a distributed model-free adaptive iterative learning bipartite containment control (DMFAILBCC) scheme is constructed by employing I/O data from MASs, and the convergence is proved by rigorous mathematical analysis In addition, the updated control method and the convergence analysis will be extended to iteration switching topologies. Finally, the performance of the two proposed schemes is validated through numerical simulations and comparisons with different control schemes.}, archive = {J_NEUCOM}, author = {Xinning He and Zhongsheng Hou}, doi = {10.1016/j.neucom.2024.128783}, journal = {Neurocomputing}, month = {1}, pages = {128783}, shortjournal = {Neurocomputing}, title = {Bipartite containment tracking for nonlinear MASs under FDI attack based on model-free adaptive iterative learning control}, volume = {614}, year = {2025}, }
• (2025). A pseudo-3D coarse-to-fine architecture for 3D medical landmark detection. NEUCOM, 614, 128782. (www)

  @article{ , abstract = {The coarse-to-fine architecture is a benchmark method designed to enhance the accuracy of 3D medical landmark detection. However, incorporating 3D convolutional neural networks into the coarse-to-fine architecture leads to a significant increase in model parameters, making it costly for deployment in clinical applications. This paper introduces a novel lightweight pseudo-3D coarse-to-fine architecture, consisting of a Plane-wise Attention Pseudo-3D (PA-P3D) model and a Spatial Separation Pseudo-3D (SS-P3D) model. The PA-P3D inherits the lightweight structure of the general pseudo-3D and enhances cross-plane feature interaction in 3D medical images. On the other hand, the SS-P3D replaces the 3D model with three spatially separated 2D models to simultaneously detect 2D landmarks on axial, sagittal, and coronal planes. In comparison to the conventional coarse-to-fine architecture, the proposed method requires only approximately a quarter of the model parameters (60% reduced by PA-P3D and 40% reduced by SS-P3D) while simultaneously improving landmark detection performance. Experimental results demonstrate the effectiveness of the proposed method, achieving state-of-the-art performance on both a public dataset for mandibular molar landmark detection and a private dataset for cephalometric landmark detection. Overall, this paper highlights the potential of the coarse-to-fine method for cost-effective model deployment, thanks to its lightweight model structure.}, archive = {J_NEUCOM}, author = {Li Cui and Boyan Liu and Guikun Xu and Jixiang Guo and Wei Tang and Tao He}, doi = {10.1016/j.neucom.2024.128782}, journal = {Neurocomputing}, month = {1}, pages = {128782}, shortjournal = {Neurocomputing}, title = {A pseudo-3D coarse-to-fine architecture for 3D medical landmark detection}, volume = {614}, year = {2025}, }
• (2025). Intelligible graph contrastive learning with attention-aware for recommendation. NEUCOM, 614, 128781. (www)

  @article{ , abstract = {Recommender systems are an important tool for information retrieval, which can aid in the solution of the issue of information overload. Recently, contrastive learning has shown remarkable performance in recommendation by data augmentation processes to address highly sparse data. Our paper proposes an Int elligible G raph C ontrastive L earning with attention-aware (IntGCL) for recommendation. Particularly, our IntGCL first introduces a novel attention-aware matrix into graph convolutional networks (GCN) to identify the importance between users and items, which is constructed to preserve the importance between users and items by a random walk with a restart strategy and can enhance the intelligibility of our model. Then, the attention-aware matrix is further utilised to guide the generation of a graph-generative model with attention-aware and a graph-denoising model for automatically generating two trainable contrastive views for data augmentation, which can de-noise and further enhance the intelligibility. Comprehensive experiments on four real-world datasets indicate the superiority of our IntGCL approach over multiple state-of-the-art methods. Our datasets and source code are available at https://github.com/restarthxr/InpGCL .}, archive = {J_NEUCOM}, author = {Xian Mo and Zihang Zhao and Xiaoru He and Hang Qi and Hao Liu}, doi = {10.1016/j.neucom.2024.128781}, journal = {Neurocomputing}, month = {1}, pages = {128781}, shortjournal = {Neurocomputing}, title = {Intelligible graph contrastive learning with attention-aware for recommendation}, volume = {614}, year = {2025}, }
• (2025). Perceptual metric for face image quality with pixel-level interpretability. NEUCOM, 614, 128780. (www)

  @article{ , abstract = {This paper tackles the shortcomings of image evaluation metrics in evaluating facial image quality. Conventional metrics do neither accurately reflect the unique attributes of facial images nor correspond with human visual perception. To address these issues, we introduce a novel metric designed specifically for faces, utilizing a learning-based adversarial framework. This framework comprises a generator for simulating face restoration and a discriminator for quality evaluation. Drawing inspiration from facial neuroscience studies, our metric emphasizes the importance of primary facial features, acknowledging that minor changes in the eyes, nose, and mouth can significantly impact perception. Another key limitation of existing image evaluation metrics is their focus on numerical values at the image level, without providing insight into how different areas of the image contribute to the overall assessment. Our proposed metric offers interpretability regarding how each region of the image is evaluated. Comprehensive experimental results confirm that our face-specific metric surpasses traditional general image quality assessment metrics for facial images, including both full-reference and no-reference methods. The code and models are available at https://github.com/AIM-SKKU/IFQA .}, archive = {J_NEUCOM}, author = {Byungho Jo and In Kyu Park and Sungeun Hong}, doi = {10.1016/j.neucom.2024.128780}, journal = {Neurocomputing}, month = {1}, pages = {128780}, shortjournal = {Neurocomputing}, title = {Perceptual metric for face image quality with pixel-level interpretability}, volume = {614}, year = {2025}, }
• (2025). Imperceptible rhythm backdoor attacks: Exploring rhythm transformation for embedding undetectable vulnerabilities on speech recognition. NEUCOM, 614, 128779. (www)

  @article{ , abstract = {Speech recognition is an essential start ring of human–computer interaction. Recently, deep learning models have achieved excellent success in this task. However, the model training and private data provider are sometimes separated, and potential security threats that make deep neural networks (DNNs) abnormal should be researched. In recent years, the typical threats, such as backdoor attacks, have been analysed in speech recognition systems. The existing backdoor methods are based on data poisoning. The attacker adds some incorporated changes to benign speech spectrograms or changes the speech components, such as pitch and timbre. As a result, the poisoned data can be detected by human hearing or automatic deep algorithms. To improve the stealthiness of data poisoning, we propose a non-neural and fast algorithm called R andom S pectrogram R hythm T ransformation (RSRT) in this paper. The algorithm combines four steps to generate stealthy poisoned utterances. From the perspective of rhythm component transformation, our proposed trigger stretches or squeezes the mel spectrograms and recovers them back to signals. The operation keeps timbre and content unchanged for good stealthiness. Our experiments are conducted on two kinds of speech recognition tasks, including testing the stealthiness of poisoned samples by speaker verification and automatic speech recognition. The results show that our method is effective and stealthy. The rhythm trigger needs a low poisoning rate and gets a very high attack success rate.}, archive = {J_NEUCOM}, author = {Wenhan Yao and Jiangkun Yang and Yongqiang He and Jia Liu and Weiping Wen}, doi = {10.1016/j.neucom.2024.128779}, journal = {Neurocomputing}, month = {1}, pages = {128779}, shortjournal = {Neurocomputing}, title = {Imperceptible rhythm backdoor attacks: Exploring rhythm transformation for embedding undetectable vulnerabilities on speech recognition}, volume = {614}, year = {2025}, }
• (2025). ERAT-DLoRA: Parameter-efficient tuning with enhanced range adaptation in time and depth aware dynamic LoRA. NEUCOM, 614, 128778. (www)

  @article{ , abstract = {Despite their potential, the industrial deployment of large language models (LLMs) is constrained by traditional fine-tuning procedures that are both resource-intensive and time-consuming. Low-Rank Adaptation (LoRA) has emerged as a pioneering methodology for addressing these challenges. By integrating low-rank decomposition matrices into network weights to reduce trainable parameters, LoRA effectively accelerates the adaptation process. While research on LoRA primarily focuses on adjusting low-rank matrices, DyLoRA optimizes the rank-setting mechanism to avoid extensive effort in rank size training and searching. However, DyLoRA rank configuration mechanism has its own limitation. First, DyLoRA sets the same rank size for all the low-rank adaptation layers at each time step. Given that layers with different depth contain distinct information, they should have varying rank values to accurately capture their unique characteristics. Second, the truncated phase selected for ordering representation based on nested dropout regulation is only half dynamic, continuously dropping tail units, thereby limiting its ability to access information. In this work, we propose a novel technique, enhanced range adaptation in time and depth aware dynamic LoRA (ERAT-DLoRA) to address these problems. The ERAT-DLoRA method introduces a dynamic range to the truncated phase that makes the truncated phase fully dynamic. Additionally, we design a time and layer-aware dynamic rank to ensure appropriate adjustments at different time steps and layer levels. We evaluate our solution on natural languages understanding and language generation tasks. Extensive evaluation results demonstrate the effectiveness of the proposed method.}, archive = {J_NEUCOM}, author = {Dan Luo and Kangfeng Zheng and Chunhua Wu and Xiujuan Wang and Jvjie Wang}, doi = {10.1016/j.neucom.2024.128778}, journal = {Neurocomputing}, month = {1}, pages = {128778}, shortjournal = {Neurocomputing}, title = {ERAT-DLoRA: Parameter-efficient tuning with enhanced range adaptation in time and depth aware dynamic LoRA}, volume = {614}, year = {2025}, }
• (2025). Robust source-free domain adaptation with anti-adversarial samples training. NEUCOM, 614, 128777. (www)

  @article{ , abstract = {Unsupervised source-free domain adaptation methods aim to transfer knowledge acquired from labeled source domain to an unlabeled target domain, where the source data are not accessible during target domain adaptation and it is prohibited to minimize domain gap by pairwise calculation of the samples from the source and target domains. Previous approaches assign pseudo label to target data using pre-trained source model to progressively train the target model in a self-learning manner. However, incorrect pseudo label may adversely affect prediction in the target domain. Furthermore, they overlook the generalization ability of the source model, which primarily affects the initial prediction of the target model. Therefore, we propose an effective framework based on adversarial training to train the target model for source-free domain adaptation. Specifically, adversarial training is an effective technique to enhance the robustness of deep neural networks. By generating anti-adversarial examples and adversarial examples, the pseudo label of target data can be corrected further by adversarial training and a more optimal performance in both accuracy and robustness is achieved. Moreover, owing to the inherent domain distribution difference between source and target domains, mislabeled target samples exist inevitably. So a target sample filtering scheme is proposed to refine pseudo label to further improve the prediction capability on the target domain. Experiments conducted on benchmark tasks demonstrate that the proposed method outperforms existing approaches.}, archive = {J_NEUCOM}, author = {Zhirui Wang and Liu Yang and Yahong Han}, doi = {10.1016/j.neucom.2024.128777}, journal = {Neurocomputing}, month = {1}, pages = {128777}, shortjournal = {Neurocomputing}, title = {Robust source-free domain adaptation with anti-adversarial samples training}, volume = {614}, year = {2025}, }
• (2025). Implicit expression recognition enhanced table-filling for aspect sentiment triplet extraction. NEUCOM, 614, 128776. (www)

  @article{ , abstract = {Aspect sentiment triplet extraction (ASTE) is a challenging task in aspect-based sentiment analysis (ABSA), involving the identification of aspect terms, opinion terms, and their corresponding sentiment polarities within comments to form triplets. The emergence of more realistic DMASTE datasets, featuring diverse domains, implicit aspect terms, and longer comments, poses challenges for existing methods. In particular, these methods struggle with recognizing implicit expressions effectively and capturing sufficient information. To overcome these hurdles, we propose an implicit expression recognition enhanced table-filling (IERET) method. This approach integrates modeling of overall implicit expression in sentences and employs a bidirectional information aggregation module to capture word pair information comprehensively. During the decoding process, a table-filling method accurately delineates aspect-opinion pair boundaries. Experimental results across in-domain, single-source cross-domain, and multi-source cross-domain on the DMASTE dataset demonstrate that our proposed IERET method achieves state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Yanbo Li and Qing He and Nisuo Du and Qingni He}, doi = {10.1016/j.neucom.2024.128776}, journal = {Neurocomputing}, month = {1}, pages = {128776}, shortjournal = {Neurocomputing}, title = {Implicit expression recognition enhanced table-filling for aspect sentiment triplet extraction}, volume = {614}, year = {2025}, }
• (2025). An efficient re-parameterization feature pyramid network on YOLOv8 to the detection of steel surface defect. NEUCOM, 614, 128775. (www)

  @article{ , abstract = {In the field of steel production, the detection of steel surface defects is one of the most important guarantees for the quality of steel production. In the process of defect detection, there are problems regarding the noise of the acquisition background, the scale of defects, and the detection speed. At present, in the face of complex steel surface defects, realizing efficient real-time steel surface defect detection has become a difficult problem. In this paper, we propose a lightweight and efficient real-time defect detection method, LDE-YOLO, based on YOLOv8. First, we propose a lightweight multi-scale feature extraction module, LighterMSMC, which not only achieves a lightweight backbone network, but also effectively guarantees the long range dependence of the features, so as to realize multi-scale feature extraction more efficiently. Secondly, we propose lightweight re-parameterized feature pyramid, DE-FPN, in which the sparse patterns of the overall features and the detailed features of the local features are efficiently captured by the DE-Block, and then efficiently fused by the PAN feature fusion structure. Finally, we propose Efficient Head, which lightens the model by group convolution while its improves the diagonal correlation of the feature maps on some specific datasets, thus enhancing the detection performance. Our proposed LDE-YOLO obtains 80.8 mAP and 75.5 FPS on NEU-DET , 80.5 mAP and 75.5 FPS on GC10-DET. It obtains 2.5 mAP and 4.7 mAP enhancement compared to the baseline model, and the detection speed is also improved by 10.4 FPS, while in terms of the number of floating point operations and parameters of the model reduced by 60.2% and 49.1%, which is sufficient to illustrate its lightweight effectiveness and realize an efficient real-time steel surface defect detection model.}, archive = {J_NEUCOM}, author = {Weining Xie and Weifeng Ma and Xiaoyong Sun}, doi = {10.1016/j.neucom.2024.128775}, journal = {Neurocomputing}, month = {1}, pages = {128775}, shortjournal = {Neurocomputing}, title = {An efficient re-parameterization feature pyramid network on YOLOv8 to the detection of steel surface defect}, volume = {614}, year = {2025}, }
• (2025). IoU-guided siamese network with high-confidence template fusion for visual tracking. NEUCOM, 614, 128774. (www)

  @article{ , abstract = {Existing IoU-guided trackers use IoU score to weight the classification score only in testing phase, this model mismatch between training and testing phases leads to poor tracking performance especially when facing background distractors. In this paper, we propose an IoU-guided Siamese network with High-confidence template fusion (SiamIH) for visual tracking. An IoU-guided distractor suppression network is proposed, which uses IoU information to guide classification in training phase and testing phase, and makes the tracking model to suppress background distractors. To cope with appearance variations, we design a high-confidence template fusion network that fuses APCE-based high-confidence template and the initial template to generate more reliable template. Experimental results on OTB2013, OTB2015, UAV123, LaSOT, and GOT10k demonstrate that the proposed SiamIH achieves state-of-the-art tracking performance.}, archive = {J_NEUCOM}, author = {Zhigang Liu and Hao Huang and Hongyu Dong and Fuyuan Xing}, doi = {10.1016/j.neucom.2024.128774}, journal = {Neurocomputing}, month = {1}, pages = {128774}, shortjournal = {Neurocomputing}, title = {IoU-guided siamese network with high-confidence template fusion for visual tracking}, volume = {614}, year = {2025}, }
• (2025). Preserving text space integrity for robust compositional zero-shot learning via mixture of pretrained experts. NEUCOM, 614, 128773. (www)

  @article{ , abstract = {In the current landscape of Compositional Zero-Shot Learning (CZSL) methods that leverage CLIP, the predominant approach is based on prompt learning paradigms. These methods encounter significant computational complexity when dealing with a large number of categories. Additionally, when confronted with new classification tasks, there is a necessity to learn the prompts again, which can be both time-consuming and resource-intensive. To address these challenges, We present a new methodology, named the M ixture of P retrained E xpert (MoPE), for enhancing Compositional Zero-shot Learning through Logit-Level Fusion with Multi Expert Fusion Module. The MoPE skillfully blends the benefits of extensive pre-trained models like CLIP, Bert, GPT-3 and Word2Vec for effectively tackling Compositional Zero-shot Learning. Firstly, we extract the text label space for each language model individually, then map the visual feature vectors to their respective text spaces. This maintains the integrity and structure of the original text space. During this process, the pre-trained expert parameters are kept frozen. The mapping of visual features to the corresponding text spaces is subject to learning and could be considered as multiple learnable visual experts. In the model fusion phase, we propose a new fusion strategy that features a gating mechanism that adjusts the contributions of various models dynamically. This enables our approach to adapt more effectively to a range of tasks and data sets. The method’s robustness is demonstrated by the fact that the language model is not tailored to specific downstream task datasets or losses. This preserves the larger model’s topology and expands the potential for application. Preliminary experiments conducted on the UT-Zappos, AO-Clever, and C-GQA datasets indicate that MoPE performs competitively when compared to existing techniques.}, archive = {J_NEUCOM}, author = {Zehua Hao and Fang Liu and Licheng Jiao and Yaoyang Du and Shuo Li and Hao Wang and Pengfang Li and Xu Liu and Puhua Chen}, doi = {10.1016/j.neucom.2024.128773}, journal = {Neurocomputing}, month = {1}, pages = {128773}, shortjournal = {Neurocomputing}, title = {Preserving text space integrity for robust compositional zero-shot learning via mixture of pretrained experts}, volume = {614}, year = {2025}, }
• (2025). Active self-semi-supervised learning for few labeled samples. NEUCOM, 614, 128772. (www)

  @article{ , abstract = {Training deep models with limited annotations poses a significant challenge when applied to diverse practical domains. Employing semi-supervised learning alongside the self-supervised model offers the potential to enhance label efficiency. However, this approach faces a bottleneck in reducing the need for labels. We observed that the semi-supervised model disrupts valuable information from self-supervised learning when only limited labels are available. To address this issue, this paper proposes a simple yet effective framework, active self-semi-supervised learning (AS3L). AS3L bootstraps semi-supervised models with prior pseudo-labels (PPL). These PPLs are obtained by label propagation over self-supervised features. Based on the observations the accuracy of PPL is not only affected by the quality of features but also by the selection of the labeled samples. We develop active learning and label propagation strategies to obtain accurate PPL. Consequently, our framework can significantly improve the performance of models in the case of limited annotations while demonstrating fast convergence. On the image classification tasks across four datasets, our method outperforms the baseline by an average of 5.4%. Additionally, it achieves the same accuracy as the baseline method in about 1/3 of the training time.}, archive = {J_NEUCOM}, author = {Ziting Wen and Oscar Pizarro and Stefan Williams}, doi = {10.1016/j.neucom.2024.128772}, journal = {Neurocomputing}, month = {1}, pages = {128772}, shortjournal = {Neurocomputing}, title = {Active self-semi-supervised learning for few labeled samples}, volume = {614}, year = {2025}, }
• (2025). Addressing imperfect symmetry: A novel symmetry-learning actor-critic extension. NEUCOM, 614, 128771. (www)

  @article{ , abstract = {Symmetry, a fundamental concept to understand our environment, often oversimplifies reality from a mathematical perspective. Humans are a prime example, deviating from perfect symmetry in terms of appearance and cognitive biases (e.g. having a dominant hand). Nevertheless, our brain can easily overcome these imperfections and efficiently adapt to symmetrical tasks. The driving motivation behind this work lies in capturing this ability through reinforcement learning. To this end, we introduce Adaptive Symmetry Learning (ASL) — a model-minimization actor-critic extension that addresses incomplete or inexact symmetry descriptions by adapting itself during the learning process. ASL consists of a symmetry fitting component and a modular loss function that enforces a common symmetric relation across all states while adapting to the learned policy. The performance of ASL is compared to existing symmetry-enhanced methods in a case study involving a four-legged ant model for multidirectional locomotion tasks. The results show that ASL can recover from large perturbations and generalize knowledge to hidden symmetric states. It achieves comparable or better performance than alternative methods in most scenarios, making it a valuable approach for leveraging model symmetry while compensating for inherent perturbations.}, archive = {J_NEUCOM}, author = {Miguel Abreu and Luís Paulo Reis and Nuno Lau}, doi = {10.1016/j.neucom.2024.128771}, journal = {Neurocomputing}, month = {1}, pages = {128771}, shortjournal = {Neurocomputing}, title = {Addressing imperfect symmetry: A novel symmetry-learning actor-critic extension}, volume = {614}, year = {2025}, }
• (2025). Observer-based adaptive neural network event-triggered quantized control for active suspensions with actuator saturation. NEUCOM, 614, 128770. (www)

  @article{ , abstract = {This paper proposes an adaptive neural network event-triggered and quantized output feedback control scheme for quarter vehicle active suspensions with actuator saturation. The scheme uses neural networks to approximate the unknown parts of the active suspension. When the system states of the suspension are not entirely available, a state observer is designed to estimate the unknown states. The measurable system states, partially estimated observer states, neural network weights, and a filtered virtual control are sequentially event-triggered, quantified, and transmitted to the controller via in-vehicle networks. The problem of non-differentiable virtual control is solved using dynamic surface control technology in the backstepping quantized control design. Integrating a Gaussian error function and a first-order auxiliary subsystem compensates for the nonlinearity caused by asymmetric saturation. Theoretical analysis proves that all error signals of the closed-loop active suspension system are semi-globally uniformly ultimately bounded, and the Zeno phenomenon can be ruled out. Simulation results validate the effectiveness of the proposed control method.}, archive = {J_NEUCOM}, author = {Tiechao Wang and Hongyang Zhang and Shuai Sui}, doi = {10.1016/j.neucom.2024.128770}, journal = {Neurocomputing}, month = {1}, pages = {128770}, shortjournal = {Neurocomputing}, title = {Observer-based adaptive neural network event-triggered quantized control for active suspensions with actuator saturation}, volume = {614}, year = {2025}, }
• (2025). PIDNODEs: Neural ordinary differential equations inspired by a proportional–integral–derivative controller. NEUCOM, 614, 128769. (www)

  @article{ , abstract = {Neural Ordinary Differential Equations (NODEs) are a novel family of infinite-depth neural-net models through solving ODEs and their adjoint equations. In this paper, we present a strategy to enhance the training and inference of NODEs by integrating a Proportional–Integral–Derivative (PID) controller into the framework of Heavy Ball NODE, resulting in the proposed PIDNODEs and its generalized version, GPIDNODEs. By leveraging the advantages of control, PIDNODEs and GPIDNODEs can address the stiff ODE challenges by adjusting the parameters (i.e., K p , K i and K d ) in the PID module. The experiments confirm the superiority of PIDNODEs/GPIDNODEs over other NODE baselines on different computer vision and pattern recognition tasks, including image classification, point cloud separation and learning long-term dependencies from irregular time-series data for a physical dynamic system. These experiments demonstrate that the proposed models have higher accuracy and fewer function evaluations while alleviating the dilemma of exploding and vanishing gradients, particularly when learning long-term dependencies from a large amount of data.}, archive = {J_NEUCOM}, author = {Pengkai Wang and Song Chen and Jiaxu Liu and Shengze Cai and Chao Xu}, doi = {10.1016/j.neucom.2024.128769}, journal = {Neurocomputing}, month = {1}, pages = {128769}, shortjournal = {Neurocomputing}, title = {PIDNODEs: Neural ordinary differential equations inspired by a proportional–integral–derivative controller}, volume = {614}, year = {2025}, }
• (2025). A bi-consolidating model for joint relational triple extraction. NEUCOM, 614, 128768. (www)

  @article{ , abstract = {Current methods to extract relational triples directly make a prediction based on a possible entity pair in a raw sentence without depending on entity recognition. The task suffers from a serious semantic overlapping problem, in which several relation triples may share one or two entities in a sentence. In this paper, based on a two-dimensional sentence representation, a bi-consolidating model is proposed to address this problem by simultaneously reinforcing the local and global semantic features relevant to a relation triple. This model consists of a local consolidation component and a global consolidation component. The first component uses a pixel difference convolution to enhance semantic information of a possible triple representation from adjacent regions and mitigate noise in neighboring neighbors. The second component strengthens the triple representation based a channel attention and a spatial attention, which has the advantage to learn remote semantic dependencies in a sentence. They are helpful to improve the performance of both entity identification and relation type classification in relation triple extraction. After evaluated on several publish datasets, the bi-consolidating model achieves competitive performance. Analytical experiments demonstrate the effectiveness of our model for relational triple extraction and give motivation for other natural language processing tasks.}, archive = {J_NEUCOM}, author = {Xiaocheng Luo and Yanping Chen and Ruixue Tang and Caiwei Yang and Ruizhang Huang and Yongbin Qin}, doi = {10.1016/j.neucom.2024.128768}, journal = {Neurocomputing}, month = {1}, pages = {128768}, shortjournal = {Neurocomputing}, title = {A bi-consolidating model for joint relational triple extraction}, volume = {614}, year = {2025}, }
• (2025). A survey of deep learning algorithms for colorectal polyp segmentation. NEUCOM, 614, 128767. (www)

  @article{ , abstract = {Early detecting and removing cancerous colorectal polyps can effectively reduce the risk of colorectal cancer. Computer intelligent segmentation techniques (CIST) can improve the detection rate of polyp by drawing the boundaries of colorectal polyps clearly and completely. Four challenges that encountered in deep learning methods for the task of colorectal polyp segmentation are considered, including the limitations of classical deep learning (DL) algorithms, the impact of data set quantity and quality, the diversity of intrinsic characteristics of lesions and the heterogeneity of images in different center datasets. The improved DL algorithms for intelligent polyp segmentation are detailed along with the key neural network modules being designed to deal with above challenges. In addition, the public and private datasets of colorectal polyp images and videos are summarized, respectively. At the end of this paper, the development trends of polyp segmentation algorithm based on deep learning are discussed.}, archive = {J_NEUCOM}, author = {Sheng Li and Yipei Ren and Yulin Yu and Qianru Jiang and Xiongxiong He and Hongzhang Li}, doi = {10.1016/j.neucom.2024.128767}, journal = {Neurocomputing}, month = {1}, pages = {128767}, shortjournal = {Neurocomputing}, title = {A survey of deep learning algorithms for colorectal polyp segmentation}, volume = {614}, year = {2025}, }
• (2025). Double machine learning for partially linear mediation models with high-dimensional confounders. NEUCOM, 614, 128766. (www)

  @article{ , abstract = {To estimate and statistically infer the direct and indirect effects of exposure and mediator variables while accounting for high-dimensional confounding variables, we propose a partially linear mediation model to incorporate a flexible mechanism of confounders. To obtain asymptotically efficient estimators for the effects of interest under the influence of the nuisance functions with high-dimensional confounders, we construct two Neyman-orthogonal score functions to remove regularization bias. Flexible machine learning methods and data splitting with cross-fitting are employed to address the overfitting issue and estimate unknown nuisance functions efficiently. We rigorously investigate the asymptotic expressions of the proposed estimators for the direct, indirect and total effects and then derive their asymptotic normality properties. In addition, two Wald statistics are constructed to test the direct and indirect effects, respectively, and their limiting distributions are obtained. The satisfactory performance of our proposed estimators is demonstrated by simulation results and a genome-wide analysis of blood DNA methylation dataset.}, archive = {J_NEUCOM}, author = {Jichen Yang and Yujing Shao and Jin Liu and Lei Wang}, doi = {10.1016/j.neucom.2024.128766}, journal = {Neurocomputing}, month = {1}, pages = {128766}, shortjournal = {Neurocomputing}, title = {Double machine learning for partially linear mediation models with high-dimensional confounders}, volume = {614}, year = {2025}, }
• (2025). Sparse discriminant manifold projections for automatic depression recognition. NEUCOM, 614, 128765. (www)

  @article{ , abstract = {In recent years, depression has become an increasingly serious problem globally. Previous research have shown that EEG-based depression recognition is a promising technique to serve as auxiliary diagnosis methods that provide assistance to clinicians. Typically, in clinical studies, due to the multichannel nature of EEG, the extracted features usually are high-dimensional and contain many redundant information. Therefore, it is necessary to perform dimensionality reduction before classification to improve the performance of machine learning algorithms. However,existing dimensionality reduction techniques do not design the objective function based on the characteristics of EEG signal and the goal of depression recognition, so they are less suitable for dimensionality reduction of EEG features. To solve this problem, in this paper we propose a novel dimensionality reduction technique called sparse discriminant manifold projections(SDMP) for depression recognition. Specifically, the use of the ℓ 2 -norm instead of the squared ℓ 2 -norm as a similarity measure in the objective function reduces sensitivity to noise and outliers. Moreover, the local geometric structure and global discriminative properties of data are integrated, which makes the extracted features more discriminative. Finally, the ℓ 2 , 1 -norm regularization is introduced to achieve feature selection. Furthermore, The formulation is extended to the ℓ 2 , p -norm regularization case, which is more likely to offer better sparsity when 0 < p < 1 . Extensive experiments on EEG data show that the SDMP achieves the competitive performance compared with other state-of-the-art dimensionality reduction methods. It also shows the practical application value of our method in detecting depression.}, archive = {J_NEUCOM}, author = {Lu Zhang and Jitao Zhong and Qinglin Zhao and Shi Qiao and Yushan Wu and Bin Hu and Sujie Ma and Hong Peng}, doi = {10.1016/j.neucom.2024.128765}, journal = {Neurocomputing}, month = {1}, pages = {128765}, shortjournal = {Neurocomputing}, title = {Sparse discriminant manifold projections for automatic depression recognition}, volume = {614}, year = {2025}, }
• (2025). Diffusion model conditioning on gaussian mixture model and negative gaussian mixture gradient. NEUCOM, 614, 128764. (www)

  @article{ , abstract = {Diffusion models (DMs) are a type of generative model that has had a significant impact on image synthesis and beyond. They can incorporate a wide variety of conditioning inputs — such as text or bounding boxes — to guide generation. In this work, we introduce a novel conditioning mechanism that applies Gaussian mixture models (GMMs) for feature conditioning, which helps steer the denoising process in DMs. Drawing on set theory, our comprehensive theoretical analysis reveals that the conditional latent distribution based on features differs markedly from that based on classes. Consequently, feature-based conditioning tends to generate fewer defects than class-based conditioning. Experiments are designed and carried out and the experimental results support our theoretical findings as well as effectiveness of proposed feature conditioning mechanism. Additionally, we propose a new gradient function named the Negative Gaussian Mixture Gradient (NGMG) and incorporate it into the training of diffusion models alongside an auxiliary classifier. We theoretically demonstrate that NGMG offers comparable advantages to the Wasserstein distance, serving as a more effective cost function when learning distributions supported by low-dimensional manifolds, especially in contrast to many likelihood-based cost functions, such as KL divergences.}, archive = {J_NEUCOM}, author = {Weiguo Lu and Xuan Wu and Deng Ding and Jinqiao Duan and Jirong Zhuang and Gangnan Yuan}, doi = {10.1016/j.neucom.2024.128764}, journal = {Neurocomputing}, month = {1}, pages = {128764}, shortjournal = {Neurocomputing}, title = {Diffusion model conditioning on gaussian mixture model and negative gaussian mixture gradient}, volume = {614}, year = {2025}, }
• (2025). CoFiNet: Unveiling camouflaged objects with multi-scale finesse. NEUCOM, 614, 128763. (www)

  @article{ , abstract = {Camouflaged Object Detection (COD) is a critical aspect of computer vision aimed at identifying concealed objects, with applications spanning military, industrial, medical and monitoring domains. To address the problem of poor detail segmentation effect, we introduce a novel method for camouflaged object detection, named CoFiNet. Our approach primarily focuses on multi-scale feature fusion and extraction, with special attention to the model’s segmentation effectiveness for detailed features, enhancing its ability to effectively detect camouflaged objects. CoFiNet adopts a coarse-to-fine strategy. A multi-scale feature integration module is laveraged to enhance the model’s capability of fusing context feature. A multi-activation selective kernel module is leveraged to grant the model the ability to autonomously alter its receptive field, enabling it to selectively choose an appropriate receptive field for camouflaged objects of different sizes. During mask generation, we employ the dual-mask strategy for image segmentation, separating the reconstruction of coarse and fine masks, which significantly enhances the model’s learning capacity for details. Comprehensive experiments were conducted on four different datasets, demonstrating that CoFiNet achieves state-of-the-art performance across all datasets. The experiment results of CoFiNet underscore its effectiveness in camouflaged object detection and highlight its potential in various practical application scenarios.}, archive = {J_NEUCOM}, author = {Cunhan Guo and Heyan Huang}, doi = {10.1016/j.neucom.2024.128763}, journal = {Neurocomputing}, month = {1}, pages = {128763}, shortjournal = {Neurocomputing}, title = {CoFiNet: Unveiling camouflaged objects with multi-scale finesse}, volume = {614}, year = {2025}, }
• (2025). Industrial and medical anomaly detection through cycle-consistent adversarial networks. NEUCOM, 614, 128762. (www)

  @article{ , abstract = {In this study, a new Anomaly Detection (AD) approach for industrial and medical images is proposed. This method leverages the theoretical strengths of unsupervised learning and the data availability of both normal and abnormal classes. Indeed, the AD is often formulated as an unsupervised task, implying only normal images during training. These normal images are devoted to be reconstructed through an autoencoder architecture, for instance. However, the information contained in abnormal data, when available, is also valuable for this reconstruction. The model would be able to identify its weaknesses by also learning how to transform an abnormal image into a normal one. This abnormal-to-normal reconstruction helps the entire model to learn better than a single normal-to-normal reconstruction. To be able to exploit abnormal images, the proposed method uses Cycle-Generative Adversarial Networks (Cycle-GAN) for (ab)normal-to-normal translation. After an input image has been reconstructed by the normal generator, an anomaly score quantifies the differences between the input and its reconstruction. Based on a threshold set to satisfy a business quality constraint, the input image is then flagged as normal or not. The proposed method is evaluated on industrial and medical datasets. The results demonstrate accurate performance with a zero false negative constraint compared to state-of-the-art methods. Quantitatively, our method reaches an accuracy under a zero false negative constraint of 79.89%, representing an improvement of about 17% compared to competitors. The code is available at https://github.com/ValDelch/CycleGANS-AnomalyDetection .}, archive = {J_NEUCOM}, author = {Arnaud Bougaham and Valentin Delchevalerie and Mohammed El Adoui and Benoît Frénay}, doi = {10.1016/j.neucom.2024.128762}, journal = {Neurocomputing}, month = {1}, pages = {128762}, shortjournal = {Neurocomputing}, title = {Industrial and medical anomaly detection through cycle-consistent adversarial networks}, volume = {614}, year = {2025}, }
• (2025). A survey of graph neural networks and their industrial applications. NEUCOM, 614, 128761. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) have emerged as a powerful tool for analyzing and modeling graph-structured data. In recent years, GNNs have gained significant attention in various domains. This review paper aims to provide an overview of the state-of-the-art graph neural network techniques and their industrial applications. First, we introduce the fundamental concepts and architectures of GNNs, highlighting their ability to capture complex relationships and dependencies in graph data. We then delve into the variants and evolution of graphs, including directed graphs, heterogeneous graphs, dynamic graphs, and hypergraphs. Next, we discuss the interpretability of GNN, and GNN theory including graph augmentation, expressivity, and over-smoothing. Finally, we introduce the specific use cases of GNNs in industrial settings, including finance, biology, knowledge graphs, recommendation systems, Internet of Things (IoT), and knowledge distillation. This review paper highlights the immense potential of GNNs in solving real-world problems, while also addressing the challenges and opportunities for further advancement in this field.}, archive = {J_NEUCOM}, author = {Haoran Lu and Lei Wang and Xiaoliang Ma and Jun Cheng and Mengchu Zhou}, doi = {10.1016/j.neucom.2024.128761}, journal = {Neurocomputing}, month = {1}, pages = {128761}, shortjournal = {Neurocomputing}, title = {A survey of graph neural networks and their industrial applications}, volume = {614}, year = {2025}, }
• (2025). MHEC: One-shot relational learning of knowledge graphs completion based on multi-hop information enhancement. NEUCOM, 614, 128760. (www)

  @article{ , abstract = {With the wide application of knowledge graphs, knowledge graph completion has garnered increasing attention in recent years. However, we find that the long tail relation is more common in the KG. These relations typically do not have a large number of triples for training and are referred to as few-shot relations. The knowledge graph completion in the few-shot scenario is a major challenge currently. The current mainstream knowledge graph completion algorithms have the following drawbacks. The metric-based methods lack interpretability of results, while the algorithms based on path interaction are not suitable for few-shot scenarios and the availability of the model is limited in sparse knowledge graphs. In this paper, we propose a one-shot relational learning of knowledge graphs completion based on multi-hop information enhancement(MHEC). Firstly, MHEC extracts entity concepts from multi-hop paths to obtain task related entity concepts and filters out noisy neighbor attributes. Then, MHEC combines multi-hop path information between head and tail to represent entity pairs. Compared to previous completion methods that only consider structural features of entities, MHEC considers the reasoning logic between entity pairs, which not only includes structural features but also possesses rich semantic features. Next, MHEC introduces a reasoning process in the completion task to address the issues of lack of interpretability in the one-shot scenario. In addition, to improve completion and reasoning quality in sparse knowledge graphs, MHEC utilizes contrastive learning to enhance pre-training vector representations of entities and relations and proposes a matching processor that leverages the semantic information of pre-training vectors to assist the reasoning model in expanding the multi-hop paths. Experiments demonstrate that MHEC outperforms the state-of-the-art completion techniques on real-world datasets NELL-One and FB15k237-One.}, archive = {J_NEUCOM}, author = {Ruixin Ma and Buyun Gao and Weihe Wang and Longfei Wang and Xiaoru Wang and Liang Zhao}, doi = {10.1016/j.neucom.2024.128760}, journal = {Neurocomputing}, month = {1}, pages = {128760}, shortjournal = {Neurocomputing}, title = {MHEC: One-shot relational learning of knowledge graphs completion based on multi-hop information enhancement}, volume = {614}, year = {2025}, }
• (2025). Subclass consistency regularization for learning with noisy labels based on contrastive learning. NEUCOM, 614, 128759. (www)

  @article{ , abstract = {A prominent effect of label noise on neural networks is the disruption of the consistency of predictions. While prior efforts primarily focused on predictions’ consistency at the individual instance level, they often fell short of fully harnessing the consistency across multiple instances. This paper introduces subclass consistency regularization (SCCR) to maximize the potential of this collective consistency of predictions. SCCR mitigates the impact of label noise on neural networks by imposing constraints on the consistency of predictions within each subclass. However, constructing high-quality subclasses poses a formidable challenge, which we formulate as a special clustering problem. To efficiently establish these subclasses, we incorporate a clustering-based contrastive learning framework. Additionally, we introduce the Q -enhancing algorithm to tailor the contrastive learning framework, ensuring alignment with subclass construction. We conducted comprehensive experiments using benchmark datasets and real datasets to evaluate the effectiveness of our proposed method under various scenarios with differing noise rates. The results unequivocally demonstrate the enhancement in classification accuracy, especially in challenging high-noise settings. Moreover, the refined contrastive learning framework significantly elevates the quality of subclasses even in the presence of noise. Furthermore, we delve into the compatibility of contrastive learning and learning with noisy labels, using the projection head as an illustrative example. This investigation sheds light on an aspect that has hitherto been overlooked in prior research efforts.}, archive = {J_NEUCOM}, author = {Xinkai Sun and Sanguo Zhang}, doi = {10.1016/j.neucom.2024.128759}, journal = {Neurocomputing}, month = {1}, pages = {128759}, shortjournal = {Neurocomputing}, title = {Subclass consistency regularization for learning with noisy labels based on contrastive learning}, volume = {614}, year = {2025}, }
• (2025). Simulation-based effective comparative analysis of neuron circuits for neuromorphic computation systems. NEUCOM, 614, 128758. (www)

  @article{ , abstract = {The spiking neural networks (SNN) that are inspired by the human brain offers wider scope for application in the growth of neuromorphic computing systems due to their brain level computational capabilities, reduced power consumption, and minimal data movement cost, among other advantages. Spike-based neurons and synapses are the essential building blocks of SNN, and their efficient implementation is vital to their performance enhancement. In this regard, the design and implementation of spiking neurons have been the major focus among the researchers. In this paper, functioning of different leaky integrate fire (LIF)-based spiking neuron circuits like frequency adaptable CMOS-based LIF, resistor-capacitor-based (RC) LIF, and volatile memristor-based LIF are subjected to comparison. The work mainly focuses on revealing analysis of spike duration and amplitude, number of spikes produced during excitation period, threshold operation, field of application, and various other significant parameters of aforementioned neuron circuits. Extensive simulations of these circuits are carried out utilizing the Cadence Virtuoso simulation environment in order to validate their behavior. Further, a brief comparative analysis is executed considering into account the attributes like circuit complexity, supply voltage, firing rate, membrane capacitance, nature of input/output, refractory mechanism, and energy consumption per spike. This work seeks to assist researchers in selecting an appropriate LIF model to efficiently construct memristors and/or non-memristors based SNN for certain application.}, archive = {J_NEUCOM}, author = {Deepthi M.S. and Shashidhara H.R. and Jayaramu Raghu and Rudraswamy S.B.}, doi = {10.1016/j.neucom.2024.128758}, journal = {Neurocomputing}, month = {1}, pages = {128758}, shortjournal = {Neurocomputing}, title = {Simulation-based effective comparative analysis of neuron circuits for neuromorphic computation systems}, volume = {614}, year = {2025}, }
• (2025). Out-of-vocabulary handling and topic quality control strategies in streaming topic models. NEUCOM, 614, 128757. (www)

  @article{ , abstract = {Topic models have become ubiquitous tools for analyzing streaming data. However, existing streaming topic models suffer from several limitations when applied to real-world data streams. This includes the inability to accommodate evolving vocabularies and control topic quality throughout the streaming process. In this paper, we propose a novel streaming topic modeling approach that dynamically adapts to the changing nature of data streams. Our method leverages Byte-Pair Encoding embedding (BPEmb) to resolve the out-of-vocabulary problem that arises with new words in the stream. Additionally, we introduce a topic change variable that provides fine-grained control over topics’ parameter updates and present a preservation approach to retain high-coherence topics at each time step, helping preserve semantic quality. To further enhance model adaptability, our method allows dynamical adjustment of topic space size as needed. To the best of our knowledge, we are the first to address the expansion of vocabulary and maintain topic quality during the streaming process. Extensive experiments show the superior effectiveness of our method.}, archive = {J_NEUCOM}, author = {Tung Nguyen and Tung Pham and Linh Ngo Van and Ha-Bang Ban and Khoat Than}, doi = {10.1016/j.neucom.2024.128757}, journal = {Neurocomputing}, month = {1}, pages = {128757}, shortjournal = {Neurocomputing}, title = {Out-of-vocabulary handling and topic quality control strategies in streaming topic models}, volume = {614}, year = {2025}, }
• (2025). Auditing privacy budget of differentially private neural network models. NEUCOM, 614, 128756. (www)

  @article{ , abstract = {In recent years, neural network models are used in various tasks. To eliminate privacy concern, differential privacy (DP) is introduced to the training phase of neural network models. However, introducing DP into neural network models is very subtle and error-prone, resulting in that some differentially private neural network models may not achieve privacy guarantee claimed. In this paper, we propose a method, which can audit privacy budget of differentially private neural network models. The proposed method is general and can be used to audit some other AI models. We elaborate on how to audit privacy budget of basic DP mechanisms and neural network models by the proposed method first. Then, we run experiments to verify our method. Experiment results indicate that the proposed method is better than the advanced method and the auditing precise is high when the privacy budget is small. In particular, when auditing privacy budget of ResNet-18 over CIFAR-10 protected by the differentially private mechanism with theoretical privacy budget 0.2, the accuracy of our method is about 17 times that of the state-of-the-art method. For the simpler dataset FMNIST, the accuracy of our method is about 32 times that of the state-of-the-art method when theoretical privacy budget is 0.2.}, archive = {J_NEUCOM}, author = {Wen Huang and Zhishuo Zhang and Weixin Zhao and Jian Peng and Wenzheng Xu and Yongjian Liao and Shijie Zhou and Ziming Wang}, doi = {10.1016/j.neucom.2024.128756}, journal = {Neurocomputing}, month = {1}, pages = {128756}, shortjournal = {Neurocomputing}, title = {Auditing privacy budget of differentially private neural network models}, volume = {614}, year = {2025}, }
• (2025). Interpretable few-shot learning with online attribute selection. NEUCOM, 614, 128755. (www)

  @article{ , abstract = {Few-shot learning (FSL) presents a challenging learning problem in which only a few samples are available for each class. Decision interpretation is more important in few-shot classification due to a greater chance of error compared to traditional classification. However, the majority of the previous FSL methods are black-box models. In this paper, we propose an inherently interpretable model for FSL based on human-friendly attributes. Previously, human-friendly attributes have been utilized to train models with the potential for human interaction and interpretability. However, such approaches are not directly extendible to the few-shot classification scenario. Moreover, we propose an online attribute selection mechanism to effectively filter out irrelevant attributes in each episode. The attribute selection mechanism improves accuracy and helps with interpretability by reducing the number of attributes that participate in each episode. We further propose a mechanism that automatically detects the episodes where the pool of available human-friendly attributes is insufficient, and subsequently augments it by engaging some learned unknown attributes. We demonstrate that the proposed method achieves results on par with black-box few-shot learning models on four widely used datasets. We also empirically evaluate the level of decision alignment between different models and human understanding and show that our model outperforms the comparison methods based on this criterion.}, archive = {J_NEUCOM}, author = {Mohammad Reza Zarei and Majid Komeili}, doi = {10.1016/j.neucom.2024.128755}, journal = {Neurocomputing}, month = {1}, pages = {128755}, shortjournal = {Neurocomputing}, title = {Interpretable few-shot learning with online attribute selection}, volume = {614}, year = {2025}, }
• (2025). Boosting multi-document summarization with hierarchical graph convolutional networks. NEUCOM, 614, 128753. (www)

  @article{ , abstract = {The input of the multi-document summarization task is usually long and has high redundancy. Encoding multiple documents is a challenge for the Seq2Seq architecture. The way of concatenating multiple documents into a sequence ignores the relation between documents. Attention-based Seq2Seq architectures have slightly improved the cross-document relation modeling for multi-document summarization. However, these methods ignore the relation between sentences, and there is little improvement that can be achieved through the attention mechanism alone. This paper proposes a hierarchical approach to leveraging the relation between words, sentences, and documents for abstractive multi-document summarization. Our model employs the Graph Convolutional Networks (GCN) for capturing the cross-document and cross-sentence relations. The GCN module can enrich semantic representations by generating high-level hidden features. Our model achieves significant improvement over the attention-based baseline, beating the Hierarchical Transformer by 3.4/1.64, 1.92/1.44 ROUGE-1/2 F1 points on the Multi-News and WikiSum datasets, respectively. Experimental results demonstrate that our delivered method brings substantial improvements over several strong baselines.}, archive = {J_NEUCOM}, author = {Yingjie Song and Li Yang and Wenming Luo and Xiong Xiao and Zhuo Tang}, doi = {10.1016/j.neucom.2024.128753}, journal = {Neurocomputing}, month = {1}, pages = {128753}, shortjournal = {Neurocomputing}, title = {Boosting multi-document summarization with hierarchical graph convolutional networks}, volume = {614}, year = {2025}, }
• (2025). Correcting the distribution of batch normalization signals for trojan mitigation. NEUCOM, 614, 128752. (www)

  @article{ , abstract = {Backdoor (Trojan) attacks represent a significant adversarial threat to deep neural networks (DNNs). In such attacks, the presence of an attacker’s backdoor trigger causes a test instance to be misclassified into the attacker’s chosen target class. Post-training mitigation methods aim to rectify these misclassifications, ensuring that poisoned models correctly classify backdoor-triggered samples. These methods require the defender to have access to a small, clean dataset and the potentially compromised DNN. However, most defenses rely on parameter fine-tuning, making their effectiveness dependent on the dataset size available to the defender. To overcome the limitations of existing approaches, we propose a method that rectifies misclassifications by correcting the altered distribution of internal layer activations of backdoor-triggered instances. Distribution alterations are corrected by applying simple transformations to internal activations. Notably, our method does not modify any trainable parameters of the DNN, yet it achieves generally good mitigation performance against various backdoor attacks and benchmarks. Consequently, our approach demonstrates robustness even with a limited amount of clean data, making it highly practical for real-world applications. The effectiveness of our approach is validated through both theoretical analysis and extensive experimentation. The appendix is provided as an electronic component and can be accessed via the link in the footnote. 2 The source codes can be found in the link 3 at the footnote.}, archive = {J_NEUCOM}, author = {Xi Li and Zhen Xiang and David J. Miller and George Kesidis}, doi = {10.1016/j.neucom.2024.128752}, journal = {Neurocomputing}, month = {1}, pages = {128752}, shortjournal = {Neurocomputing}, title = {Correcting the distribution of batch normalization signals for trojan mitigation}, volume = {614}, year = {2025}, }
• (2025). Advances in vehicle re-identification techniques: A survey. NEUCOM, 614, 128745. (www)

  @article{ , abstract = {The development of vehicle re-identification technology has significantly enhanced the operational efficiency of intelligent transportation systems and smart cities, attributed to the advancement of artificial intelligence technologies such as deep learning and transformer models. By accurately tracking and identifying the same vehicle under different cameras, the technology not only greatly enhances the ability of urban safety monitoring, traffic management and accident investigation, but also provides powerful technical support for the development of intelligent transportation. This paper explores the shift from traditional to deep learning approaches in vehicle re-identification, highlighting the rise of Transformer models. We assess both non-visual and vision-based re-identification technologies, with a special focus on the deep feature-based methods across supervised, unsupervised, and semi-supervised learning. And we summarize the performance of supervised and unsupervised methods on the VeRi-776 and VehicleID datasets. Finally, this paper outlines six directions for the future development of vehicle Re-ID technology, highlighting its potential applications in various areas such as smart city traffic management.}, archive = {J_NEUCOM}, author = {Xiaoying Yi and Qi Wang and Qi Liu and Yikang Rui and Bin Ran}, doi = {10.1016/j.neucom.2024.128745}, journal = {Neurocomputing}, month = {1}, pages = {128745}, shortjournal = {Neurocomputing}, title = {Advances in vehicle re-identification techniques: A survey}, volume = {614}, year = {2025}, }
• (2025). Multi-contrast image clustering via multi-resolution augmentation and momentum-output queues. NEUCOM, 614, 128738. (www)

  @article{ , abstract = {Contrastive clustering has emerged as an efficacious technique in the domain of deep clustering, leveraging the interplay between paired samples and the learning capabilities of deep network architectures. However, the augmentation strategies employed in the existing methods do not fully utilize the information of images, coupled with the limitation of the number of negative samples makes the clustering performance suffer. In this study, we propose a novel clustering approach that incorporates momentum-output queues and multi-resolution augmentation strategies to effectively address these limitations. Initially, we deploy a multi-resolution augmentation strategy, transforming conventional augmentations into distinct global and local perspectives across various resolutions. This approach comprehensively harnesses inter-image information to construct a multi-contrast model with multi-view inputs. Subsequently, we introduce momentum-output queues, which are designed to store a large number of negative samples without increasing the computational cost, thereby enhancing the clustering effect. Within our joint optimization framework, sample features are derived from both the original and momentum encoders for instance-level contrastive learning. Simultaneously, features produced exclusively by the original encoder within the same batch are employed for cluster-level contrastive learning. Our experimental results on five challenging datasets substantiate the superior performance of our method over existing state-of-the-art techniques.}, archive = {J_NEUCOM}, author = {Sheng Jin and Shuisheng Zhou and Dezheng Kong and Banghe Han}, doi = {10.1016/j.neucom.2024.128738}, journal = {Neurocomputing}, month = {1}, pages = {128738}, shortjournal = {Neurocomputing}, title = {Multi-contrast image clustering via multi-resolution augmentation and momentum-output queues}, volume = {614}, year = {2025}, }
• (2025). Cross-view action recognition understanding from exocentric to egocentric perspective. NEUCOM, 614, 128731. (www)

  @article{ , abstract = {Understanding action recognition in egocentric videos has emerged as a vital research topic with numerous practical applications. With the limitation in the scale of egocentric data collection, learning robust deep learning-based action recognition models remains difficult. Transferring knowledge learned from the large-scale exocentric data to the egocentric data is challenging due to the difference in videos across views. Our work introduces a novel cross-view learning approach to action recognition (CVAR) that effectively transfers knowledge from the exocentric to the selfish view. First, we present a novel geometric-based constraint into the self-attention mechanism in Transformer based on analyzing the camera positions between two views. Then, we propose a new cross-view self-attention loss learned on unpaired cross-view data to enforce the self-attention mechanism learning to transfer knowledge across views. Finally, to further improve the performance of our cross-view learning approach, we present the metrics to measure the correlations in videos and attention maps effectively. Experimental results on standard egocentric action recognition benchmarks, i.e., Charades-Ego, EPIC-Kitchens-55, and EPIC-Kitchens-100, have shown our approach’s effectiveness and state-of-the-art performance.}, archive = {J_NEUCOM}, author = {Thanh-Dat Truong and Khoa Luu}, doi = {10.1016/j.neucom.2024.128731}, journal = {Neurocomputing}, month = {1}, pages = {128731}, shortjournal = {Neurocomputing}, title = {Cross-view action recognition understanding from exocentric to egocentric perspective}, volume = {614}, year = {2025}, }
• (2025). Enhancing video frame interpolation with region of motion loss and self-attention mechanisms: A dual approach to address large, nonlinear motions. NEUCOM, 614, 128728. (www)

  @article{ , abstract = {Video frame interpolation is particularly challenging when dealing with large and non-linear object motions, often resulting in poor frame quality and motion artifacts. In this study, we introduce a novel dual-approach methodology for video frame interpolation that effectively addresses these complexities. Our method consists of two key components: a Region of Motion (RoM) loss and self-attention mechanisms. The RoM loss is designed to spotlight significant movements within frames. This is achieved by employing feature-matching techniques that assign tailored weights during the training process, ensuring that areas of intense motion are given priority. This is facilitated by the computation of optical flow, which identifies crucial feature points and highlights regions of significant motion for targeted enhancement. Our method incorporates self-attention mechanisms to maintain inter-frame continuity while emphasizing the unique attributes of individual frames. The self-attention scores reduce motion discrepancies and enhance the distinctiveness and texture quality of each frame. We validate the efficacy of our approach through extensive evaluations on benchmark datasets, including Vimeo-90K, Middlebury, UCF101, and SNU-Film.}, archive = {J_NEUCOM}, author = {Yeongjoon Kim and Sunkyu Kwon and Donggoo Kang and Hyunmin Lee and Joonki Paik}, doi = {10.1016/j.neucom.2024.128728}, journal = {Neurocomputing}, month = {1}, pages = {128728}, shortjournal = {Neurocomputing}, title = {Enhancing video frame interpolation with region of motion loss and self-attention mechanisms: A dual approach to address large, nonlinear motions}, volume = {614}, year = {2025}, }
• (2025). Easy and effective! data augmentation for knowledge-aware dialogue generation via multi-perspective sentences interaction. NEUCOM, 614, 128724. (www)

  @article{ , abstract = {In recent years, knowledge-based dialogue generation has garnered significant attention due to its capacity to produce informative and coherent responses through the integration of external knowledge into models. However, obtaining high-quality knowledge that aligns with the dialogue content poses a considerable challenge, necessitating substantial time and resources. To tackle the issue of limited dialogue data, a majority of research endeavors concentrate on data augmentation to augment the volume of training data. Regrettably, these methods overlook knowledge augmentation, leading to a restricted diversity in input data and yielding enhancements solely in specific metrics. Real-world conversations exhibit a spectrum of characteristics, including repetitions, reversals, and interruptions, demanding a heightened level of data diversity. In this study, we introduce a straightforward yet effective data augmentation technique known as Multi-perspective Sentence Interaction to bolster the connections among sentences from varied viewpoints. Through an examination of target responses from multiple dialogue perspectives, we enhance our comprehension of the relationships between dialogue sentences, thereby facilitating the expansion of knowledge-based dialogue data. Through experiments conducted on various knowledge-based dialogue datasets and utilizing different models, our findings illustrate a notable enhancement in the quality of model generation facilitated by our method. Specifically, we observed a 3.5% enhancement in reply accuracy and a 0.1506 increase in diversity (DIST-2). Moreover, there was a substantial improvement in knowledge selection accuracy by 19.04% and a reduction in model perplexity by 31.48%.}, archive = {J_NEUCOM}, author = {Sisi Peng and Dan Qu and Wenlin Zhang and Hao Zhang and Shunhang Li and Minchen Xu}, doi = {10.1016/j.neucom.2024.128724}, journal = {Neurocomputing}, month = {1}, pages = {128724}, shortjournal = {Neurocomputing}, title = {Easy and effective! data augmentation for knowledge-aware dialogue generation via multi-perspective sentences interaction}, volume = {614}, year = {2025}, }
• (2025). Graph-based similarity of deep neural networks. NEUCOM, 614, 128722. (www)

  @article{ , abstract = {Understanding the enigmatic black-box representations within Deep Neural Networks (DNNs) is an essential problem in the community of deep learning. An initial step towards tackling this conundrum lies in quantifying the degree of similarity between these representations. Various approaches have been proposed in prior research, however, as the field of representation similarity continues to develop, existing metrics are not compatible with each other and struggling to meet the evolving demands. To address this, we propose a comprehensive similarity measurement framework inspired by the natural graph structure formed by samples and their corresponding features within the neural network. Our novel Graph-Based Similarity (GBS) framework gauges the similarity of DNN representations by constructing a weighted, undirected graph based on the output of hidden layers. In this graph, each node represents an input sample, and the edges are weighted in accordance with the similarity between pairs of nodes. Consequently, the measure of representational similarity can be derived through graph similarity metrics, such as layer similarity. We observe that input samples belonging to the same category exhibit dense interconnections within the deep layers of the DNN. To quantify this phenomenon, we employ a motif-based approach to gauge the extent of these interconnections. This serves as a metric to evaluate whether the representation derived from one model can be accurately classified by another. Experimental results show that GBS gets state-of-the-art performance in the sanity check. We also extensively evaluate GBS on downstream tasks to demonstrate its effectiveness, including measuring the transferability of pretrained models and model pruning.}, archive = {J_NEUCOM}, author = {Zuohui Chen and Yao Lu and JinXuan Hu and Qi Xuan and Zhen Wang and Xiaoniu Yang}, doi = {10.1016/j.neucom.2024.128722}, journal = {Neurocomputing}, month = {1}, pages = {128722}, shortjournal = {Neurocomputing}, title = {Graph-based similarity of deep neural networks}, volume = {614}, year = {2025}, }
• (2025). Breaking the gap between label correlation and instance similarity via new multi-label contrastive learning. NEUCOM, 614, 128719. (www)

  @article{ , abstract = {Multi-label text classification (MLTC) is a fundamental yet challenging task in natural language processing. Existing MLTC models mostly learn text representations and label correlations, separately; while the instance-level correlation, which is crucial for the classification is ignored. To rectify this, we propose a new multi-label contrastive learning model, that captures instance-level correlations, for the MLTC task. Specifically, we first learn label representations by using Graph Convolutional Network (GCN) on label co-occurrence graphs. We next learn text representations by taking label correlations into consideration. Through an attention mechanism, instance-level correlation can be established. To better utilize label correlations, we propose a new contrastive learning model, whose learning is guided by a new learning objective, to further refine label representations. We finally implement a k -NN mechanism, that identifies k nearest neighbors of a given text for final prediction. Intensive experimental studies over benchmark multi-label datasets demonstrate the effectiveness of our approach.}, archive = {J_NEUCOM}, author = {Xin Wang and Wang Zhang and Yuhong Wu and Xingpeng Zhang and Chao Wang and Huayi Zhan}, doi = {10.1016/j.neucom.2024.128719}, journal = {Neurocomputing}, month = {1}, pages = {128719}, shortjournal = {Neurocomputing}, title = {Breaking the gap between label correlation and instance similarity via new multi-label contrastive learning}, volume = {614}, year = {2025}, }
• (2025). Deep belief network with fuzzy parameters and its membership function sensitivity analysis. NEUCOM, 614, 128716. (www)

  @article{ , abstract = {Over the last few years, deep belief networks (DBNs) have been extensively utilized for efficient and reliable performance in several complex systems. One critical factor contributing to the enhanced learning of the DBN layers is the handling of network parameters, such as weights and biases. The efficient training of these parameters significantly influences the overall enhanced performance of the DBN. However, the initialization of these parameters is often random, and the data samples are normally corrupted by unwanted noise. This causes the uncertainty to arise among weights and biases of the DBNs, which ultimately hinders the performance of the network. To address this challenge, we propose a novel DBN model with weights and biases represented using fuzzy sets. The approach systematically handles inherent uncertainties in parameters resulting in a more robust and reliable training process. We show the working of the proposed algorithm considering four widely used benchmark datasets such as: MNSIT, n-MNIST (MNIST with additive white Gaussian noise (AWGN) and MNIST with motion blur) and CIFAR-10. The experimental results show superiority of the proposed approach as compared to classical DBN in terms of robustness and enhanced performance. Moreover, it has the capability to produce equivalent results with a smaller number of nodes in the hidden layer; thus, reducing the computational complexity of the network architecture. Additionally, we also study the sensitivity analysis for stability and consistency by considering different membership functions to model the uncertain weights and biases. Further, we establish the statistical significance of the obtained results by conducting both one-way and Kruskal-Wallis analyses of variance tests.}, archive = {J_NEUCOM}, author = {Amit K. Shukla and Pranab K. Muhuri}, doi = {10.1016/j.neucom.2024.128716}, journal = {Neurocomputing}, month = {1}, pages = {128716}, shortjournal = {Neurocomputing}, title = {Deep belief network with fuzzy parameters and its membership function sensitivity analysis}, volume = {614}, year = {2025}, }
• (2025). TeST: Temporal–spatial separated transformer for temporal action localization. NEUCOM, 614, 128688. (www)

  @article{ , abstract = {Temporal action localization is a fundamental task in video understanding. Existing methods fall into three categories: anchor-based, actionness-guided, and anchor-free. Anchor-based and actionness-guided models need huge computation resources to process redundant proposals or enumerate every possible proposal. Anchor-free models with lighter parameters become a more attractive option as temporal actions become more complex. However, they typically struggle to achieve high performance due to the need to aggregate global temporal–spatial features at every time step. To overcome this limitation, we design three efficient transformer-based architectures, bringing two advantages: (i) the global receptive field of transformers enables models to aggregate spatial and temporal at each time step, and (ii) the transformers could capture the moment-level feature, enhancing localization performance. Our designed architectures are adapted to any framework, thus we propose a simple but effective anchor-free framework named TeST . Compared to strong baselines, TeST achieves 0.96% to 3.20% improvement on two real-world datasets. Meanwhile, it improves time efficiency by 1.36 times and space efficiency by 1.08 times. Further experiments prove the effectiveness of TeST ’s modules. Implementation of our work is available at https://github.com/whr000001/TeST .}, archive = {J_NEUCOM}, author = {Herun Wan and Minnan Luo and Zhihui Li and Yang Wang}, doi = {10.1016/j.neucom.2024.128688}, journal = {Neurocomputing}, month = {1}, pages = {128688}, shortjournal = {Neurocomputing}, title = {TeST: Temporal–spatial separated transformer for temporal action localization}, volume = {614}, year = {2025}, }
• (2025). Global and local semantic enhancement of samples for cross-modal hashing. NEUCOM, 614, 128678. (www)

  @article{ , abstract = {Hashing becomes popular in cross-modal retrieval due to its exceptional performance in both search and storage. However, existing cross-modal hashing (CMH) methods may (a) neglect to learn sufficient modal-specific information, and (b) fail to fully exploit sample semantics. To address these issues, we propose a method called Semantic Enhancement of Sample Hashing (SESH). First, SESH employs a global modal-specific learning strategy to draw overall shared information and global modal-specific information by factoring the mapping matrix. Second, SESH introduces manifold learning and a local modal-specific learning strategy to extract additional local modal-specific and modal-shared data under label guidance. Meanwhile, local modal-specific information is integrated with global modal-specific details to add rich modal-specific information. Third, SESH uses discrete maximum similarity and orthogonal constraint transformation to enhance both global and local semantic information, embedding more discriminative information into the Hamming space. Finally, an efficient discrete optimization algorithm is proposed to generate the hash codes directly. Experiments on three datasets demonstrate the superior performance of SESH. The source code will be available at https://github.com/kokorording/SESH .}, archive = {J_NEUCOM}, author = {Shaohua Teng and Yongqi Chen and Zefeng Zheng and Wei Zhang and Peipei Kang and Naiqi Wu}, doi = {10.1016/j.neucom.2024.128678}, journal = {Neurocomputing}, month = {1}, pages = {128678}, shortjournal = {Neurocomputing}, title = {Global and local semantic enhancement of samples for cross-modal hashing}, volume = {614}, year = {2025}, }
• (2025). Whole slide cervical cancer classification via graph attention networks and contrastive learning. NEUCOM, 613, 128787. (www)

  @article{ , abstract = {Cervical cancer is one of the most common cancers among women, which seriously threatens women’s health. Early screening can reduce the incidence rate and mortality. Thinprep cytologic test (TCT) is one of the important means of cytological screening, which has high sensitivity and specificity, and has been widely used in the early screening of cervical cancer. The automatic diagnosis of whole slide images (WSIs) by computers can effectively improve the efficiency and accuracy of doctors’ diagnoses. However, current methods ignore the intrinsic relationships between cervical cells in WSIs and neglect contextual information from the surrounding suspicious areas, and therefore limit their robustness and generalizability. In this paper, we propose a novel two-stage method to implement the automatic diagnosis of WSIs, which constructs Graph Attention Networks (GAT) based on local and global fields respectively to capture their contextual information in a hierarchical manner. In the first stage, we extract representative patches from each WSI through suspicious cell detection, and then employ a Local GAT to classify cervical cells by capturing correlations between suspicious cells in image tiles. This classification process provides the confidence and feature vectors for each suspicious cell. In the second stage, we perform WSI classification using a Global GAT model. We construct graphs corresponding to top- K g and bottom- K g cells for each WSI based on results from Local GAT, and introduce a supervised contrastive learning strategy to enhance the discriminative power of the extracted features. Experimental results demonstrate that our proposed method outperforms conventional approaches and effectively showcases the benefits of supervised contrastive learning. Our source code and example data are available at https://github.com/feimanman/Whole-Slide-Cervical-Cancer-Classification.}, archive = {J_NEUCOM}, author = {Manman Fei and Xin Zhang and Dongdong Chen and Zhiyun Song and Qian Wang and Lichi Zhang}, doi = {10.1016/j.neucom.2024.128787}, journal = {Neurocomputing}, month = {1}, pages = {128787}, shortjournal = {Neurocomputing}, title = {Whole slide cervical cancer classification via graph attention networks and contrastive learning}, volume = {613}, year = {2025}, }
• (2025). Multi-level discriminator based contrastive learning for multiplex networks. NEUCOM, 613, 128754. (www)

  @article{ , abstract = {Graph embedding is a technique for obtaining low-dimensional representations of nodes across diverse networks, which may then be used for various downstream tasks and applications. When it applies to heterogeneous networks, it is hard to handle heterogeneous networks because they usually contain different types of nodes and edges with more semantic and structural information. Recently, contrastive learning has developed as the preferred strategy for dealing with unsupervised heterogeneous graph embedding to reduce the cost of human label annotation. However, most multi-view contrastive learning approaches calculate the model’s loss only based on the mutual dependence between the node representation and graph representation. These approaches ignore that both node attributes and node clustering contain discriminative content. To solve this issue, we propose a model called Multi-Level Discriminator-based Contrastive Learning for Multiplex Networks (MLDCL). This model adopts a multi-level multi-discriminator-based approach that can simultaneously learn the global-level structural information, node-level attribute information, and local-level clustering information. Moreover, an augmentation strategy in the contrast learning process from the spectral domain is proposed to improve the representation and discriminative ability of MLDCL. Numerous tests with node clustering and classification tasks on widely used datasets demonstrate the efficacy of the proposed approach.}, archive = {J_NEUCOM}, author = {Hongrun Wu and MingJie Zhang and Zhenglong Xiang and Yingpin Chen and Fei Yu and Xuewen Xia and Yuanxiang Li}, doi = {10.1016/j.neucom.2024.128754}, journal = {Neurocomputing}, month = {1}, pages = {128754}, shortjournal = {Neurocomputing}, title = {Multi-level discriminator based contrastive learning for multiplex networks}, volume = {613}, year = {2025}, }
• (2025). Mamba-in-mamba: Centralized mamba-cross-scan in tokenized mamba model for hyperspectral image classification. NEUCOM, 613, 128751. (www)

  @article{ , abstract = {Hyperspectral image (HSI) classification plays a crucial role in remote sensing (RS) applications, enabling the precise identification of materials and land cover based on spectral information. This supports tasks such as agricultural management and urban planning. While sequential neural models like Recurrent Neural Networks (RNNs) and Transformers have been adapted for this task, they present limitations: RNNs struggle with feature aggregation and are sensitive to noise from interfering pixels, whereas Transformers require extensive computational resources and tend to underperform when HSI datasets contain limited or unbalanced training samples. To address these challenges, Mamba architectures have emerged, offering a balance between RNNs and Transformers by leveraging lightweight, parallel scanning capabilities. Although models like Vision Mamba (ViM) and Visual Mamba (VMamba) have demonstrated improvements in visual tasks, their application to HSI classification remains underexplored, particularly in handling land-cover semantic tokens and multi-scale feature aggregation for patch-wise classifiers. In response, this study introduces the Mamba-in-Mamba (MiM) architecture for HSI classification, marking a pioneering effort in this domain. The MiM model features: (1) a novel centralized Mamba-Cross-Scan (MCS) mechanism for efficient image-to-sequence data transformation; (2) a Tokenized Mamba (T-Mamba) encoder that incorporates a Gaussian Decay Mask (GDM), Semantic Token Learner (STL), and Semantic Token Fuser (STF) for enhanced feature generation; and (3) a Weighted MCS Fusion (WMF) module with a Multi-Scale Loss Design for improved training efficiency. Experimental results on four public HSI datasets—Indian Pines, Pavia University, Houston2013, and WHU-Hi-Honghu—demonstrate that our method achieves an overall accuracy improvement of up to 3.3%, 2.7%, 1.5%, and 2.3% over state-of-the-art approaches (i.e., SSFTT, MAEST, etc.) under both fixed and disjoint training-testing settings.}, archive = {J_NEUCOM}, author = {Weilian Zhou and Sei-ichiro Kamata and Haipeng Wang and Man Sing Wong and Huiying (Cynthia) Hou}, doi = {10.1016/j.neucom.2024.128751}, journal = {Neurocomputing}, month = {1}, pages = {128751}, shortjournal = {Neurocomputing}, title = {Mamba-in-mamba: Centralized mamba-cross-scan in tokenized mamba model for hyperspectral image classification}, volume = {613}, year = {2025}, }
• (2025). Multi-attribute balanced dataset generation framework AutoSyn and KinFace channel-spatial feature extractor for kinship recognition. NEUCOM, 613, 128750. (www)

  @article{ , abstract = {In the field of kinship verification, facial recognition technology is becoming increasingly vital due to privacy concerns, ethical disputes, and the high costs associated with DNA testing. We have developed a novel method, the AutoSyn framework, to synthesize facial images and enhance kinship image datasets, effectively addressing the challenges of scale and quality in existing datasets. By employing a strategy that mixes ages and genders in the synthesized images, we minimize the impact of these factors on kinship recognition tasks. We have enhanced the original KinFaceW-I dataset by integrating ten distinct styles, including diverse combinations of gender, ethnicity, and age. This enrichment significantly improves both the quality and quantity of the images. Furthermore, this paper introduces an efficient feature extractor for kinship tasks, KinFace-CSFE, within a siamese neural network framework. This model not only utilizes meticulously designed channel feature extraction but also incorporates mixed kernel size spatial attention mechanisms to better focus on local features. We have also integrated YOCO data augmentation techniques to simulate complex imaging conditions, enhancing the model’s robustness and accuracy. The effectiveness of these innovations has been validated through experiments on the KinFaceW-I, KinFaceW-II, and synthesized Syn-KinFaceW-I datasets, achieving accuracy rates of 82.7%, 94.1%, and 83.2% respectively. These results significantly surpass both traditional models and current advanced models.}, archive = {J_NEUCOM}, author = {Jia-Xuan Jiang and Hongsheng Jing and Ling Zhou and Yuee Li and Zhong Wang}, doi = {10.1016/j.neucom.2024.128750}, journal = {Neurocomputing}, month = {1}, pages = {128750}, shortjournal = {Neurocomputing}, title = {Multi-attribute balanced dataset generation framework AutoSyn and KinFace channel-spatial feature extractor for kinship recognition}, volume = {613}, year = {2025}, }
• (2025). ThyFusion: A lightweight attribute enhancement module for thyroid nodule diagnosis using gradient and frequency-domain awareness. NEUCOM, 613, 128749. (www)

  @article{ , abstract = {Accurate classification of thyroid nodules is a critical step in clinical diagnosis and plays a crucial role in guiding subsequent treatment planning. However, current deep learning methods lack the ability to effectively perceive the attributes of thyroid nodules. Specifically, convolutional neural network-based methods struggle to capture fine-grained attributes, such as echogenic foci, due to repeated downsampling operations. On the other hand, Transformer-based methods partition the image into blocks, which hinders the capture of continuous attributes such as margins, and require a large number of parameters. To overcome these limitations, this paper proposes a novel lightweight attribute enhancement module called ThyFusion. Firstly, a multi-scale Gaussian filtering attention module is developed to accurately capture fine-grained information. Secondly, a frequency-domain global regulation module is proposed to regulate global features and capture coarse-grained attributes. Finally, a cross domain mutual learning module is presented to fuse fine-grained and coarse-grained attributes. Extensive experiments were conducted to demonstrate the effectiveness of ThyFusion, and it was compared against 15 benchmark methods. ThyFusion achieves higher accuracy, F1 score, recall, and precision (87.18%, 87.12%, 87.36% and 87.10%). ThyFusion is a plug-and-play lightweight module that can be embedded behind any layer of a backbone network to enhance the accuracy of thyroid nodule classification. The advantages of ThyFusion in thyroid nodule classification can greatly streamline the process of automated diagnosis and improve the accuracy of thyroid diagnosis.}, archive = {J_NEUCOM}, author = {Guanyuan Chen and Ningbo Zhu and Jianxin Lin and Bin Pu and Hongxia Luo and Kenli Li}, doi = {10.1016/j.neucom.2024.128749}, journal = {Neurocomputing}, month = {1}, pages = {128749}, shortjournal = {Neurocomputing}, title = {ThyFusion: A lightweight attribute enhancement module for thyroid nodule diagnosis using gradient and frequency-domain awareness}, volume = {613}, year = {2025}, }
• (2025). H3NI: Non-target-specific node injection attacks on hypergraph neural networks via genetic algorithm. NEUCOM, 613, 128746. (www)

  @article{ , abstract = {Node injection attack is widely used in graph neural networks (GNNs) attacks, which misleads GNNs by injecting nodes. Though hypergraph neural networks (HNNs) are an extension of GNNs, node injection attacks have not yet been studied in HNNs. Since each edge of a hypergraph can connect more than two nodes, existing node injection methods designed for GNNs cannot effectively select the hyperedges connected to the injected nodes when applied to hypergraphs. In this paper, we propose a H ypergraph N eural N etwork N ode I njection attack method called H3NI, which utilizes a genetic algorithm and a predefined budget model to implement the first black-box node injection framework designed for HNNs attacks. We conducted experiments on the datasets of Cora co-authorship and co-citation. Experimental results show the effectiveness and superior performance of H3NI in attacking HNNs, which reduces the model accuracy by 18%–20% within 5% of the total injected nodes and achieves a 2-4X improvement compared to existing gradient-based node injection methods.}, archive = {J_NEUCOM}, author = {Heyuan Shi and Binqi Zeng and Ruishi Yu and Yulin Yang and Zijian Zouxia and Chao Hu and Ronghua Shi}, doi = {10.1016/j.neucom.2024.128746}, journal = {Neurocomputing}, month = {1}, pages = {128746}, shortjournal = {Neurocomputing}, title = {H3NI: Non-target-specific node injection attacks on hypergraph neural networks via genetic algorithm}, volume = {613}, year = {2025}, }
• (2025). Domain adaptation for semantic segmentation of road scenes via two-stage alignment of traffic elements. NEUCOM, 613, 128744. (www)

  @article{ , abstract = {Unsupervised domain adaptation has been used to reduce the domain shift, which would improve the performance of semantic segmentation on unlabeled real-world data. However, existing methodologies fall short in effectively addressing the domain shift issue prevalent in traffic scenarios, leading to less than satisfactory segmentation results. In this paper, we propose a novel domain adaptation method for semantic segmentation via unsupervised alignment of traffic elements. Firstly, we introduce a two-stage self-training framework that leverages a blended set of training samples to enhance the training process. In the first stage, we leverage generated mixup training samples as inputs within our two-stage self-training framework and have developed corresponding loss functions for both the source and target domains to direct the training process. Then, the alignment modules for dynamic and static traffic elements are designed to achieve accurate matching between the source and the target domain images. The cosine similarity maximization is applied to the alignment of dynamic traffic elements, while the prototype learning is utilized for the static traffic elements. Additionally, we present a new technique for reducing noise in pseudo labels by constructing thresholds that adjust to each class. Meanwhile, we formulate the associated target domain loss function for vacant pseudo label pixels. The experimental results demonstrate that the proposed method is superior to the existing methods on five different domain adaptation tasks, which is more applicable to semantic segmentation of road scenes.}, archive = {J_NEUCOM}, author = {Yuan Gao and Yaochen Li and Hao Liao and Tenweng Zhang and Chao Qiu}, doi = {10.1016/j.neucom.2024.128744}, journal = {Neurocomputing}, month = {1}, pages = {128744}, shortjournal = {Neurocomputing}, title = {Domain adaptation for semantic segmentation of road scenes via two-stage alignment of traffic elements}, volume = {613}, year = {2025}, }
• (2025). 3D skeleton aware driver behavior recognition framework for autonomous driving system. NEUCOM, 613, 128743. (www)

  @article{ , abstract = {The recognition of the driver’s behaviors inside an autonomous vehicle can effectively address emergency handling in autonomous driving and is crucial for ensuring the driver’s safety. Driver behavior recognition is a challenging task due to factors such as variations, diversities, complexities, and strong interferences in behaviors. In this paper, to realize the application in the autonomous driving scenes, a novel 3D skeleton aware behavior recognition framework is proposed to recognize various driver behaviors in autonomous driving systems. First, a 3D human pose estimation network (Pose-GTFNet) with temporal Transformer and spatial graph convolutional network (GCN) is designed to infer 3D human poses from 2D pose sequences. Second, based on the obtained 3D human pose sequences, a behavior recognition network (Beh-MSFNet) with multi-skeleton feature fusion is designed to recognize driver behaviors. In the experiments, the Pose-GTFNet and Beh-MSFNet can get the best performance compared with most state-of-the-art (SOTA) methods on the Human3.6M human pose dataset, JHMDB and SHREC action recognition dataset, respectively. In addition, the proposed driver behavior recognition framework can achieve SOTA performance on the Drive&Act and Driver-Skeleton driver behavior datasets.}, archive = {J_NEUCOM}, author = {Rongtian Huo and Junkang Chen and Ye Zhang and Qing Gao}, doi = {10.1016/j.neucom.2024.128743}, journal = {Neurocomputing}, month = {1}, pages = {128743}, shortjournal = {Neurocomputing}, title = {3D skeleton aware driver behavior recognition framework for autonomous driving system}, volume = {613}, year = {2025}, }
• (2025). Fixed-time output synchronization of multilayered coupled networks with quaternion: An exponential quantized scheme. NEUCOM, 613, 128742. (www)

  @article{ , abstract = {In this article, the output synchronization of multi-layer coupled quaternion-valued neural networks with or without target state is discussed in fixed or prescribed time by developing a nonseparation approach and applying quantitative control. Firstly, considering the inaccessibility of network states and the diversity of individual functions, a type of quaternion-valued coupled networks with output coupling and multi-layer structure is introduced. In terms of fixed-time or preassigned-time controller design, an exponential quantization protocol without linear feedback is directly designed in the field of quaternion for the addressed networks with the presence of synchronous target, and a distributed quaternion-valued exponential control scheme with finite control gains is developed for the multilayered networks with the absence of synchronous target. In the process of convergence analysis, without utilizing the traditional separation method, some different forms of Lyapunov functions are directly constructed and the technique of Taylor expansion is used to derive the output synchronization criteria and the estimate of convergence time. One specific example is shown at last to verify the developed results.}, archive = {J_NEUCOM}, author = {Kailong Xiong and Cheng Hu}, doi = {10.1016/j.neucom.2024.128742}, journal = {Neurocomputing}, month = {1}, pages = {128742}, shortjournal = {Neurocomputing}, title = {Fixed-time output synchronization of multilayered coupled networks with quaternion: An exponential quantized scheme}, volume = {613}, year = {2025}, }
• (2025). A dual encoder LDCT image denoising model based on cross-scale skip connections. NEUCOM, 613, 128741. (www)

  @article{ , abstract = {LDCT image denoising is crucial in medical imaging as it aims to minimize patient radiation exposure while maintaining diagnostic image quality. However, current convolutional neural network-based denoising methods struggle to incorporate global contexts, often focusing solely on local features. This limitation poses a significant challenge. To address this, a dual encoder denoising model is introduced that utilizes the Transformer model’s proficiency in capturing long-range dependencies and global context. This model integrates the Transformer branch and the convolutional branch in the encoder. By concatenating the features of these two different branches, the model can capture both global and local image features, substantially enhancing denoising efficacy. A cross-scale skip connection mechanism is introduced to integrate the encoder’ s low-level features with the decoder’ s high-level features, enriching contextual information and preserving image details. In addition, to meet the requirements of multi-scale feature fusion, the decoder is equipped with different multi-scale convolution modules to optimize feature processing. The number of layers in these modules gradually decreases as the depth of the decoder increases. In order to enhance the discriminative ability of the model, a multi-scale discriminator is also introduced, which effectively improves the recognition ability of the image by extracting features from four different scales. Consequently, our approach demonstrates remarkable performance in reducing noise and improving LDCT image quality, as evidenced by the substantial improvements in PSNR (17.75%) and SSIM (7.31%) values.}, archive = {J_NEUCOM}, author = {Lifang Wang and Yali Wang and Wenjing Ren and Jing Yu and Xiaoyan Chang and Xiaodong Guo and Lihua Hu}, doi = {10.1016/j.neucom.2024.128741}, journal = {Neurocomputing}, month = {1}, pages = {128741}, shortjournal = {Neurocomputing}, title = {A dual encoder LDCT image denoising model based on cross-scale skip connections}, volume = {613}, year = {2025}, }
• (2025). A comprehensive review of deep learning for medical image segmentation. NEUCOM, 613, 128740. (www)

  @article{ , abstract = {Medical image segmentation provides detailed mappings of regions of interest, facilitating precise identification of critical areas and greatly aiding in the diagnosis, treatment, and understanding of diverse medical conditions. However, conventional techniques frequently rely on hand-crafted feature-based approaches, posing challenges when dealing with complex medical images, leading to issues such as low accuracy and sensitivity to noise. Recent years have seen substantial research focused on the effectiveness of deep learning models for segmenting medical images. In this study, we present a comprehensive review of the various deep learning-based approaches for medical image segmentation and provide a detailed analysis of their contributions to the domain. These methods can be broadly categorized into five groups: CNN-based methods, Transformer-based methods, Mamba-based methods, semi-supervised learning methods, and weakly supervised learning methods. Convolutional Neural Networks (CNNs), with their efficient feature self-learning, have driven major advances in medical image segmentation. Subsequently, Transformers, leveraging self-attention mechanisms, have achieved performance on par with or surpassing Convolutional Neural Networks. Mamba-based methods, as a novel selective state-space model, are emerging as a promising direction. Furthermore, due to the limited availability of annotated medical images, research in weakly supervised and semi-supervised learning continues to evolve. This review covers common evaluation methods, datasets, and deep learning applications in diagnosing and treating skin lesions, hippocampus, tumors, and polyps. Finally, we identify key challenges such as limited data, diverse modalities, noise, and clinical applicability, and propose future research in zero-shot segmentation, transfer learning, and multi-modal techniques to advance the development of medical image segmentation technology.}, archive = {J_NEUCOM}, author = {Qingling Xia and Hong Zheng and Haonan Zou and Dinghao Luo and Hongan Tang and Lingxiao Li and Bin Jiang}, doi = {10.1016/j.neucom.2024.128740}, journal = {Neurocomputing}, month = {1}, pages = {128740}, shortjournal = {Neurocomputing}, title = {A comprehensive review of deep learning for medical image segmentation}, volume = {613}, year = {2025}, }
• (2025). Learning accurate neighborhood- and self-information for higher-order relation prediction in heterogeneous information networks. NEUCOM, 613, 128739. (www)

  @article{ , abstract = {Heterogeneous Information Networks (HINs) are commonly employed to model complex real-world scenarios with diverse node and edge types. However, due to constraints in data collection and processing, constructed networks often lack certain relations. Consequently, various methods have emerged, particularly recently, leveraging heterogeneous graph neural networks (HGNNs) to predict missing relations. Nevertheless, these methods primarily focus on pairwise relations between two nodes. Real-world interactions, however, often involve multiple nodes and diverse types, extending beyond simple pairwise relations. For instance, academic collaboration networks may entail interactions among authors, papers, and conferences simultaneously. Despite their prevalence, higher-order relations are often overlooked. While HGNNs are effective at learning network structures, they may suffer from over-smoothing, resulting in similar representations for nodes and their neighbors. The learned inaccurate proximity among nodes impedes the discernment of higher-order relations. Furthermore, observed edges among a target group of nodes can provide valuable evidence for predicting higher-order relations. To address these challenges, we propose a novel model called Accurate Neighborhood- and Self-information Enhanced Heterogeneous Graph Neural Network (ANSE-HGN). Building upon HGNNs to encode network structure and attributes, we introduce a relation-based neighborhood encoder to capture information within multi-hop neighborhoods in heterogeneous higher-order relations. This enables the calculation of accurate proximity among target groups of nodes, thereby enhancing prediction accuracy. Additionally, we leverage self-information from observed higher-order relations as an auxiliary loss to reinforce the learning process. Extensive experiments on four real-world datasets demonstrate the superiority of our proposed method in higher-order relation prediction tasks.}, archive = {J_NEUCOM}, author = {Jie Li and Xuan Guo and Pengfei Jiao and Wenjun Wang}, doi = {10.1016/j.neucom.2024.128739}, journal = {Neurocomputing}, month = {1}, pages = {128739}, shortjournal = {Neurocomputing}, title = {Learning accurate neighborhood- and self-information for higher-order relation prediction in heterogeneous information networks}, volume = {613}, year = {2025}, }
• (2025). Word vector embedding and self-supplementing network for generalized few-shot semantic segmentation. NEUCOM, 613, 128737. (www)

  @article{ , abstract = {Under the condition of sufficient base class samples and a few novel class samples, Generalized Few-shot Semantic Segmentation (GFSS) classifies each pixel in the query image as base class, novel class, or background. A standard GFSS approach involves two training stages: base class learning and novel class updating. However, inter-class interference and information loss which contribute to the poor performance of GFSS, have not been synthetical considered. To address the problem, we propose an Embedded-Self-Supplementing Network (ESSNet), i.e., semantic word embedding and query set self-supplementing information to enhance segmentation accuracy. Specifically, the semantic word embedding module employs distance information between word vectors to assist the model in learning the distance between class prototypes. In order to transform the semantic word vector prototypes from the semantic space to the visual embedding space, we designed a triplet loss function to supervise the word vector embedding module, where the word vector prototype serves as an anchor and positive-negative samples are collected among the general features of the support image. To compensate for the information loss caused by using prototypes to represent classes, we propose a self-supplementing module to mine the information contained in the query image. Specifically, this module first makes a preliminary prediction on the query image, then selects high-confidence area to form pseudo labels, and finally uses pseudo labels to extract query prototypes to supplement the missing information. Extensive experiments on PASCAL-5 i and COCO-20 i show that ESSNet has superior performance and outperforms state-of-the-art methods in all settings.}, archive = {J_NEUCOM}, author = {Xiaowei Wang and Qiong Chen and Yong Yang}, doi = {10.1016/j.neucom.2024.128737}, journal = {Neurocomputing}, month = {1}, pages = {128737}, shortjournal = {Neurocomputing}, title = {Word vector embedding and self-supplementing network for generalized few-shot semantic segmentation}, volume = {613}, year = {2025}, }
• (2025). Causal-relationship representation enhanced joint extraction model for elements and relationships. NEUCOM, 613, 128736. (www)

  @article{ , abstract = {The joint extraction of elements such as entities, relations, events, and their arguments, along with their specific interrelationships, is a critical task in natural language processing. Existing research often employs implicit handling of interactions between tasks through techniques like shared encodings or parameter sharing, lacking explicit modeling of the specific relationships between tasks. This limitation hinders the full utilization of inter-task correlation information and impacts effective collaboration between tasks. To address this, we propose a model for joint extraction of elements and relations based on causal relationship representation enhancement (CRE). This model aims to capture specific relationships between tasks in multiple stages, facilitating finer adjustments and optimization of subtasks and thereby enhancing the overall model performance. Specifically, CRE comprises three key modules: feature adaptation, feature interaction, and feature fusion. The feature adaptation module selects and adjusts features from shared encodings based on the requirements of specific tasks to better adapt to semantic differences between different tasks. The feature interaction module employs causal reasoning to comprehensively capture the causal relationships between tasks while mitigating negative transfer brought about by interference from semantic information. The feature fusion module further integrates features to obtain optimized task-specific representations. Ultimately, CRE exhibits a significant improvement in average performance across multiple information extraction tasks.}, archive = {J_NEUCOM}, author = {Hang Wang and Xiaoming Liu and Guan Yang and Sensen Dong and Xingang Hu and Jie Liu}, doi = {10.1016/j.neucom.2024.128736}, journal = {Neurocomputing}, month = {1}, pages = {128736}, shortjournal = {Neurocomputing}, title = {Causal-relationship representation enhanced joint extraction model for elements and relationships}, volume = {613}, year = {2025}, }
• (2025). Answering, fast and slow: Strategy enhancement of visual understanding guided by causality. NEUCOM, 613, 128735. (www)

  @article{ , abstract = {In his classic book Thinking, Fast and Slow (Daniel, 2017), Kahneman points out that human thinking can be categorized into two main modes of thinking: a system that displays intuition and emotion (i.e., System 1), and a system that is more planned and relies more on logic, defined as System 2. This idea explains both rational and irrational motivations. In this paper, we revisit visual comprehension tasks based on this idea. At the theoretical level, we focus on the relationship between intuitive thinking, prior knowledge, and environmental information, and build a causal graph between the three. Further, inspired by the constructed causal graph, an intuitive optimization strategy with clear interpretability is proposed. In the validation session, we provide conclusions consistent with the theoretical analyses through extensive experiments on public datasets based on a visual quizzing task. Excitingly, our scheme demonstrates strong competitiveness in terms of generalizability without adding new technologies.}, archive = {J_NEUCOM}, author = {Chao Wang and Zihao Wang and Yang Zhou}, doi = {10.1016/j.neucom.2024.128735}, journal = {Neurocomputing}, month = {1}, pages = {128735}, shortjournal = {Neurocomputing}, title = {Answering, fast and slow: Strategy enhancement of visual understanding guided by causality}, volume = {613}, year = {2025}, }
• (2025). Threshold-optimized and features-fused semi-supervised domain adaptation method for rotating machinery fault diagnosis. NEUCOM, 613, 128734. (www)

  @article{ , abstract = {In the field of intelligent fault diagnosis, domain adaptation (DA) technology achieves significant breakthroughs, particularly in reducing reliance on large volumes of labeled samples. Despite these advancements, challenges persist when unlabeled data do not accurately represent actual application scenarios. Additionally, the impact of pseudo-labels on conditional domain adaptation raises concerns. To overcome the above challenges, a novel DA approach based on chaos sparrow search algorithm (CSSOA) optimized threshold parameters and feature fusion deep belief network (DBN) is proposed, named CSS-DADBN. Firstly, this method, by integrating pseudo-label updating with semi-supervised domain adaptation (SSDA) and employing confidence and entropy threshold parameters as corrective rules for pseudo-label filtering, along with the introduction of iterative conditions as an additional selection criterion, effectively alleviates the aforementioned issues. Furthermore, combining the feature extraction capabilities of DBN with a domain feature fusion strategy significantly enhances cross-domain feature learning, thereby substantially improving diagnostic accuracy. Ultimately, to validate the effectiveness and practicality of the CSS-DADBN method, a series of experiments conducted on the PT700 and Case Western Reserve University (CWRU) rolling bearing test platform clearly demonstrate its utility and efficiency in intelligent fault diagnosis.}, archive = {J_NEUCOM}, author = {Shenquan Wang and Fangyuan Zhao and Chao Cheng and Hongtian Chen and Yulian Jiang}, doi = {10.1016/j.neucom.2024.128734}, journal = {Neurocomputing}, month = {1}, pages = {128734}, shortjournal = {Neurocomputing}, title = {Threshold-optimized and features-fused semi-supervised domain adaptation method for rotating machinery fault diagnosis}, volume = {613}, year = {2025}, }
• (2025). Neighbor patches merging reduces spatial redundancy to accelerate vision transformer. NEUCOM, 613, 128733. (www)

  @article{ , abstract = {Vision Transformers (ViTs) deliver outstanding performance but often require substantial computational resources. Various token pruning methods have been developed to enhance throughput by removing redundant tokens; however, these methods do not address the peak memory consumption, which remains equivalent to that of the unpruned networks. In this study, we introduce Neighbor Patches Merging (NEPAM), a method that significantly reduces the maximum memory footprint of ViTs while pruning tokens. NEPAM targets spatial redundancy within images and prunes redundant patches at the onset of the model, thereby achieving the optimal throughput-accuracy trade-off without fine-tuning. Experimental results demonstrate that NEPAM can accelerate the inference speed of the Vit-Base-Patch16-384 model by 25% with a negligible accuracy loss of 0.07% and a notable 18% reduction in memory usage. When applied to VideoMAE, NEPAM doubles the throughput with a 0.29% accuracy loss and a 48% reduction in memory usage. These findings underscore the efficacy of NEPAM in mitigating computational requirements while maintaining model performance.}, archive = {J_NEUCOM}, author = {Kai Jiang and Peng Peng and Youzao Lian and Weihui Shao and Weisheng Xu}, doi = {10.1016/j.neucom.2024.128733}, journal = {Neurocomputing}, month = {1}, pages = {128733}, shortjournal = {Neurocomputing}, title = {Neighbor patches merging reduces spatial redundancy to accelerate vision transformer}, volume = {613}, year = {2025}, }
• (2025). Symbolic equation solving via reinforcement learning. NEUCOM, 613, 128732. (www)

  @article{ , abstract = {Machine-learning methods are rapidly being adopted in a wide variety of social, economic, and scientific contexts, yet they are notorious for struggling with exact mathematics. A typical example is computer algebra, which includes tasks like simplifying mathematical terms, calculating formal derivatives, or finding exact solutions of algebraic equations. Traditional software packages for these purposes are commonly based on a huge database of rules for how a specific operation (e.g., differentiation) transforms a certain term (e.g., sine function) into another one (e.g., cosine function). These rules have usually needed to be discovered and subsequently programmed by humans. Efforts to automate this process by machine-learning approaches are faced with challenges like the singular nature of solutions to mathematical problems, when approximations are unacceptable, as well as hallucination effects leading to flawed reasoning. We propose a novel deep-learning interface involving a reinforcement-learning agent that operates a symbolic stack calculator to explore mathematical relations. By construction, this system is capable of exact transformations and immune to hallucination. Using the paradigmatic example of solving linear equations in symbolic form, we demonstrate how our reinforcement-learning agent autonomously discovers elementary transformation rules and step-by-step solutions.}, archive = {J_NEUCOM}, author = {Lennart Dabelow and Masahito Ueda}, doi = {10.1016/j.neucom.2024.128732}, journal = {Neurocomputing}, month = {1}, pages = {128732}, shortjournal = {Neurocomputing}, title = {Symbolic equation solving via reinforcement learning}, volume = {613}, year = {2025}, }
• (2025). VG-CALF: A vision-guided cross-attention and late-fusion network for radiology images in medical visual question answering. NEUCOM, 613, 128730. (www)

  @article{ , abstract = {Image and question matching is essential in Medical Visual Question Answering (MVQA) in order to accurately assess the visual-semantic correspondence between an image and a question. However, the recent state-of-the-art methods focus solely on the contrastive learning between an entire image and a question. Though contrastive learning successfully model the global relationship between an image and a question, it is less effective to capture the fine-grained alignments conveyed between image regions and question words. In contrast, large-scale pre-training poses significant drawbacks, including extended training times, handling substantial data volumes, and necessitating high computational power. To address these challenges, we propose the Vision-Guided Cross-Attention based Late Fusion (VG-CALF) network, which integrates image and question features into a unified deep model without relying on pre-training for MVQA tasks. In our proposed approach, we use self-attention to effectively leverage intra-modal relationships within each modality and implement vision-guided cross-attention to emphasize the inter-modal relationships between image regions and question words. By simultaneously considering intra-modal and inter-modal relationships, our proposed method significantly improves the overall performance of MVQA without the need for pre-training on extensive image-question pairs. Experimental results on benchmark datasets, such as, SLAKE and VQA-RAD demonstrate that our proposed approach performs competitively with existing state-of-the-art methods.}, archive = {J_NEUCOM}, author = {Aiman Lameesa and Chaklam Silpasuwanchai and Md. Sakib Bin Alam}, doi = {10.1016/j.neucom.2024.128730}, journal = {Neurocomputing}, month = {1}, pages = {128730}, shortjournal = {Neurocomputing}, title = {VG-CALF: A vision-guided cross-attention and late-fusion network for radiology images in medical visual question answering}, volume = {613}, year = {2025}, }
• (2025). Synchronization of nonlinear neural networks with hybrid couplings and uncertain time-varying perturbations: A novel distributed-delay impulsive comparison principle. NEUCOM, 613, 128729. (www)

  @article{ , abstract = {This paper investigates the synchronization of nonlinear drive-response neural networks subject to uncertain time-varying perturbations, non-delayed coupling, and distributed delay coupling. To address the influence of distributed and discrete delays on the system, we establish a novel impulsive comparison principle, extending the Halanay inequality. By leveraging Lyapunov stability theory, we derive sufficient conditions for the exponential synchronization of the neural networks using a delayed impulsive controller with historical status information. This approach relaxes the conventional constraint that impulsive delays must be smaller than impulsive intervals, thereby generalizing existing synchronization results for distributed delay networks. Numerical simulations for chaotic neural networks validate the theoretical results and demonstrate the sensitivity of the control gain matrix.}, archive = {J_NEUCOM}, author = {Hongguang Fan and Kaibo Shi and Yi Zhao}, doi = {10.1016/j.neucom.2024.128729}, journal = {Neurocomputing}, month = {1}, pages = {128729}, shortjournal = {Neurocomputing}, title = {Synchronization of nonlinear neural networks with hybrid couplings and uncertain time-varying perturbations: A novel distributed-delay impulsive comparison principle}, volume = {613}, year = {2025}, }
• (2025). SF-GPT: A training-free method to enhance capabilities for knowledge graph construction in LLMs. NEUCOM, 613, 128726. (www)

  @article{ , abstract = {Knowledge graphs (KGs) are constructed by extracting knowledge triples from text and fusing knowledge, enhancing information retrieval efficiency. Current methods for knowledge triple extraction include ”Pretrain and Fine-tuning” and Large Language Models (LLMs). The former shifts effort from manual extraction to dataset annotation and suffers from performance degradation with different test and training set distributions. LLMs-based methods face errors and incompleteness in extraction. We introduce SF-GPT, a training-free method to address these issues. Firstly, we propose the Entity Extraction Filter (EEF) module to filter triple generation results, addressing evaluation and cleansing challenges. Secondly, we introduce a training-free Entity Alignment Module based on Entity Alias Generation (EAG), tackling semantic richness and interpretability issues in LLM-based knowledge fusion. Finally, our Self-Fusion Subgraph strategy uses multi-response self-fusion and a common entity list to filter triple results, reducing noise from LLMs’ multi-responses. In experiments, SF-GPT showed a 55.5% increase in recall and a 32.6% increase in F1 score on the BDNC dataset compared to the UniRel model trained on the NYT dataset and achieved a 5% improvement in F1 score compared to GPT-4+EEF baseline on the WebNLG dataset in the case of a fusion round of three. SF-GPT offers a promising way to extract knowledge from unstructured information.}, archive = {J_NEUCOM}, author = {Lizhuang Sun and Peng Zhang and Fang Gao and Yuan An and Zhixing Li and Yuanwei Zhao}, doi = {10.1016/j.neucom.2024.128726}, journal = {Neurocomputing}, month = {1}, pages = {128726}, shortjournal = {Neurocomputing}, title = {SF-GPT: A training-free method to enhance capabilities for knowledge graph construction in LLMs}, volume = {613}, year = {2025}, }
• (2025). Practical consensus of T-S fuzzy positive multi-agent systems using linear programming. NEUCOM, 613, 128725. (www)

  @article{ , abstract = {This paper presents the practical consensus of T-S fuzzy positive multi-agent systems with interval type-1 and type-2 fuzzy sets using observer-based control protocols. A fuzzy positive observer is first constructed for the systems. An observer-based fuzzy control protocol is designed, where an additional constant term is introduced. Some linear programming conditions are established to achieve the positivity of the original system and its observer. Then, the practical consensus of the original system is transformed into the stability of a dynamic system, where a set of new variables is defined by combining the constant term and the states of the agents. In the first step, the positivity of the new variables is guaranteed. In the second step, the stability of the dynamic system consisting new variables is addressed by using a fuzzy copositive Lyapunov function. The gain matrices of observer and control protocols are formulated based on a matrix decomposition approach. All positive and consensus conditions are described via linear programming. Finally, two examples are provided to verify the validity of the obtained results.}, archive = {J_NEUCOM}, author = {Junfeng Zhang and Chongxiang Yu and Baochen Zhang and Weidong Zhang}, doi = {10.1016/j.neucom.2024.128725}, journal = {Neurocomputing}, month = {1}, pages = {128725}, shortjournal = {Neurocomputing}, title = {Practical consensus of T-S fuzzy positive multi-agent systems using linear programming}, volume = {613}, year = {2025}, }
• (2025). Learning differentiable categorical regions with gumbel-softmax for person re-identification. NEUCOM, 613, 128723. (www)

  @article{ , abstract = {Locating diverse body parts and perceiving part visibility are essential to person re-identification (re-ID). Most existing methods employ an extra model, e.g. , pose estimation or human parsing, to locate parts, or generate pseudo labels to train the part locator incorporated with the re-ID model. In this paper, we aim at learning diverse horizontal stripes with foreground refinement to pursue pixel-level part alignment via only using person identity labels. Specifically, we proposed a Gumbel-Softmax based Differential Categorical Region (DCR) learning method and make two contributions. (1) A stripe-wise regularization. Given an image, the part locator produce part probability maps. The continuous values in the probability maps are discretized into zero or arg max value in the horizontal stripes by the Gumbel-Softmax. Gumbel-Softmax allows us to use the arg max discrete value for part diversity regularization in the forward pass, but can still estimate gradients in the backward pass. (2) A self-refinement method to suppress the background noise in the stripes. We employ a lightweight foreground perception head to produce foreground probability map with only person identity labels supervision. Benefits from discretization of the categorical stripes, we can conveniently obtain the part pseudo label by element-wise multiplying the categorical stripes with foreground probability map. Finally, DCR can locate the body parts at pixel-level and extract part-aligned representation. Experimental results on both holistic and occluded re-ID datasets confirm that our approach significantly improves the learned representation and the achieved performance is on par with the state-of-the-art methods. The code is available at https://github.com/deepalchemist/differentiable-categorical-region}, archive = {J_NEUCOM}, author = {Wenjie Yang and Pei Xu}, doi = {10.1016/j.neucom.2024.128723}, journal = {Neurocomputing}, month = {1}, pages = {128723}, shortjournal = {Neurocomputing}, title = {Learning differentiable categorical regions with gumbel-softmax for person re-identification}, volume = {613}, year = {2025}, }
• (2025). Research on application of knowledge graph in industrial control system security situation awareness and decision-making: A survey. NEUCOM, 613, 128721. (www)

  @article{ , abstract = {Knowledge graph as a powerful tool for knowledge organization and representation, can integrate scattered data into a unified knowledge network, enabling knowledge association and knowledge reasoning, thus improving the accuracy of security situation awareness. In the context of industrial control system security decision-making, knowledge graphs can provide comprehensive knowledge support, assisting decision-makers in making intelligent decisions. This paper reviews the current research status of knowledge graphs in the field of industrial control system security situation awareness and decision-making, covering data-driven techniques, rule-based methods, and knowledge graph-based approaches. Existing knowledge graph technologies and their practical applications in industrial control system security situation awareness and decision-making are discussed. Finally, a series of challenges in the application of knowledge graphs in industrial control system security situation awareness and decision are summarized, such as data dispersion, correlation and information security, and the future prospects.}, archive = {J_NEUCOM}, author = {Lixin Liu and Peihang Xu and Kefeng Fan and Mingyan Wang}, doi = {10.1016/j.neucom.2024.128721}, journal = {Neurocomputing}, month = {1}, pages = {128721}, shortjournal = {Neurocomputing}, title = {Research on application of knowledge graph in industrial control system security situation awareness and decision-making: A survey}, volume = {613}, year = {2025}, }
• (2025). Integrating reinforcement learning and supervisory control theory for optimal directed control of discrete-event systems. NEUCOM, 613, 128720. (www)

  @article{ , abstract = {Directed control is crucial for implementing controllers on programmable logic controllers. A directed controller is one that actively selects at most one controllable event (control command) to execute at any instant. This study investigates the numerical optimization problem of directed control for discrete-event systems, such as traffic systems and robotic systems. By integrating supervisory control theory and reinforcement learning, an algorithmic procedure is designed to synthesize an optimal directed controller, which minimizes the total cost associated with event execution while ensuring the safety and liveness of the controlled system. Two improved Q-learning algorithms incorporating dynamic parameter adaptation strategies are developed to enhance the global search ability of basic Q-learning. To demonstrate applicability and effectiveness, we apply the proposed method to control a guideway system and a multi-train traffic system, respectively. The experimental results indicate the superiority of the proposed method over the comparison methods.}, archive = {J_NEUCOM}, author = {Yuhong Hu and Deguang Wang and Ming Yang and Jiahan He}, doi = {10.1016/j.neucom.2024.128720}, journal = {Neurocomputing}, month = {1}, pages = {128720}, shortjournal = {Neurocomputing}, title = {Integrating reinforcement learning and supervisory control theory for optimal directed control of discrete-event systems}, volume = {613}, year = {2025}, }
• (2025). RI-L1Approx: A novel resnet-inception-based fast l1-approximation method for face recognition. NEUCOM, 613, 128708. (www)

  @article{ , abstract = {Performance of deep learning methods for face recognition often relies on abundant data, posing challenges in surveillance and security where data availability is limited and environments are unconstrained. To address this challenge, we propose a novel few shot learning approach termed Resnet Inception-based Fast L 1 approximation ( R I - L 1 Approx) for face recognition with limited number of image samples. The method relies on L 1 norm approximation of test sample with known class samples. Initially, facial features are extracted by leveraging ResNet-Inception hybrid network’s abilities to learn rich hierarchical representations from facial images. The extracted features are subsequently employed for L 1 norm approximation of known features, which are referred to as approximation samples. The L 1 norm approximation promotes sparsity by encouraging a subset of approximation samples to possess zero coefficients. This process helps in selecting the most discriminative and informative approximation samples, leading to improved classification capabilities. The proposed method is evaluated on benchmark facial recognition datasets, demonstrating its effectiveness. Comparative experiments with state-of-the-art techniques highlight its superior recognition accuracy. Remarkably, the R I - L 1 Approx model achieved high accuracy rates of 84.86% with just one sample per class and 96.144% with thirteen samples per class during experimentation. This is significantly better than existing deep learning approaches, which require a large amount of data to train the model for similar performance.}, archive = {J_NEUCOM}, author = {Supriya Bajpai and Gargi Mishra and Rachna Jain and Deepak Kumar Jain and Dharmender Saini and Amir Hussain}, doi = {10.1016/j.neucom.2024.128708}, journal = {Neurocomputing}, month = {1}, pages = {128708}, shortjournal = {Neurocomputing}, title = {RI-L1Approx: A novel resnet-inception-based fast l1-approximation method for face recognition}, volume = {613}, year = {2025}, }
• (2025). Improving rule-based classifiers by bayes point aggregation. NEUCOM, 613, 128699. (www)

  @article{ , abstract = {The widespread adoption of artificial intelligence systems with continuously higher capabilities is causing ethical concerns. The lack of transparency, particularly for state-of-the-art models such as deep neural networks, hinders the applicability of such black-box methods in many domains, like the medical or the financial ones, where model transparency is a mandatory requirement, and hence white-box models are largely preferred over potentially more accurate but opaque techniques. For this reason, in this paper, we focus on ruleset learning, arguably the most interpretable class of learning techniques. Specifically, we propose Bayes Point Rule Classifier, an ensemble methodology inspired by the Bayes Point Machine, to improve the performance and robustness of rule-based classifiers. In addition, to improve interpretability, we propose a technique to retain the most relevant rules based on their importance, thus increasing the transparency of the ensemble, making it easier to understand its decision-making process. We also propose FIND-RS, a greedy ruleset learning algorithm that, under mild conditions, guarantees to learn hypothesis with perfect accuracy on the training set while preserving a good generalization capability to unseen data points. We performed extensive experimentation showing that FIND-RS achieves state-of-the-art classification performance at the cost of a slight increase in the ruleset complexity w.r.t. the competitors. However, when paired with the Bayes Point Rule Classifier, FIND-RS outperforms all the considered baselines.}, archive = {J_NEUCOM}, author = {Luca Bergamin and Mirko Polato and Fabio Aiolli}, doi = {10.1016/j.neucom.2024.128699}, journal = {Neurocomputing}, month = {1}, pages = {128699}, shortjournal = {Neurocomputing}, title = {Improving rule-based classifiers by bayes point aggregation}, volume = {613}, year = {2025}, }
• (2025). A lightweight video anomaly detection model with weak supervision and adaptive instance selection. NEUCOM, 613, 128698. (www)

  @article{ , abstract = {Video anomaly detection is to determine whether there are any abnormal events, behaviors or objects in a given video, which enables effective and intelligent public safety management. As video anomaly labeling is both time-consuming and expensive, most existing works employ unsupervised or weakly supervised learning methods. This paper focuses on weakly supervised video anomaly detection, in which the training videos are labeled whether or not they contain any anomalies, but lack information about the specific frames and quantities of anomalies. However, the uncertainty of weakly labeled data and the large model size prevent existing methods from wide deployment in real scenarios, especially the resource-limit situations such as edge-computing. In this paper, we develop a lightweight video anomaly detection model. On the one hand, we propose an adaptive instance selection strategy, which is based on the model’s current status to select confident instances, thereby mitigating the uncertainty of weakly labeled data and subsequently promoting the model’s performance. On the other hand, we design a lightweight multi-level temporal correlation attention module and an hourglass-shaped fully connected layer to construct the model, which can reduce the model parameters to only 0.56% of the existing methods (e.g. RTFM). Extensive experiments on three public datasets UCF-Crime, ShanghaiTech and XD-Violence show that our model performs better than or equally to the existing lightweight methods, while with a significantly reduced number of model parameters. Furthermore, by integrating the improved module designed in this paper with the VadCLIP method proposed by Wu et al., we achieve the state-of-the-art performance of non-lightweight models on the UCF-Crime and XD-Violence datasets.}, archive = {J_NEUCOM}, author = {Yang Wang and Jiaogen Zhou and Jihong Guan}, doi = {10.1016/j.neucom.2024.128698}, journal = {Neurocomputing}, month = {1}, pages = {128698}, shortjournal = {Neurocomputing}, title = {A lightweight video anomaly detection model with weak supervision and adaptive instance selection}, volume = {613}, year = {2025}, }
• (2025). Adept: Annotation-denoising auxiliary tasks with discrete cosine transform map and keypoint for human-centric pretraining. NEUCOM, 613, 128694. (www)

  @article{ , abstract = {Human-centric perception is the core of diverse computer vision tasks and has been a long-standing research focus. However, previous research studied these human-centric tasks individually, whose performance is largely limited to the size of the public task-specific datasets. Recent human-centric methods leverage the additional modalities, e.g., depth, to learn fine-grained semantic information, which limits the benefit of pretraining models due to their sensitivity to camera views and the scarcity of RGB-D data on the Internet. This paper improves the data scalability of human-centric pretraining methods by discarding depth information and exploring semantic information of RGB images in the frequency space by Discrete Cosine Transform (DCT). We further propose new annotation denoising auxiliary tasks with keypoints and DCT maps to enforce the RGB image extractor to learn fine-grained semantic information of human bodies. Our extensive experiments show that when pretrained on large-scale datasets (COCO and AIC datasets) without depth annotation, our model achieves better performance than state-of-the-art methods by +0.5 mAP ( ↑ ) on COCO, +1.4 PCKh ( ↑ ) on MPII and -0.51 EPE ( ↓ ) on Human3.6M for pose estimation, by +4.50 mIoU ( ↑ ) on Human3.6M for human parsing, by -3.14 MAE ( ↓ ) on SHA and -0.07 MAE ( ↓ ) on SHB for crowd counting, by +1.1 F1 score ( ↑ ) on SHA and +0.8 F1 score ( ↑ ) on SHA for crowd localization, and by +0.1 mAP ( ↑ ) on Market1501 and +0.8 mAP ( ↑ ) on MSMT for person ReID. We also validate the effectiveness of our method on MPII+NTURGBD datasets.}, archive = {J_NEUCOM}, author = {Weizhen He and Yunfeng Yan and Shixiang Tang and Yiheng Deng and Yangyang Zhong and Pengxin Luo and Donglian Qi}, doi = {10.1016/j.neucom.2024.128694}, journal = {Neurocomputing}, month = {1}, pages = {128694}, shortjournal = {Neurocomputing}, title = {Adept: Annotation-denoising auxiliary tasks with discrete cosine transform map and keypoint for human-centric pretraining}, volume = {613}, year = {2025}, }
• (2025). FedATA: Adaptive attention aggregation for federated self-supervised medical image segmentation. NEUCOM, 613, 128691. (www)

  @article{ , abstract = {Pre-trained on large-scale datasets has profoundly promoted the development of deep learning models in medical image analysis. For medical image segmentation, collecting a large number of labeled volumetric medical images from multiple institutions is an enormous challenge due to privacy concerns. Self-supervised learning with mask image modeling (MIM) can learn general representation without annotations. Integrating MIM into FL enables collaborative learning of an efficient pre-trained model from unlabeled data, followed by fine-tuning with limited annotations. However, setting pixels as reconstruction targets in traditional MIM fails to facilitate robust representation learning due to the medical image's complexity and distinct characteristics. On the other hand, the generalization of the aggregated model in FL is also impaired under the heterogeneous data distributions among institutions. To address these issues, we proposed a novel self-supervised federated learning, which combines masked self-distillation with adaptive attention federated learning. Such incorporation enjoys two vital benefits. First, masked self-distillation sets high-quality latent representations of masked tokens as the target, improving the descriptive capability of the learned presentation rather than reconstructing low-level pixels. Second, adaptive attention aggregation with Personalized federate learning effectively captures specific-related representation from the aggregated model, thus facilitating local fine-tuning performance for target tasks. We conducted comprehensive experiments on two medical segmentation tasks using a large-scale dataset consisting of volumetric medical images from multiple institutions, demonstrating superior performance compared to existing federated self-supervised learning approaches.}, archive = {J_NEUCOM}, author = {Jian Dai and Hao Wu and Huan Liu and Liheng Yu and Xing Hu and Xiao Liu and Daoying Geng}, doi = {10.1016/j.neucom.2024.128691}, journal = {Neurocomputing}, month = {1}, pages = {128691}, shortjournal = {Neurocomputing}, title = {FedATA: Adaptive attention aggregation for federated self-supervised medical image segmentation}, volume = {613}, year = {2025}, }
• (2025). Language-guided temporal primitive modeling for skeleton-based action recognition. NEUCOM, 613, 128636. (www)

  @article{ , abstract = {Human actions describe the complex body dynamics, which are defined by an ordered set of sub-action primitives. Existing methods in skeleton-based action recognition primarily focus on designing various networks to learn the entire action representation, in alignment with the action label. However, training the heterogeneous action as a whole increases the burden on the network and is not conducive to learning the temporal structure of the action. In this paper, the traditional action label is extended into the ordered primitive language description by employing a large language model, which not only unveils the temporal structure of the action but also enriches its connotation. Based on this, we propose a language-guided skeleton representation learning network (LGS-Net) that leverages the language description to guide the skeleton feature learning from the ordered primitive and global action levels. Especially, the primitive guidance aligns the features of skeleton clips with those of the ordered primitive descriptions while preserving temporal order, which promotes the network to model the temporal primitive and capture the internal structure within the action from skeletons. To enhance the cross-modal alignment performance, we develop an innovative temporal alignment loss supplemented with diversity and sparsity regularization terms to generate the discriminative multi-modal representation. Evaluated on three benchmark datasets, NTU-60, NTU-120 and N-UCLA, the proposed LGS-Net achieves comparable results to state-of-the-art methods, which proves the effectiveness of the language-guided learning mechanism.}, archive = {J_NEUCOM}, author = {Qingzhe Pan and Xuemei Xie}, doi = {10.1016/j.neucom.2024.128636}, journal = {Neurocomputing}, month = {1}, pages = {128636}, shortjournal = {Neurocomputing}, title = {Language-guided temporal primitive modeling for skeleton-based action recognition}, volume = {613}, year = {2025}, }
• (2025). Hierarchical collaboration for referring image segmentation. NEUCOM, 613, 128632. (www)

  @article{ , abstract = {In the field of referring segmentation, top-down methods and bottom-up methods are the two prevailing approaches. Both of these methods inevitably exhibit certain drawbacks. Top-down methods are susceptible to Polar Negative (PN) errors due to their limited understanding of multi-modal fine-grained features. Bottom-up methods lack macro-level object positional information, making them susceptible to Inferior Positive (IP) errors. However, we find that the two approaches are highly complementary in addressing their respective weaknesses, but combining them directly through a simple average does not yield complementary advantages. Therefore, we proposed a hierarchical collaboration approach to explore the complementary characteristics of the existing two methods from the perspectives of fusion and interaction, aiming to achieve more precise segmentation results. We proposed the Complementary Feature Interaction (CFI) module, which enables top-down methods to access fine-grained information and allows bottom-up approaches to obtain object positional information interactively. Regarding integration, Gaussian Scoring Integration (GSI) models the Gaussian performance distributions of two branches and performs weighted integration by sampling confidence scores from these distributions. We integrate various top-down and bottom-up methods within the proposed architecture and conduct experiments on three standard datasets. The experimental results demonstrate that our method outperforms the state-of-the-art independent segmentation algorithms. On the RefCOCO validation, test A and test B datasets, our proposed method achieved IoU scores of 77.51, 79.12, and 72.79, respectively. Extensive experiments demonstrate that our method can significantly improve segmentation accuracy when fusing different sub-methods.}, archive = {J_NEUCOM}, author = {Wei Zhang and Zesen Cheng and Jie Chen and Wen Gao}, doi = {10.1016/j.neucom.2024.128632}, journal = {Neurocomputing}, month = {1}, pages = {128632}, shortjournal = {Neurocomputing}, title = {Hierarchical collaboration for referring image segmentation}, volume = {613}, year = {2025}, }
• (2025). LAAP: Learning the argument of an entity with event prompts for document-level event extraction. NEUCOM, 613, 128584. (www)

  @article{ , abstract = {Document-level Event Extraction (DEE) aims to identify event types within a document and extract their corresponding arguments, which is essential for structured information provision in various NLP applications. Unlike sentence-level extraction, DEE requires handling events and arguments scattered across a document. Existing methods often focus on intricate feature interactions, neglecting explicit argument–entity relationships. We introduce a novel method, Learning the Argument of an Entity with Event Prompts (LAAP), which constructs event prompts for type detection, incorporating sentence placeholders to elicit event-specific information. Additionally, we propose an entity argument learning strategy that narrows down entity types to find the most suitable one. Experiments on the ChFinAnn and three other public datasets show that our method surpasses state-of-the-art approaches in accuracy and effectiveness.}, archive = {J_NEUCOM}, author = {Jinghan Xu and Cheng Yang and Xiaojun Kang}, doi = {10.1016/j.neucom.2024.128584}, journal = {Neurocomputing}, month = {1}, pages = {128584}, shortjournal = {Neurocomputing}, title = {LAAP: Learning the argument of an entity with event prompts for document-level event extraction}, volume = {613}, year = {2025}, }
• (2025). A diverse knowledge perception and fusion network for detecting targets and key parts in UAV images. NEUCOM, 612, 128748. (www)

  @article{ , abstract = {Detecting targets and their key parts in UAV images is crucial for both military and civilian applications, including optimizing damage assessment, evaluating infrastructure, and facilitating disaster response efforts. Traditional top-down approaches impose excessive constraints that struggle to address challenges such as variable definitions and quantities of key parts, potential target occlusion, and model redundancy. Conversely, end-to-end approaches often overlook the relationships between targets and key parts, resulting in low detection accuracy. Inspired by the remarkable human reasoning process, we propose the Diverse Knowledge Perception and Fusion (DKPF) network, which skillfully balances the trade-offs between stringent constraints and unconstrained methods while ensuring both detection precision and real-time performance. Specifically, our model integrates reasoning guided by three distinct forms of knowledge: contextual knowledge at the image level in an unsupervised manner; explicit semantic knowledge regarding the interactions between targets and key parts at the instance level; and implicit comprehensive knowledge about the relationships among different types of targets or key parts, such as shape similarity. These specific knowledge forms are extracted through a novel adaptive fusion strategy for multi-scale features, a binary region-to-region semantic knowledge graph, and a data-driven self-attention architecture, respectively. Experiments conducted on both simulated and real-world datasets reveal that our method significantly outperforms state-of-the-art techniques, regardless of the number of key parts in the target. Furthermore, extensive ablation studies and visualization analyses validate both the efficacy of our approach and the interpretability of the generated features.}, archive = {J_NEUCOM}, author = {Hanyu Wang and Qiang Shen and Zilong Deng}, doi = {10.1016/j.neucom.2024.128748}, journal = {Neurocomputing}, month = {1}, pages = {128748}, shortjournal = {Neurocomputing}, title = {A diverse knowledge perception and fusion network for detecting targets and key parts in UAV images}, volume = {612}, year = {2025}, }
• (2025). UCC: A unified cascade compression framework for vision transformer models. NEUCOM, 612, 128747. (www)

  @article{ , abstract = {In recent years, Vision Transformer (ViT) and its variants have dominated many computer vision tasks. However, the high computational consumption and training data requirements of ViT make it challenging to be deployed directly on resource-constrained devices and environments. Model compression is an effective approach to accelerate deep learning networks, but existing methods for compressing ViT models are limited in their scopes and struggle to strike a balance between performance and computational cost. In this paper, we propose a novel Unified Cascaded Compression Framework (UCC) to compress ViT in a more precise and efficient manner. Specifically, we first analyze the frequency information within tokens and prune them based on a joint score of their both spatial and spectral characteristics. Subsequently, we propose a similarity-based token aggregation scheme that combines the abundant contextual information contained in all pruned tokens with the host tokens according to their weights. Additionally, we introduce a novel cumulative cascaded pruning strategy that performs bottom-up cascaded pruning of tokens based on cumulative scores, avoiding information loss caused by individual idiosyncrasies of blocks. Finally, we design a novel two-level distillation strategy, incorporating imitation and exploration, to ensure the diversity of knowledge and better performance recovery. Extensive experiments demonstrate that our unified cascaded compression framework outperforms most existing state-of-the-art approaches, compresses the floating-point operations of ViT-Base as well as DeiT-Base models by 22 % and 54.1 %, and improves the recognition accuracy of the models by 3.74 % and 1.89 %, respectively, significantly reducing model computational consumption while enhancing performance, which enables efficient end-to-end training of compact ViT models.}, archive = {J_NEUCOM}, author = {Dingfu Chen and Kangwei Lin and Qingxu Deng}, doi = {10.1016/j.neucom.2024.128747}, journal = {Neurocomputing}, month = {1}, pages = {128747}, shortjournal = {Neurocomputing}, title = {UCC: A unified cascade compression framework for vision transformer models}, volume = {612}, year = {2025}, }
• (2025). Audio feature enhancement based on quaternion filtering and deep hashing. NEUCOM, 612, 128727. (www)

  @article{ , abstract = {This paper aims to solve the problems of difficult convergence of audio model training, large data demand, and large dimensionality of storage space for audio-generated feature vectors. To this end, this paper proposes the use of quaternion Gabor filtering to suppress the background information of the spectrogram and reduce the interference of the data for the case of insufficient data alignment between audio data and image data after shifting the domain. In addition, different scales of window lengths and frame shifts are used to capture the connections between different vocal objects. To address the problem that the generated feature vectors are large dimensional, we use a deep hash module to map high-dimensional features to low-dimensional features and use a probability function to make the learned samples more consistent with the overall distribution. In the experimental evaluation, the proposed method was evaluated on the environmental sound classification dataset and the music genre classification dataset. The proposed method uses only a common backbone network and improves the accuracy of audio recognition.}, archive = {J_NEUCOM}, author = {Xun Jin and Bingkui Sun and De Li}, doi = {10.1016/j.neucom.2024.128727}, journal = {Neurocomputing}, month = {1}, pages = {128727}, shortjournal = {Neurocomputing}, title = {Audio feature enhancement based on quaternion filtering and deep hashing}, volume = {612}, year = {2025}, }
• (2025). A novel ensemble over-sampling approach based chebyshev inequality for imbalanced multi-label data. NEUCOM, 612, 128717. (www)

  @article{ , abstract = {With the development of intelligent technology, data exhibits characteristics of multi-label and imbalanced distribution, which lead to the degradation of classification model performance. Therefore, addressing multi-label class imbalance has become a hot research topic. Nowadays, over-sampling approaches aim to generate a superset of the original dataset to deal with imbalanced data. However, traditional over-sampling methods only employ the central data point and its nearest neighbor samples to synthesize samples without considering the impact of data distribution. To address these issues, in this paper, we propose an ensemble multi-label over-sampling algorithm (MLCIO) based on Chebyshev inequality and a group optimization strategy. Firstly, to generate more representative and diverse samples, with the seed sample serving as the sphere’s center, Chebyshev inequality is utilized to ensure that synthetic samples fall within its m times the standard deviation. Secondly, a group optimization ranking weighting approach is employed to obtain more reliable and stable label information. Finally, comparative experiments are conducted on 11 imbalanced datasets from various domains using different evaluation metrics. The results demonstrate that our proposal achieves better performance than other approaches.}, archive = {J_NEUCOM}, author = {Weishuo Ren and Yifeng Zheng and Wenjie Zhang and Depeng Qing and Xianlong Zeng and Guohe Li}, doi = {10.1016/j.neucom.2024.128717}, journal = {Neurocomputing}, month = {1}, pages = {128717}, shortjournal = {Neurocomputing}, title = {A novel ensemble over-sampling approach based chebyshev inequality for imbalanced multi-label data}, volume = {612}, year = {2025}, }
• (2025). Artificial immunofluorescence in a flash: Rapid synthetic imaging from brightfield through residual diffusion. NEUCOM, 612, 128715. (www)

  @article{ , abstract = {Immunofluorescent (IF) imaging is crucial for visualising biomarker expressions, cell morphology and assessing the effects of drug treatments on sub-cellular components. IF imaging needs extra staining process and often requiring cell fixation, therefore it may also introduce artefacts and alter endogenous cell morphology. Some IF stains are expensive or not readily available hence hindering experiments. Recent diffusion models, which synthesise high-fidelity IF images from easy-to-acquire brightfield (BF) images, offer a promising solution but are hindered by training instability and slow inference times due to the noise diffusion process. This paper presents a novel method for the conditional synthesis of IF images directly from BF images along with cell segmentation masks. Our approach employs a Residual Diffusion process that enhances stability and significantly reduces inference time. We performed a critical evaluation against other image-to-image synthesis models, including UNets, GANs, and advanced diffusion models. Our model demonstrates significant improvements in image quality ( p < 0 . 05 in MSE, PSNR, and SSIM), inference speed (26 times faster than competing diffusion models), and accurate segmentation results for both nuclei and cell bodies (0.77 and 0.63 mean IOU for nuclei and cell true positives, respectively). This paper is a substantial advancement in the field, providing robust and efficient tools for cell image analysis.}, archive = {J_NEUCOM}, author = {Xiaodan Xing and Chunling Tang and Siofra Murdoch and Giorgos Papanastasiou and Yunzhe Guo and Xianglu Xiao and Jan Cross-Zamirski and Carola-Bibiane Schönlieb and Kristina Xiao Liang and Zhangming Niu and Evandro Fei Fang and Yinhai Wang and Guang Yang}, doi = {10.1016/j.neucom.2024.128715}, journal = {Neurocomputing}, month = {1}, pages = {128715}, shortjournal = {Neurocomputing}, title = {Artificial immunofluorescence in a flash: Rapid synthetic imaging from brightfield through residual diffusion}, volume = {612}, year = {2025}, }
• (2025). Deep learning-based image encryption techniques: Fundamentals, current trends, challenges and future directions. NEUCOM, 612, 128714. (www)

  @article{ , abstract = {In recent years, the number of digital images has grown exponentially because of the widespread use of fast internet and smart devices. The integrity authentication of these images is a major concern for the research community. So, the encryption schemes that are commonly used to protect these images are an important subject for many potential applications. This paper presents a comprehensive survey of recent image encryption techniques using deep learning models. First, we explain the reasons that image encryption using deep learning models is beneficial to researchers and the public. Second, we discuss various state-of-art encryption techniques using deep learning models and offer technical summaries of popular techniques. Third, we provide a comparative analysis of our survey and existing state-of-the-art surveys. Finally, by investigating existing deep learning-based encryption, we identify several important research challenges and possible solutions including standard security metrics. To the best of our knowledge, we are the first researchers to do a detailed survey of deep learning-based image encryption for digital images.}, archive = {J_NEUCOM}, author = {Om Prakash Singh and Kedar Nath Singh and Amit Kumar Singh and Amrit Kumar Agrawal}, doi = {10.1016/j.neucom.2024.128714}, journal = {Neurocomputing}, month = {1}, pages = {128714}, shortjournal = {Neurocomputing}, title = {Deep learning-based image encryption techniques: Fundamentals, current trends, challenges and future directions}, volume = {612}, year = {2025}, }
• (2025). MARA: A deep learning based framework for multilayer graph simplification. NEUCOM, 612, 128712. (www)

  @article{ , abstract = {In many scientific fields, complex systems are characterized by a multitude of heterogeneous interactions/relationships that are challenging to model. Multilayer graphs constitute valuable tools that can represent such complex systems, thus making possible their analysis for downstream decision-making processes. Nevertheless, modeling such complex information still remains challenging in real-world scenarios. On the one hand, holistically including all relationships may lead to noisy or computationally intensive graphs. On the other hand, limiting the amount of information to model through the selection of a portion of the available relationships can introduce boundary specification biases. However, the current research studies are demonstrating that it is more beneficial to retain as much information as possible and at a later stage perform graph simplification i.e., removing uninformative or redundant parts of the graph to facilitate the final analysis. While simplification strategies, based on deep learning methods, have been already extensively explored in the context of single-layer graphs, only a limited amount of efforts have been devoted to simplification strategies for multilayer graphs. In this work, we propose the MultilAyer gRaph simplificAtion ( MARA ) framework, a GNN-based approach designed to simplify multilayer graphs based on the downstream task. MARA generates node embeddings for a specific task by training jointly two main components: (i) an edge simplification module and (ii) a (multilayer) graph neural network. We tested MARA on different real-world multilayer graphs for node classification tasks. Experimental results show the effectiveness of the proposed approach: MARA reduces the dimension of the input graph while keeping and even improving the performance of node classification tasks in different domains and across graphs characterized by different structures. Moreover, deep learning-based simplification allows MARA to preserve and enhance important graph properties for the downstream task. To our knowledge, MARA represents the first simplification framework especially tailored for multilayer graphs analysis.}, archive = {J_NEUCOM}, author = {Cheick Tidiane Ba and Roberto Interdonato and Dino Ienco and Sabrina Gaito}, doi = {10.1016/j.neucom.2024.128712}, journal = {Neurocomputing}, month = {1}, pages = {128712}, shortjournal = {Neurocomputing}, title = {MARA: A deep learning based framework for multilayer graph simplification}, volume = {612}, year = {2025}, }
• (2025). ESERNet: Learning spectrogram structure relationship for effective speech emotion recognition with swin transformer in classroom discourse analysis. NEUCOM, 612, 128711. (www)

  @article{ , abstract = {Speech emotion recognition (SER) has received increased attention due to its extensive applications in many fields, especially in the analysis of teacher-student dialogue in classroom environment. It can help teachers to better learn about students’ emotions and thereby adjust teaching activities. However, SER has faced several challenges, such as the intrinsic ambiguity of emotions and the complex task of interpreting emotions from speech in noisy environments. These issues can result in reduced recognition accuracy due to a focus on less relevant or insignificant features. To address these challenges, this paper presents ESERNet, a Transformer-based model designed to effectively extract crucial clues from speech data by capturing both pivotal cues and long-range relationships in speech signal. The major contribution of our approach is a two-pathway SER framework. By leveraging the Transformer architecture, ESERNet captures long-range dependencies within speech mel-spectrograms, enabling a refined understanding of the emotional cues embedded in speech signals. Extensive experiments were conducted on the IEMOCAP and EmoDB datasets, the results show that ESERNet achieves state-of-the-art performance in SER and outperforms existing methods by effectively leveraging critical clues and capturing long-range dependencies in speech data. These results highlight the effectiveness of the model in addressing the complex challenges associated with SER tasks.}, archive = {J_NEUCOM}, author = {Tingting Liu and Minghong Wang and Bing Yang and Hai Liu and Shaoxin Yi}, doi = {10.1016/j.neucom.2024.128711}, journal = {Neurocomputing}, month = {1}, pages = {128711}, shortjournal = {Neurocomputing}, title = {ESERNet: Learning spectrogram structure relationship for effective speech emotion recognition with swin transformer in classroom discourse analysis}, volume = {612}, year = {2025}, }
• (2025). Ranking-based adaptive query generation for DETRs in crowded pedestrian detection. NEUCOM, 612, 128710. (www)

  @article{ , abstract = {Variants of DEtection TRansformer (DETRs) have shown promising performance in crowded pedestrian detection. However, we observe that DETRs are sensitive to the hyper-parameter (the number of queries). Adjusting this hyper-parameter is crucial for achieving competitive performance across different crowded pedestrian datasets. Existing query generation methods are limited to generate a fixed number of queries based on this hyper-parameter, which often leads to missed detections and incorrect detections due to the varied number and density of pedestrians in crowded scenes. To address this challenge, we propose an adaptive query generation method called Ranking-based Adaptive Query Generation (RAQG). RAQG comprises three components: a ranking prediction head, a query supplementer, and Soft Gradient L1 Loss (SGL1). Specifically, we leverage the ranking of the lowest confidence score positive training sample to generate queries adaptively. The ranking prediction head predicts this ranking, which guides our query generation. Additionally, to refine the query generation process, we introduce a query supplementer that adjusts the number of queries based on the predicted ranking. Furthermore, we introduce SGL1, a novel loss function for training the ranking prediction head over a wide regression range. Our method is designed to be lightweight and universal, suitable for integration into any DETRs framework for crowded pedestrian detection. Experimental results on Crowdhuman and Citypersons datasets demonstrate that our RAQG method can generate queries adaptively and achieves competitive results. Notably, our approach achieves a state-of-the-art 39.4% MR on Crowdhuman.}, archive = {J_NEUCOM}, author = {Feng Gao and Jiaxu Leng and Ji Gan and Xinbo Gao}, doi = {10.1016/j.neucom.2024.128710}, journal = {Neurocomputing}, month = {1}, pages = {128710}, shortjournal = {Neurocomputing}, title = {Ranking-based adaptive query generation for DETRs in crowded pedestrian detection}, volume = {612}, year = {2025}, }
• (2025). Enhancing aspect-based sentiment analysis with linking words-guided emotional augmentation and hybrid learning. NEUCOM, 612, 128705. (www)

  @article{ , abstract = {Aspect-based sentiment analysis (ABSA) is a sophisticated task in the field of natural language processing that aims to identify emotional tendencies related to specific aspects of text. However, ABSA often faces significant data shortages, which limit the availability of annotated data for training and affects the robustness of models. Moreover, when a text contains multiple emotional dimensions, these dimensions can interact, complicating the judgments of emotional polarity. In response to these challenges, this study proposes an innovative training framework: Linking words-guided multidimensional emotional data augmentation and adversarial contrastive training (LWEDA-ACT). Specifically, this method alleviates the issue of data scarcity by synthesizing additional training samples using four different text generators. To obtain the most representative samples, we selected them by calculating sentence entropy. Meanwhile, to reduce potential noise, we introduced linking words to ensure text coherence. Additionally, by applying adversarial training, the model is able to learn generalized feature representations to handle minor input perturbations, thereby enhancing its robustness and accuracy in complex emotional dimension interactions. Through contrastive learning, we constructed positive and negative sample pairs, enabling the model to more accurately identify and distinguish the sentiment polarity of different aspect terms. We conducted comprehensive experiments on three popular ABSA datasets, namely Restaurant, Laptop, and Twitter, and compared our method against the current state-of-the-art techniques. The experimental results demonstrate that our approach achieved an accuracy improvement of +0.98% and a macro F1 score increase of +0.52% on the Restaurant dataset. Additionally, on the challenging Twitter dataset, our method improved accuracy by +0.77% and the macro F1 score by +1.14%.}, archive = {J_NEUCOM}, author = {Deling Huang and Ling Ren and Zanxiong Li}, doi = {10.1016/j.neucom.2024.128705}, journal = {Neurocomputing}, month = {1}, pages = {128705}, shortjournal = {Neurocomputing}, title = {Enhancing aspect-based sentiment analysis with linking words-guided emotional augmentation and hybrid learning}, volume = {612}, year = {2025}, }
• (2025). Advancing MRI segmentation with CLIP-driven semi-supervised learning and semantic alignment. NEUCOM, 612, 128690. (www)

  @article{ , abstract = {Precise segmentation and reconstruction of multi-structures within MRI are crucial for clinical applications such as surgical navigation. However, medical image segmentation faces several challenges. Although semi-supervised methods can reduce the annotation workload, they often suffer from limited robustness. To address this issue, this study proposes a novel CLIP-driven semi-supervised model, that includes two branches and a module. In the image branch, copy-paste is used as data augmentation method to enhance consistency learning. In the text branch, patient-level information is encoded via CLIP to drive the image branch. Notably, a novel cross-modal fusion module is designed to enhance the alignment and representation of text and image. Additionally, a semantic spatial alignment module is introduced to register segmentation results from different axial MRIs into a unified space. Three multi-modal datasets (one private and two public) were constructed to demonstrate the model’s performance. Compared to previous state-of-the-art methods, this model shows a significant advantage with both 5% and 10% labeled data. This study constructs a robust semi-supervised medical segmentation model, particularly effective in addressing label inconsistency and abnormal organ deformations. It also tackles the axial non-orthogonality challenges inherent in MRI, providing a consistent view of multi-structures.}, archive = {J_NEUCOM}, author = {Bo Sun and Kexuan Li and Jingjuan Liu and Zhen Sun and Xuehao Wang and Yuanbo He and Xin Zhao and Huadan Xue and Aimin Hao and Shuai Li and Yi Xiao}, doi = {10.1016/j.neucom.2024.128690}, journal = {Neurocomputing}, month = {1}, pages = {128690}, shortjournal = {Neurocomputing}, title = {Advancing MRI segmentation with CLIP-driven semi-supervised learning and semantic alignment}, volume = {612}, year = {2025}, }
• (2025). Dual-level adaptive incongruity-enhanced model for multimodal sarcasm detection. NEUCOM, 612, 128689. (www)

  @article{ , abstract = {Multimodal sarcasm detection leverages multimodal information, such as image, text, etc. to identify special instances whose superficial emotional expression is contrary to the actual emotion. Existing methods primarily focused on the incongruity between text and image information for sarcasm detection. Existing sarcasm methods in which the tendency of image encoders to encode similar images into similar vectors, and the introduction of noise in graph-level feature extraction due to negative correlations caused by the accumulation of GAT layers and the lack of representations for non-neighboring nodes. To address these limitations, we propose a Dual-Level Adaptive Incongruity-Enhanced Model (DAIE) to extract the incongruity between the text and image at both token and graph levels. At the token level, we bolster token-level contrastive learning with patch-based reconstructed image to capture common and specific features of images, thereby amplifying incongruities between text and images. At the graph level, we introduce adaptive graph contrast learning, coupled with negative pair similarity weights, to refine the feature representation of the model’s textual and visual graph nodes, while also enhancing the information exchange among neighboring nodes. We conduct experiments using a publicly available sarcasm detection dataset. The results demonstrate the effectiveness of our method, outperforming several state-of-the-art approaches by 3.33% and 4.34% on accuracy and F1 score, respectively.}, archive = {J_NEUCOM}, author = {Qiaofeng Wu and Wenlong Fang and Weiyu Zhong and Fenghuan Li and Yun Xue and Bo Chen}, doi = {10.1016/j.neucom.2024.128689}, journal = {Neurocomputing}, month = {1}, pages = {128689}, shortjournal = {Neurocomputing}, title = {Dual-level adaptive incongruity-enhanced model for multimodal sarcasm detection}, volume = {612}, year = {2025}, }
• (2025). EATSA-GNN: Edge-aware and two-stage attention for enhancing graph neural networks based on teacher–student mechanisms for graph node classification. NEUCOM, 612, 128686. (www)

  @article{ , abstract = {Graph Neural Networks (GNNs) have fundamentally transformed the way in which we handle and examine data originating from non-Euclidean domains. Traditional approaches to imbalanced node classification problems, such as resampling, are ineffective because they do not take into account the underlying network structure of the edges. The limited methods available to capture the intricate connections encoded in the edges of a graph pose a significant challenge for GNNs in accurately classifying nodes. We propose EATSA-GNN model to enhance GNN node classification using Edge-Aware and Two-Stage Attention Mechanisms (EATSA-GNN). EATSA-GNN focuses its initial attention on edge traits, enabling the model to differentiate the variable significance of different connections between nodes, referred to as Teacher-Attention (TA). In the second step, attention is directed towards the nodes, incorporating the knowledge obtained from the edge-level analysis referred to as Student-Attention (SA). Employing this dual strategy ensures a more sophisticated comprehension of the graph’s framework, resulting in improved classification precision. The EATSA-GNN model’s contribution to the field of GNNs lies in its ability to utilise both node and edge information in a cohesive manner, resulting in more accurate node classifications. This highlights the essence of the model and its potential. Comparing the EATSA-GNN model to state-of-the-arts methods with two different variants shows how strong it is and how well it can handle complex problems for node classification. This solidifies its position as one of leading solution in the field of GNN architectures and their use in complex networked systems. The exceptional performance of EATSA-GNN not only showcases its effectiveness but also underscores its potential to greatly influence the future advancement of the GNN framework. Implementation of the proposed EATSA-GNN can be accessed here https://github.com/afofanah/EATSA-GNN .}, archive = {J_NEUCOM}, author = {Abdul Joseph Fofanah and Alpha Omar Leigh}, doi = {10.1016/j.neucom.2024.128686}, journal = {Neurocomputing}, month = {1}, pages = {128686}, shortjournal = {Neurocomputing}, title = {EATSA-GNN: Edge-aware and two-stage attention for enhancing graph neural networks based on teacher–student mechanisms for graph node classification}, volume = {612}, year = {2025}, }
• (2025). AP-net: Attention-fused volume and progressive aggregation for accurate stereo matching. NEUCOM, 612, 128685. (www)

  @article{ , abstract = {Stereo matching plays an important role in advancing substantial applications such as 3D reconstruction, autonomous driving, depth sensing, and has recently made tremendous progress. It still poses a significant challenge, however, in the extraction of disparities from a pair of rectified images to achieve accurate stereo matching. In this paper, we propose an attention-fused cost volume construction and progressive cost aggregation network, namely AP-Net, to elevate the accuracy of stereo matching. Specifically, we introduce attention fusion to learn feature similarity between corresponding points on the left and right images while capturing long-dependent contextual information to construct attention-fused cost volumes. Moreover, we propose progressive cost aggregation, which consists of two stages to fully exploit information from different cost volumes. Furthermore, to make more accurate adjustments to the initial disparity, we implement an information entropy-guided disparity refinement module, which leverages information entropy as an explicit confidence measure of the initial disparity estimation to guide the refinement module in calculating the disparity residual more accurately and improving the overall accuracy of disparity estimation. Our implemented disparity refinement module can also be seamlessly embedded into a variety of stereo matching networks, significantly improving the model’s performance with only a small increase in computation and parameters. Experimental results across various stereo datasets confirm that our proposed AP-Net consistently delivers competitive performance. The code is available at https://github.com/zhuys-bupt/AP-Net .}, archive = {J_NEUCOM}, author = {Yansong zhu and Songwei Pei and BingFeng Liu and Jun Gao}, doi = {10.1016/j.neucom.2024.128685}, journal = {Neurocomputing}, month = {1}, pages = {128685}, shortjournal = {Neurocomputing}, title = {AP-net: Attention-fused volume and progressive aggregation for accurate stereo matching}, volume = {612}, year = {2025}, }
• (2025). RFCSC: Communication efficient reinforcement federated learning with dynamic client selection and adaptive gradient compression. NEUCOM, 612, 128672. (www)

  @article{ , abstract = {In the field of public safety, high-quality data is often owned by governments, companies, and organizations, making it difficult to train effective models through centralized datasets. This paper leverages federated learning as a mechanism to address data privacy concerns. Within the federated learning framework, the data across different clients is typically Non-IID (non-Independently and Identically Distributed). Furthermore, The complexity of sensitive image recognition tasks in public safety, along with the large number of model parameters, can lead to communication congestion issues in federated learning. To solve these challenges, this paper proposes a Reinforcement Federated Client Selection and Gradient Compression Method(RFCSC). By integrating client data prototypes with accuracy metrics, we dynamically assess the contribution of each client in federated learning. Through an intelligent dynamic incentive mechanism based on reinforcement learning, high-quality client nodes are dynamically selected to participate in federated learning, achieving dynamic adaptive aggregation, reducing the influence of Non-IID data, reducing communication cost, enhancing model accuracy, and achieving a balance between quality and efficiency. To address the issue of high communication cost for parameter transmission, we propose an adaptive model compression strategy in reinforcement federated learning which make local clients find an appropriate compression rate. This approach not only reduces gradient communication overhead but also minimizes the impact of compression on model accuracy. The effectiveness of our proposed approach are corroborated through comprehensive experiments conducted on one private dataset and two public datasets.}, archive = {J_NEUCOM}, author = {Zhenhui Pan and Yawen Li and Zeli Guan and Meiyu Liang and Ang Li and Jia Wang and Feifei Kou}, doi = {10.1016/j.neucom.2024.128672}, journal = {Neurocomputing}, month = {1}, pages = {128672}, shortjournal = {Neurocomputing}, title = {RFCSC: Communication efficient reinforcement federated learning with dynamic client selection and adaptive gradient compression}, volume = {612}, year = {2025}, }
• (2025). Modality-specific adaptive scaling and attention network for cross-modal retrieval. NEUCOM, 612, 128664. (www)

  @article{ , abstract = {There are huge differences in data distribution and feature representation of different modalities. How to flexibly and accurately retrieve data from different modalities is a challenging problem. The mainstream common subspace methods only focus on the heterogeneity gap, and use a unified method to jointly learn the common representation of different modalities, which can easily lead to the difficulty of multi-modal unified fitting. In this work, we innovatively propose the concept of multi-modal information density discrepancy, and propose a modality-specific adaptive scaling method incorporating prior knowledge, which can adaptively learn the most suitable network for different modalities. Secondly, for the problem of efficient semantic fusion and interference features, we propose a multi-level modal feature attention mechanism, which realizes the efficient fusion of text semantics through attention mechanism, explicitly captures and shields the interference features from multiple scales. In addition, to address the bottleneck of cross-modal retrieval task caused by the insufficient quality of multimodal common subspace and the defects of Transformer structure, this paper proposes a cross-level interaction injection mechanism to fuse multi-level patch interactions without affecting the pre-trained model to construct higher quality latent representation spaces and multimodal common subspaces. Comprehensive experimental results on four widely used cross-modal retrieval datasets show the proposed MASAN achieves the state-of-the-art results and significantly outperforms other existing methods.}, archive = {J_NEUCOM}, author = {Xiao Ke and Baitao Chen and Yuhang Cai and Hao Liu and Wenzhong Guo and Weibin Chen}, doi = {10.1016/j.neucom.2024.128664}, journal = {Neurocomputing}, month = {1}, pages = {128664}, shortjournal = {Neurocomputing}, title = {Modality-specific adaptive scaling and attention network for cross-modal retrieval}, volume = {612}, year = {2025}, }
• (2025). Salient event detection via hypergraph convolutional network with cross-view self-supervised learning. NEUCOM, 612, 128639. (www)

  @article{ , abstract = {Identifying the most dominant and central events in a document is critical for holistically understanding its important information. To measure the importance of an event, it is critical to understand its context: who is involved, where it happened, which other events it is related to, and what kind of relationship it has, among other factors. Although existing studies have achieved some accomplishments, they are still not fully effective for two main reasons: (1) They incorporate only the discrete global features of the document, which is insufficient for effectively capturing the contextual information of events; (2) They inadequately model the dependency relationships between events. According to previous research findings, it has been shown that hypergraphs effectively capture the global context (i.e., the document-level context) of long text. However, their potential for salient event detection has remained unexplored. To address this, we propose a novel framework called Salient Event Detection via Hypergraph Convolutional Network with Graph Self-supervised Learning (SEDGS). More specifically, we first construct two hypergraphs: one in the event argument view and another in the event view. We then propose a hypergraph convolutional network to model the event context and discourse relations between events. Moreover, to enhance hypergraph modeling and ensure consistency between argument-view and event-view event representations, we employ contrastive self-supervised learning (SSL) in our model training. Experimental results on a standard event salience dataset verify the superiority of SEDGS, advancing state-of-the-art models.}, archive = {J_NEUCOM}, author = {Enchang Zhu and Zhengtao Yu and Yuxin Huang and Shengxiang Gao and Yantuan Xian}, doi = {10.1016/j.neucom.2024.128639}, journal = {Neurocomputing}, month = {1}, pages = {128639}, shortjournal = {Neurocomputing}, title = {Salient event detection via hypergraph convolutional network with cross-view self-supervised learning}, volume = {612}, year = {2025}, }
• (2025). A hybrid novel SWARA-ELECTRE-I method using probabilistic uncertain linguistic information for feature selection in image recognition. NEUCOM, 612, 128615. (www)

  @article{ , abstract = {In the digital age, the exponential growth of data poses significant challenges for analysts and machine learning algorithms in pattern detection due to its high dimensionality. This study addresses the dimensionality problem by leveraging Probabilistic Uncertain Linguistic Term Set (PULTS), which combine Uncertain Linguistic Term Set (ULTS) with associated probabilities to handle uncertainty in decision-making. We introduce the PUL-weighted average operator to integrate the opinions of multiple decision-makers and propose a novel ELimination and Choice Translating REality (ELECTRE-I) method for optimizing alternatives in multiple attribute group decision-making (MAGDM) scenarios. This method is enhanced by the Stepwise Weight Assessment Ratio Analysis (SWARA) method to determine the relative weight of each attribute. By integrating SWARA with the ELECTRE-I method, we develop a comprehensive approach to tackle MAGDM problems using PULTS. A numerical example involving feature selection in image recognition demonstrates the method’s effectiveness and accuracy. Comparative studies highlight the advantages of our approach in producing a small feature set with high classification accuracy. The proposed method offers a robust solution for feature selection in image recognition and other MAGDM problems, significantly improving decision-making accuracy and efficiency. The methodology’s simplicity and computational ease make it applicable across various domains requiring effective dimensionality reduction.}, archive = {J_NEUCOM}, author = {Sumera Naz and Shariq Aziz Butt and Muhammad Muneeb ul Hassan and José Escorcia-Gutierrez and Areej Fatima and Farhat ul Ain}, doi = {10.1016/j.neucom.2024.128615}, journal = {Neurocomputing}, month = {1}, pages = {128615}, shortjournal = {Neurocomputing}, title = {A hybrid novel SWARA-ELECTRE-I method using probabilistic uncertain linguistic information for feature selection in image recognition}, volume = {612}, year = {2025}, }
• (2025). Encoding–decoding-based distributed state estimation over sensor networks with limited sensing range under DoS attacks. NEUCOM, 611, 128713. (www)

  @article{ , abstract = {This paper is concerned with encoding–decoding-based distributed state estimation over sensor networks under DoS attacks. Different from most of the existing research results on distributed state estimation for sensor networks, where all sensors are assumed to have sufficiently wide sensing ranges, this paper considers sensors with limited sensing ranges. Therefore, the problem studied has more practical significance. To save the limited bandwidth resources of sensor networks, a two-channel encoding–decoding scheme (EDS) based on probability is proposed for each node to compress the transmitted data to an acceptable range, where independent DoS attacks are launched randomly on the communication channels between nodes. Then, a distributed state estimator with limited sensing ranges under DoS attacks in the presence of both the sensor-estimator channel EDS and the node–node channel EDS is constructed under the criterion of minimum mean-square error. Furthermore, considering the real-time changes of the communication topology resulting from independent DoS attacks and the uncertainty introduced by the node–node channel EDS, the upper bound of the expected estimation error covariance is derived and the boundedness of the upper bound is analyzed under given assumption conditions. Finally, a numerical example is exhibited to illustrate the effectiveness of the designed algorithm.}, archive = {J_NEUCOM}, author = {Xufeng Lin and Yanyan Hu and Xuechun Zhang and Kaixiang Peng}, doi = {10.1016/j.neucom.2024.128713}, journal = {Neurocomputing}, month = {1}, pages = {128713}, shortjournal = {Neurocomputing}, title = {Encoding–decoding-based distributed state estimation over sensor networks with limited sensing range under DoS attacks}, volume = {611}, year = {2025}, }
• (2025). Predicting functional connectivity network from routinely acquired t1-weighted imaging-based brain network by generative U-GCNet. NEUCOM, 611, 128709. (www)

  @article{ , abstract = {Predicting the function magnetic resonance imaging (fMRI)-based brain network (fMRI-BN) from structure MRI-based brain network is imperative in clinical practice because fMRIs are not routinely acquired in a vast majority of hospitals. In this study, a generative U-GCNet (U-shaped graph convolutional Network) was proposed to predict fMRI-BN from radiomics-based morphological brain network (radMBN) derived from routinely acquired T1-WI image. Specifically, the U-GCNet consisted of a graph convolutional network (GCN) encoder module (En-GCN), a deep feature connectivity construction module (DF2C), and a GCN decoder module (De-GCN). Both En-GCN and De-GCN employed mixed local-and-long distance node feature aggregation strategy to enhance the graph encoding and decoding ability, and the DF2C reshaped the deep feature matrix into the connectivity matrix for outputting the brain network prediction. Additionally, a multi-scale network similarity loss function was conducted on full values, upper triangular values, and each row values of connectivity matrix. Experiments on 3169 subjects from three publicly available databases demonstrated that the U-GCNet could predict the fMRI-BN from radMBN with a promising performance (MSE [0.0002 0.0025], PCC [0.956 0.991]) over eight alternative methods under comparison. The results exhibited a significant correlation (PCC [0.796, 0.897], P<0.05) between the estimated and real radiomics-function coupling values. The individual-level and group-level brain network visualization was displayed with high consistency. The TOP brain regions identified by four graph-based metrics also exhibited with consistency. These results demonstrated that the proposed U-GCNet could achieve promising prediction of fMRI-BN from radMBN which could alleviate the limited availability of fMRI and boost its usage in clinical practice.}, archive = {J_NEUCOM}, author = {Zhiwei Song and Chuanzhen Zhu and Minbo Jiang and Minhui Ouyang and Qiang Zheng}, doi = {10.1016/j.neucom.2024.128709}, journal = {Neurocomputing}, month = {1}, pages = {128709}, shortjournal = {Neurocomputing}, title = {Predicting functional connectivity network from routinely acquired t1-weighted imaging-based brain network by generative U-GCNet}, volume = {611}, year = {2025}, }
• (2025). A novel label-aware global graph construction method and spiking-coded graph neural network for intelligent process fault diagnosis. NEUCOM, 611, 128707. (www)

  @article{ , abstract = {Fault diagnosis plays a crucial role in ensuring the safety and efficiency of industrial processes. However, traditional techniques often face difficulties in handling large-scale data characterized by complex structures and relationships. To efficiently represent industrial data as graphs and develop a low-energy-cost feature extraction model, a novel label-aware global graph construction method and a spiking graph convolutional network (SGCN) are proposed in this study to achieve intelligent process fault diagnosis. The label-aware method enhances graph data representation by capturing intrinsic correlations and global features. The SGCN integrates graph convolutional layers with spiking encoding, enabling effective feature extraction while offering computational efficiency advantages. A weighted loss function is introduced to mitigate data imbalance issues. Experiments on the Tennessee Eastman process, the Three-phase Flow Facility, and the de-propanizer distillation process demonstrate SGCN’s superior performance over baseline models in various fault scenarios, while significantly reducing computational costs. The proposed method offers promising potential for reliable and efficient fault diagnosis in complex real-world industrial environments.}, archive = {J_NEUCOM}, author = {Dazi Li and Yurui Zhu and Zhihuan Song and Hamid Reza Karimi}, doi = {10.1016/j.neucom.2024.128707}, journal = {Neurocomputing}, month = {1}, pages = {128707}, shortjournal = {Neurocomputing}, title = {A novel label-aware global graph construction method and spiking-coded graph neural network for intelligent process fault diagnosis}, volume = {611}, year = {2025}, }
• (2025). Graphical model for mixed data types. NEUCOM, 611, 128706. (www)

  @article{ , abstract = {With the development of data collection technologies, data types have become more diverse. Additionally, graphical models, as tools for describing variable network relationships, have become increasingly popular in recent years. Previous studies have focused on graphical models tailored to specific types of data. However, these existing methods fail to identify graphical models for mixed data types. The difficulty of constructing graphical models for mixed data types lies in the fact that each type of data has its own space, which challenges the estimation of network relationships in a graphical model when the data are combined. To address this issue, this study presents a novel method that utilizes a vectorization and alignment strategy developed particularly for mixed data types, including scalar, interval-valued, compositional, and functional data, to estimate a graphical model. By iteratively employing a block-sparse graphical lasso method on aligned data, the method can achieve satisfactory results, as shown by numerous simulation experiments. The results also validate the superiority of our proposed method over potential competing methods. Furthermore, this method was applied to an engine damage propagation network as an illustrative example. Our method provides a novel modeling approach for graphical models in the case of mixed data types.}, archive = {J_NEUCOM}, author = {Qiying Wu and Huiwen Wang and Shan Lu and Hui Sun}, doi = {10.1016/j.neucom.2024.128706}, journal = {Neurocomputing}, month = {1}, pages = {128706}, shortjournal = {Neurocomputing}, title = {Graphical model for mixed data types}, volume = {611}, year = {2025}, }
• (2025). Pruning convolutional neural networks for inductive conformal prediction. NEUCOM, 611, 128704. (www)

  @article{ , abstract = {Neural network pruning is a popular approach to reducing model storage size and inference time by removing redundant parameters in the neural network. However, the uncertainty of predictions from pruned models is unexplored. In this paper we study neural network pruning in the context of conformal predictors (CP). The conformal prediction framework built on top of machine learning algorithms supplements their predictions with reliable uncertainty measure in the form of prediction sets, under the independent and identically distributed assumption on the data. Convolutional neural networks (CNNs) have complicated architectures and are widely used in various applications nowadays. Therefore, we focus on pruning CNNs and, in particular, filter-level pruning. We first propose a brute force method that estimates the contribution of a filter to the CP’s predictive efficiency and removes those with the least contribution. Given the computation inefficiency of the brute force method, we also propose the Taylor expansion to approximate the filter’s contribution. Furthermore, we improve the global pruning method by protecting the most important filters within each layer from being pruned. In addition, we explore the ConfTr loss function which is optimized to yield maximal CP efficiency in the context of neural network pruning. We have conducted extensive experimental studies and compared the results regarding the trade-offs between predictive efficiency, computational efficiency, and network sparsity. These results are instructive for deploying pruned neural networks with applications using conformal prediction where reliable predictions and reduced computational cost are relevant, such as in safety-critical applications.}, archive = {J_NEUCOM}, author = {Xindi Zhao and Amin Farjudian and Anthony Bellotti}, doi = {10.1016/j.neucom.2024.128704}, journal = {Neurocomputing}, month = {1}, pages = {128704}, shortjournal = {Neurocomputing}, title = {Pruning convolutional neural networks for inductive conformal prediction}, volume = {611}, year = {2025}, }
• (2025). Robust and stochastic sparse subspace clustering. NEUCOM, 611, 128703. (www)

  @article{ , abstract = {Sparse subspace clustering (SSC) has been widely employed in machine learning and pattern recognition, but it still faces scalability challenges when dealing with large-scale datasets. Recently, stochastic SSC (SSSC) has emerged as an effective solution by leveraging the dropout technique. However, SSSC cannot robustly handle noise, especially non-Gaussian noise, leading to unsatisfactory clustering performance. To address the above issues, we propose a novel robust and stochastic method called stochastic sparse subspace clustering with the Huber function (S3CH). The key idea is to introduce the Huber surrogate to measure the loss of the stochastic self-expression framework, thus S3CH inherits the advantage of the stochastic framework for large-scale problems while mitigating sensitivity to non-Gaussian noise. In algorithms, an efficient proximal alternating minimization (PAM)-based optimization scheme is developed. In theory, the convergence of the generated sequence is rigorously proved. Extensive numerical experiments on synthetic and six real datasets validate the advantages of the proposed method in clustering accuracy, noise robustness, parameter sensitivity, post-hoc analysis, and model stability.}, archive = {J_NEUCOM}, author = {Yanjiao Zhu and Xinrong Li and Xianchao Xiu and Wanquan Liu and Chuancun Yin}, doi = {10.1016/j.neucom.2024.128703}, journal = {Neurocomputing}, month = {1}, pages = {128703}, shortjournal = {Neurocomputing}, title = {Robust and stochastic sparse subspace clustering}, volume = {611}, year = {2025}, }
• (2025). Online parallel multi-task relationship learning via alternating direction method of multipliers. NEUCOM, 611, 128702. (www)

  @article{ , abstract = {Online multi-task learning (OMTL) enhances streaming data processing by leveraging the inherent relations among multiple tasks. It can be described as an optimization problem in which a single loss function is defined for multiple tasks. Existing gradient-descent-based methods for this problem might suffer from gradient vanishing and poor conditioning issues. Furthermore, the centralized setting hinders their application to online parallel optimization, which is vital to big data analytics. Therefore, this study proposes a novel OMTL framework based on the alternating direction multiplier method (ADMM), a recent breakthrough in optimization suitable for the distributed computing environment because of its decomposable and easy-to-implement nature. The relations among multiple tasks are modeled dynamically to fit the constant changes in an online scenario. In a classical distributed computing architecture with a central server, the proposed OMTL algorithm with the ADMM optimizer outperforms SGD-based approaches in terms of accuracy and efficiency. Because the central server might become a bottleneck when the data scale grows, we further tailor the algorithm to a decentralized setting, so that each node can work by only exchanging information with local neighbors. Experimental results on a synthetic and several real-world datasets demonstrate the efficiency of our methods.}, archive = {J_NEUCOM}, author = {Ruiyu Li and Peilin Zhao and Guangxia Li and Zhiqiang Xu and Xuewei Li}, doi = {10.1016/j.neucom.2024.128702}, journal = {Neurocomputing}, month = {1}, pages = {128702}, shortjournal = {Neurocomputing}, title = {Online parallel multi-task relationship learning via alternating direction method of multipliers}, volume = {611}, year = {2025}, }
• (2025). Generalization in neural networks: A broad survey. NEUCOM, 611, 128701. (www)

  @article{ , abstract = {This paper reviews concepts, modeling approaches, and recent findings along a spectrum of different levels of abstraction of neural network models including generalization across (1) Samples, (2) Distributions, (3) Domains, (4) Tasks, (5) Modalities, and (6) Scopes. Strategies for (1) sample generalization from training to test data are discussed, with suggestive evidence presented that, at least for the ImageNet dataset, popular classification models show substantial overfitting. An empirical example and perspectives from statistics highlight how models’ (2) distribution generalization can benefit from consideration of causal relationships and counterfactual scenarios. Transfer learning approaches and results for (3) domain generalization are summarized, as is the wealth of domain generalization benchmark datasets available. Recent breakthroughs surveyed in (4) task generalization include few-shot meta-learning approaches and the emergence of transformer-based foundation models such as those used for language processing. Studies performing (5) modality generalization are reviewed, including those that integrate image and text data and that apply a biologically-inspired network across olfactory, visual, and auditory modalities. Higher-level (6) scope generalization results are surveyed, including graph-based approaches to represent symbolic knowledge in networks and attribution strategies for improving networks’ explainability. Additionally, concepts from neuroscience are discussed on the modular architecture of brains and the steps by which dopamine-driven conditioning leads to abstract thinking.}, archive = {J_NEUCOM}, author = {Chris Rohlfs}, doi = {10.1016/j.neucom.2024.128701}, journal = {Neurocomputing}, month = {1}, pages = {128701}, shortjournal = {Neurocomputing}, title = {Generalization in neural networks: A broad survey}, volume = {611}, year = {2025}, }
• (2025). An efficient CNN accelerator for pattern-compressed sparse neural networks on FPGA. NEUCOM, 611, 128700. (www)

  @article{ , abstract = {Currently, the sparsity of weights and activations are mainly utilized to improve the energy efficiency and computational performance of CNN accelerators. However, the irregular sparsity of weights and activations can lead to problems such as mapping access conflict and imbalanced workload, which poses great challenges to the efficient computation of sparse networks. To address those problems, this paper proposes a CNN accelerator for pattern-compressed sparse neural networks on FPGA to achieve efficient inference acceleration. It mainly includes three parts: first, we propose a convolutional kernel grouping fusion method to achieve a balanced workload of weights. Second, we propose an activation hash mapping conflict reduction method to drastically reduce the mapping access conflict rate. Finally, we further propose a hierarchical mapping computing architecture to achieve efficient processing of mapping access conflicts and balanced adaptation of weight loads. To verify the effectiveness of the method proposed in this paper, our accelerator is designed and implemented on the Zynq UltraScale+ MPSoC ZCU102 evaluation board and evaluated by running VGG16 and ResNet50 networks. The experimental results show that the method proposed in this paper can reduce the mapping conflicts by more than 97 % and improve the utilization rate of processing element (PE) by 1.67×. Furthermore, the accelerator implemented in this paper can achieve 393.2 GOPs throughput, 1.40–3.25× computational performance, and 1.11–7.34× energy efficiency improvement.}, archive = {J_NEUCOM}, author = {Yonghua Zhang and Haojie Wang and Zhenhua Pan}, doi = {10.1016/j.neucom.2024.128700}, journal = {Neurocomputing}, month = {1}, pages = {128700}, shortjournal = {Neurocomputing}, title = {An efficient CNN accelerator for pattern-compressed sparse neural networks on FPGA}, volume = {611}, year = {2025}, }
• (2025). A generic shared attention mechanism for various backbone neural networks. NEUCOM, 611, 128697. (www)

  @article{ , abstract = {The self-attention mechanism is crucial for enhancing various backbone neural networks’ performance. However, current methods add self-attention modules (SAMs) to each network layer without fully utilizing their potential, resulting in suboptimal performance and higher parameter consumption as network depth increases. In this paper, we reveal an inherent phenomenon: SAMs produce highly correlated attention maps across layers, with an average Pearson correlation coefficient of 0.85. Inspired by this inherent observation, we propose Dense-and-Implicit Attention (DIA), which shares SAMs across layers and uses a long short-term memory module to calibrate and connect these correlated attention maps, improving parameter efficiency. This approach aligns with the neural network’s dynamic system perspective. Extensive experiments show DIA consistently enhances various backbones like ResNet, Transformer, and UNet in tasks such as image classification, object detection, and image generation with diffusion models. Our analysis indicates that DIA’s effectiveness stems from its dense inter-layer information connections, absent in conventional mechanisms, stabilizing training and providing regularization effects. This paper’s insights advance our understanding of attention mechanisms, optimizing them and paving the way for future developments across diverse neural networks.}, archive = {J_NEUCOM}, author = {Zhongzhan Huang and Senwei Liang and Mingfu Liang}, doi = {10.1016/j.neucom.2024.128697}, journal = {Neurocomputing}, month = {1}, pages = {128697}, shortjournal = {Neurocomputing}, title = {A generic shared attention mechanism for various backbone neural networks}, volume = {611}, year = {2025}, }
• (2025). Adaptive search for broad attention based vision transformers. NEUCOM, 611, 128696. (www)

  @article{ , abstract = {Vision Transformer (ViT) has prevailed among computer vision tasks for its powerful capability of image representation recently. Frustratingly, the manual design of efficient architectures for ViTs can be laborious, often involving repetitive trial and error processes. Furthermore, the exploration of lightweight ViTs remains limited, resulting in inferior performance compared to convolutional neural networks. To tackle these challenges, we propose Adaptive Search for Broad attention based Vision Transformers, called ASB, which automates the design of efficient ViT architectures by utilizing the broad search space and an adaptive evolutionary algorithm. The broad search space facilitates the exploration of a novel connection paradigm, enabling more comprehensive integration of attention information to improve ViT performance. Additionally, an efficient adaptive evolutionary algorithm is developed to efficiently explore architectures by dynamically learning the probability distribution of candidate operators. Our experimental results demonstrate that the adaptive evolution in ASB efficiently learns excellent lightweight models, achieving a 55% improvement in convergence speed over traditional evolutionary algorithms. Moreover, the effectiveness of ASB is validated across several visual tasks. For instance, on ImageNet classification, the searched model attains a performance of 77.8% with 6.5M parameters and outperforms state-of-the-art models, including EfficientNet and EfficientViT networks. On mobile COCO panoptic segmentation, our approach delivers 43.7% PQ. On mobile ADE20K semantic segmentation, our method attains 40.9% mIoU. The code and pre-trained models will be available soon in ASB-Code .}, archive = {J_NEUCOM}, author = {Nannan Li and Yaran Chen and Dongbin Zhao}, doi = {10.1016/j.neucom.2024.128696}, journal = {Neurocomputing}, month = {1}, pages = {128696}, shortjournal = {Neurocomputing}, title = {Adaptive search for broad attention based vision transformers}, volume = {611}, year = {2025}, }
• (2025). Cross-attention multi-perspective fusion network based fake news censorship. NEUCOM, 611, 128695. (www)

  @article{ , abstract = {Current fake news censorship models mostly use only one single semantic perspective, which contains insufficient information and may result in biases. However, news inherently encompasses multiple perspectives, including internal elements such as semantics, emotion, and style, as well as the news environment which may offer valuable information for news censorship. Combining multiple perspectives helps mitigate the biases and provides comprehensive information for news censorship. Thus, we propose the Cross-Attention Multi-Perspective Fusion Network Based Fake News Censorship (MPFNC). In this model, we design an Internal Multi-Perspective Fusion Network to integrate the internal perspectives of news, including semantic, emotion, and style, and equip it with Cross-Attention to capture the correlations between these perspectives. Additionally, we construct the news environment based on the fuzzy cluster and apply Cross-Attention to uncover associations with the internal perspectives. Compared to the current SOTA models, our model achieves a 3.04% and 3.86% improvement in F1 score on Chinese and English datasets, respectively. On fake news datasets generated by LLMs, our model also outperforms the best-performing LLM by 10% in accuracy.}, archive = {J_NEUCOM}, author = {Weishan Zhang and Mingli Zhang and Zhicheng Bao and Zhenqi Wang}, doi = {10.1016/j.neucom.2024.128695}, journal = {Neurocomputing}, month = {1}, pages = {128695}, shortjournal = {Neurocomputing}, title = {Cross-attention multi-perspective fusion network based fake news censorship}, volume = {611}, year = {2025}, }
• (2025). Back-propagation-based multivariate state estimation technique: A lightweight adaptive condition monitoring approach for wind turbine. NEUCOM, 611, 128693. (www)

  @article{ , abstract = {Aimed at the challenge of wind turbine condition monitoring with insufficient available data caused by the harsh conditions, this paper presents a lightweight adaptive condition monitoring approach based on the back-propagation algorithm and the multivariate state estimation technique. In order to maximize the utility of available data, the multivariate state estimation technique is used to extract the features from multiple historical data. Moreover, buffer space is adopted to update dynamically while avoiding the effect of outliers. Then, the operator is modified so that the model can be trained and updated by the back-propagation algorithm, thereby achieving lightweight. The model performs well in condition monitoring with small samples, even discontinuous samples, that are verified through experiments based on field wind turbine datasets. The experimental results and comparative analyses demonstrate the effectiveness, generality, and advantage over others.}, archive = {J_NEUCOM}, author = {Dongsheng Yang and Huanying Han and Hamid Reza Karimi and Yesheng Zhu}, doi = {10.1016/j.neucom.2024.128693}, journal = {Neurocomputing}, month = {1}, pages = {128693}, shortjournal = {Neurocomputing}, title = {Back-propagation-based multivariate state estimation technique: A lightweight adaptive condition monitoring approach for wind turbine}, volume = {611}, year = {2025}, }
• (2025). CoMaSa: Context multi-aware self-attention for emotional response generation. NEUCOM, 611, 128692. (www)

  @article{ , abstract = {Dialogue generation is an important research direction in natural language generation, and generating utterances with appropriate contextual emotion is a challenging task. The previous work incorporates commonsense knowledge as an auxiliary information source by simply concatenating or adding it to the embeddings of the dialogue context. They did not consider updating the commonsense knowledge from the perspective of context semantic information or selecting context from the perspective of commonsense knowledge semantic information. At the same time, they did not extract emotional information from dialogue context and commonsense knowledge to enhance the embeddings of the dialogue text. These defects may lead to the dialogue model’s inability to judge the semantic consistency between context and generated utterance, and result in the generation of utterance unassociated with the dialogue context and the corresponding context emotion being untrue. Our goal is to build an emotional dialogue generation model that can deeply integrate commonsense knowledge with dialogue context and achieve bidirectional semantic interaction and enhancement. Therefore, in this paper, we propose a context multi-aware self-attention emotion dialogue generation model that models from multiple aspects such as the semantic information, emotional information of context and external knowledge, it effectively alleviates the problems of generating utterance unrelated to the context and disordered emotional expressions in utterance. First, we adopt the transformer encoder as the basic framework and carefully construct a context-aware encoder capable of deeply extracting semantic information. Specifically, we ingeniously integrate different levels of positional information to simulate the natural flow of conversation between speakers. Then, we effectively integrate these positional information with the basic word embedding vectors, and enrich the dialogue’s embedding representation. Lastly, by introducing the attention mechanism, the model can precisely capture and extract context information closely related to the current situation and achieve a deep understanding and accurate grasp of the dialogue content. Second, we consider that the relationship between context and commonsense knowledge is not just a unidirectional selection but a more intimate bidirectional interaction. Therefore, we designed a bidirectional interaction attention component for context and external knowledge. Specifically, we use the method of iterative similarity matrices to calculate the bidirectional similarity relationship between context and external knowledge. Then, using the attention algorithm, we calculate the connectivity from context to commonsense knowledge and from commonsense knowledge to context. Finally, by combining the semantic embeddings of both aspects, we achieve bidirectional updating of context and external knowledge. Third, we establish an emotion-aware decoder based on multi-task learning, which includes an emotion classifier and an utterance generator. In the emotion classifier, we establish an auxiliary task for emotion recognition to help the model understand the process of emotional transition. We use the output of the emotion recognition model as an additional supervisory signal to better control the generation of emotional responses. In the utterance generator, we integrate the output of the emotion classifier with the interaction vector from the bidirectional interaction attention component. This fusion allows us to generate utterance that aligns with the emotional context. Our experiments on DailyDialog and ESConv datasets show that CoMaSa outperforms the baselines in terms of Perplexity, Distinct-1, Distinct-2, and human evaluation. The experiment demonstrates the effectiveness of our Context Multi-aware Self-attention model in the task of emotion generation.}, archive = {J_NEUCOM}, author = {Xin Zheng and YaJun Du and XiaoLin Qin}, doi = {10.1016/j.neucom.2024.128692}, journal = {Neurocomputing}, month = {1}, pages = {128692}, shortjournal = {Neurocomputing}, title = {CoMaSa: Context multi-aware self-attention for emotional response generation}, volume = {611}, year = {2025}, }
• (2025). Consensus and diversity-fusion partial-view-shared multi-view learning. NEUCOM, 611, 128687. (www)

  @article{ , abstract = {Due to the multi-perspective of data, multi-view learning (MVL) is usually employed. Although existing MVL approaches gain fruitful achievements, they may neglect to learn (a) partial-view-shared knowledge between views; and (b) differences across views. Consequently, inadequate complementary knowledge and weak discriminability may be gained. To address the above problems, Consensus and Diversity-fusion Partial-view-shared Multi-view Learning (CDPMVL) is proposed, which includes two components: (a) Consensus, Partial-view-shared and Specific Component Learning (CPSCL) that partitions the samples into consensual, partial-view-shared, and specific parts, and learns the consensual, partial-view-shared, and specific knowledge of views; and (b) Diversity-fusion Partial-view-shared Knowledge Enhancement (DPKE) that imposes a diversity constraint on partial-view-shared parts and employs a heuristic-based auto-weighting mechanism to highlight the differences among views. By CDPMVL, more complementary relationships between and across views are explored, and the discriminability of the model is enhanced. Extensive experiments performed with eleven algorithms on nine datasets verify the superiority of CDPMVL, which indicates that the incorporation of partial-view-shared knowledge indeed enhances the complementary ability of views. The source code of CDPMVL is available at https://github.com/zzf495/CDPMVL .}, archive = {J_NEUCOM}, author = {Luyao Teng and Zefeng Zheng}, doi = {10.1016/j.neucom.2024.128687}, journal = {Neurocomputing}, month = {1}, pages = {128687}, shortjournal = {Neurocomputing}, title = {Consensus and diversity-fusion partial-view-shared multi-view learning}, volume = {611}, year = {2025}, }
• (2025). Fine-grained detoxification framework via instance-level prefixes for large language models. NEUCOM, 611, 128684. (www)

  @article{ , abstract = {Large Language Models (LLMs) have demonstrated remarkable performance across various natural language processing (NLP) tasks. However, their practical usability is often compromised by a propensity to generate toxic content, such as insults, threats, and profanity, particularly in response to adversarial prompts. Several fine-tuning and decoding approaches have been employed to address this challenge to mitigate toxicity. Nevertheless, these methods typically necessitate additional resources, such as high-quality training data or auxiliary models, thereby incurring extra costs. In this paper, we propose a novel detoxification framework, Fine-Grained Detoxification via Instance-Level Prefixes (FGDILP), which effectively mitigates the generation of toxic text without incurring additional training costs. Specifically, FGDILP leverages contextualized representations in attention space by contrasting a positive prefix-prepended prompt against multiple negative prefix-prepended prompts at the instance level. This methodology facilitates the construction of fine-grained subtoxicity vectors, which are subsequently fused to adjust the original generation pathway when responding to raw prompts. Our results demonstrate that FGDILP enables controlled text generation concerning detoxification at both the utterance and context levels. While our method slightly impacts generation fluency and diversity, it significantly outperforms prompt-based baselines regarding detoxification effectiveness. Our code is available at https://github.com/xinykou/FGDILP .}, archive = {J_NEUCOM}, author = {Xin Yi and Linlin Wang and Xiaoling Wang and Liang He}, doi = {10.1016/j.neucom.2024.128684}, journal = {Neurocomputing}, month = {1}, pages = {128684}, shortjournal = {Neurocomputing}, title = {Fine-grained detoxification framework via instance-level prefixes for large language models}, volume = {611}, year = {2025}, }
• (2025). Semi-supervised action recognition with dynamic temporal information fusion. NEUCOM, 611, 128683. (www)

  @article{ , abstract = {The most advanced semi-supervised models available are based on images for innovation, and the use of semi-supervised learning models augmented with temporal data for video-level action recognition still suffers from severe model mismatches, and the models are not sufficiently capable of capturing both local and global information about the action. Secondly the use of constant-threshold pseudo-labeling leads to low utilization of unlabeled data for difficult actions in the early stages of training, poor pseudo-labeling quality and affects recognition accuracy. To make the semi-supervised framework FixMatch more suitable for action recognition, we propose Time-Mixer and Dynamic Threshold, respectively. Time-Mixer explores complementary information between time sequences through the fusion of two-channel temporal context information. Dynamic Threshold utilizes a new core mapping function (Normal Distribution Function) to enhance pseudo-labeling quality. Extensive experiments were conducted on three action recognition datasets (Kinetics-400, UCF-101, and HMDB-51). Comprehensive experiments show that the performance of the semi-supervised model in action recognition improves considerably after using dynamic thresholding and temporal context information fusion, with a 14.4% improvement over the baseline and a 1.8% improvement over the TG (with a labeling rate of 10%) in UCF101, whereas an overall good performance is obtained for DTIF.}, archive = {J_NEUCOM}, author = {Huifang Qian and Jialun Zhang and Zhenyu Shi and Yimin Zhang}, doi = {10.1016/j.neucom.2024.128683}, journal = {Neurocomputing}, month = {1}, pages = {128683}, shortjournal = {Neurocomputing}, title = {Semi-supervised action recognition with dynamic temporal information fusion}, volume = {611}, year = {2025}, }
• (2025). Enhancing bilingual lexicon induction via harnessing polysemous words. NEUCOM, 611, 128682. (www)

  @article{ , abstract = {Bilingual Lexicon Induction (BLI) aims to induce word translation pairs across different languages. We find there is a large proportion of polysemous words in the dataset used by BLI, which impacts the quality of the induced translation pairs. In particular, the mainstream BLI models map two monolingual word embeddings (WEs) into a shared space and regard the closest cross-lingual words as translation pairs. However, when a word has multiple translations, it becomes challenging to ensure that the polysemous word is the closest neighbor to all its translations. Based on the analysis of polysemy, we propose a simple yet effective method by harnessing polysemous words to enhance BLI. We first detect polysemous words in the seed lexicons and filter their multiple translations by comparing their semantic similarities. Therefore, a refined seed lexicon mitigates the confusion caused by translations with divergent semantics for the BLI model. Moreover, in scenarios where only single translations are available in the seed lexicon, we propose to employ a Large Language Model (LLM) to acquire monolingual synonyms, thereby refining the structure of the seed lexicon more effectively. On benchmark BLI datasets, our method demonstrates superior performance compared to five state-of-the-art (SOTA) baseline systems across almost all experimental language pairs, encompassing both sides of three similar language pairs and three distant language pairs.}, archive = {J_NEUCOM}, author = {Qiuyu Ding and Hailong Cao and Zihao Feng and Muyun Yang and Tiejun Zhao}, doi = {10.1016/j.neucom.2024.128682}, journal = {Neurocomputing}, month = {1}, pages = {128682}, shortjournal = {Neurocomputing}, title = {Enhancing bilingual lexicon induction via harnessing polysemous words}, volume = {611}, year = {2025}, }
• (2025). DM-HAP: Diffusion model for accurate hand pose prediction. NEUCOM, 611, 128681. (www)

  @article{ , abstract = {Forecasting hand poses is a challenging task due to inherent uncertainties, occlusions, and inaccuracies in 3D pose estimation. Diffusion models provide a promising direction for predicting precise 3D hand poses under noisy conditions. In this work, we introduce Dual-diffusion, an innovative framework for the precise prediction of future hand poses. Our approach leverages the strengths of diffusion models by framing hand pose forecasting as a reverse diffusion process, effectively addressing the complexities of noisy hand joint movements and their subtleties. To enhance the learning of hand pose representations, we propose a unique neural architecture that simultaneously captures both local and global features. This is achieved through the deployment of Global and Local Diffusion (GLD) blocks within our network, which facilitate the exchange of information between local and global features. This diffusion-based interaction enables the integration of global hand actions and local finger actions, leading to a more powerful representation learning approach. We evaluate the effectiveness of our Dual-diffusion method on three public 3D hand pose estimation datasets (MSRA, F-PHAB, and BigHand2.2M), and our approach outperforms previous methods.}, archive = {J_NEUCOM}, author = {Zhifeng Wang and Kaihao Zhang and Ramesh Sankaranarayana}, doi = {10.1016/j.neucom.2024.128681}, journal = {Neurocomputing}, month = {1}, pages = {128681}, shortjournal = {Neurocomputing}, title = {DM-HAP: Diffusion model for accurate hand pose prediction}, volume = {611}, year = {2025}, }
• (2025). GazeGCN: Gaze-aware graph convolutional network for text classification. NEUCOM, 611, 128680. (www)

  @article{ , abstract = {Graph convolutional networks (GCNs) are capable of capturing contextual relationships in text classification. In this paper, we propose a novel Gaze-aware Graph Convolutional Network (GazeGCN) for text classification, where the gaze signals from humans are incorporated into the GCN architecture, empowering our GazeGCN to use gaze information that can be shared across all data to model the intricate relationships between words and documents. To be specific, we first build a gaze prediction model to obtain five gaze signals to form the gaze distribution of each word. Then, Wasserstein Distance is employed to derive the gaze-aware word–word weight by calculating the gaze distribution between words, so as to improve the capture of luxuriant and similar relationships between words. Furthermore, to enhance the extraction of contextual syntactic information, we introduce a Gaze-enhanced TF–IDF method integrating gaze signals and term frequency–inverse document frequency (TF–IDF) for gaze-aware word-document weight derivation, thus making up for the lack of TF–IDF that does not consider syntactic information. Afterward, the rich source of graph edge information including gaze-aware word–word weight and gaze-aware word-document weight is incorporated to construct graphs, so as to leverage the relationship of word–word and word-document. Experiments show encouraging results on seven benchmark datasets that our approach outperforms the state-of-the-art baseline methods.}, archive = {J_NEUCOM}, author = {Bingbing Wang and Bin Liang and Zhixin Bai and Min Yang and Lin Gui and Ruifeng Xu}, doi = {10.1016/j.neucom.2024.128680}, journal = {Neurocomputing}, month = {1}, pages = {128680}, shortjournal = {Neurocomputing}, title = {GazeGCN: Gaze-aware graph convolutional network for text classification}, volume = {611}, year = {2025}, }
• (2025). Comparison of reservoir computing topologies using the recurrent kernel approach. NEUCOM, 611, 128679. (www)

  @article{ , abstract = {Reservoir Computing (RC) has become popular in recent years thanks to its fast and efficient computational capabilities. Standard RC has been shown to be equivalent in the asymptotic limit to Recurrent Kernels, which helps in analyzing its expressive power. However, many well-established RC paradigms, such as Leaky RC, Sparse RC, and Deep RC, are yet to be systematically analyzed in such a way. We define the Recurrent Kernel limit of all these RC topologies and conduct a convergence study for a wide range of activation functions and hyperparameters. Our findings provide new insights into various aspects of Reservoir Computing. First, we demonstrate that there is an optimal sparsity level which grows with the reservoir size. Furthermore, our analysis suggests that Deep RC should use reservoir layers of decreasing sizes. Finally, we perform a benchmark demonstrating the efficiency of Structured Reservoir Computing compared to vanilla and Sparse Reservoir Computing.}, archive = {J_NEUCOM}, author = {Giuseppe Alessio D’Inverno and Jonathan Dong}, doi = {10.1016/j.neucom.2024.128679}, journal = {Neurocomputing}, month = {1}, pages = {128679}, shortjournal = {Neurocomputing}, title = {Comparison of reservoir computing topologies using the recurrent kernel approach}, volume = {611}, year = {2025}, }
• (2025). Off-policy safe reinforcement learning for nonlinear discrete-time systems. NEUCOM, 611, 128677. (www)

  @article{ , abstract = {This paper presents a control barrier function-based approach for learning safe optimal controllers for discrete-time (DT) nonlinear systems such that safety, stability, and performance are guaranteed on an infinite time horizon. The paper investigates the fusion of reinforcement learning (RL) and control barrier functions (CBFs) for DT systems by considering Taylor expansion around operating point taking into account corresponding gradient(s) and Hessian(s) to develop a comprehensive approach for DT systems. To that end, on one hand, the cost-to-go function is augmented with a CBF, and mathematically rigorous proofs are provided to establish the invariance of the safe set, ensuring the safety of system states as well as stability and optimal performance. Formulation of a novel Generalized Safety-aware Hamilton Jacobi Bellman (G-SHJB) equation is proposed to assure safety, stability and optimality during the exploitation phase. The G-SHJB is solved in an iterative sense to obtain improved control policy. The invariance of safe set as well as stability and optimality of the system under the learned control law is established using mathematically rigorous proofs. Additionally, this work proposes an off-policy Safe RL approach and develops off-policy expressions to execute the same. An off-policy G-SHJB expression is developed for DT systems to execute the proposed safe RL approach using neural networks. The effectiveness of the approach is demonstrated using simulation examples.}, archive = {J_NEUCOM}, author = {Mayank Shekhar Jha and Bahare Kiumarsi}, doi = {10.1016/j.neucom.2024.128677}, journal = {Neurocomputing}, month = {1}, pages = {128677}, shortjournal = {Neurocomputing}, title = {Off-policy safe reinforcement learning for nonlinear discrete-time systems}, volume = {611}, year = {2025}, }
• (2025). Growing evolutional deep echo state network. NEUCOM, 611, 128676. (www)

  @article{ , abstract = {Deep echo state network (Deep ESN) is a powerful method for time-series prediction, but developing the optimal structure of Deep ESN remains challenging. A novel generative framework, growing evolutional Deep ESN (GE Deep ESN) is proposed, where networks are generated through alternating growth and evolution. A model compression method is implemented during the evolution phase to optimize the inner structure of the reservoirs, and to automatically determine reservoir size. A simple termination criterion of growth is introduced to determine network depth. The feasibility and superiority of GE Deep ESN were validated experimentally with two benchmark prediction tasks. For real-time trajectory prediction in photoelectric tracking systems, an iterative learning model based on GE Deep ESN is proposed, the real time prediction result showed that network evolution enhances the model’s suitability for this task, and improves prediction precision.}, archive = {J_NEUCOM}, author = {Qingyu Shen and Junzhe Wang and Hanwen Zhang and Jinjin Peng and Minxing Sun and Yao Mao}, doi = {10.1016/j.neucom.2024.128676}, journal = {Neurocomputing}, month = {1}, pages = {128676}, shortjournal = {Neurocomputing}, title = {Growing evolutional deep echo state network}, volume = {611}, year = {2025}, }
• (2025). An architectural analysis of DeepOnet and a general extension of the physics-informed DeepOnet model on solving nonlinear parametric partial differential equations. NEUCOM, 611, 128675. (www)

  @article{ , abstract = {The Deep Neural Operator, as proposed by Lu et al. (2021), marks a considerable advancement in solving parametric partial differential equations. This paper examines the DeepOnet model’s neural network design, focusing on the effectiveness of its trunk-branch structure in operator learning tasks. Three key advantages of the trunk-branch structure are identified: the global learning strategy, the independent operation of the trunk and branch networks, and the consistent representation of solutions. These features are especially beneficial for operator learning. Building upon these findings, we have evolved the traditional DeepOnet into a more general form from a network perspective, allowing a nonlinear interfere of the branch net on the trunk net than the linear combination limited by the conventional DeepOnet. The operator model also incorporates physical information for enhanced integration. In a series of experiments tackling partial differential equations, the extended DeepOnet consistently outperforms than the traditional DeepOnet, particularly in complex problems. Notably, the extended DeepOnet model shows substantial advancements in operator learning with nonlinear parametric partial differential equations and exhibits a remarkable capacity for reducing physics loss.}, archive = {J_NEUCOM}, author = {Haolin Li and Yuyang Miao and Zahra Sharif Khodaei and M.H. Aliabadi}, doi = {10.1016/j.neucom.2024.128675}, journal = {Neurocomputing}, month = {1}, pages = {128675}, shortjournal = {Neurocomputing}, title = {An architectural analysis of DeepOnet and a general extension of the physics-informed DeepOnet model on solving nonlinear parametric partial differential equations}, volume = {611}, year = {2025}, }
• (2025). A novel deep learning algorithm for phaeocystis counting and density estimation based on feature reconstruction and multispectral generator. NEUCOM, 611, 128674. (www)

  @article{ , abstract = {Phaeocystis proliferation is a primary instigator of algal blooms, commonly known as red tides, posing a significant threat to marine life and severely disrupting marine ecosystems. Currently, no effective method exists for estimating Phaeocystis density, underscoring an urgent need for preventative measures against Phaeocystis blooms. Given the challenges associated with the varying sizes and frequent overlapping of Phaeocystis colonies, we propose an innovative counting algorithm that leverages feature reconstruction and multispectral generator modules. Utilizing deep learning, our method achieves accurately real-time density estimation and prediction of Phaeocystis colonies. The algorithm operates in two stages: first, a multispectral reconstruction block is trained to function as a multispectral generator; second, spectral and spatial features are integrated to predict density and perform counting. Our approach surpasses existing algorithms in accuracy for Phaeocystis counting and demonstrates the utility of multispectral data in enhancing the neural network’s ability to discern targets from their background.}, archive = {J_NEUCOM}, author = {Yifan Si and Shuo Li and Sailing He}, doi = {10.1016/j.neucom.2024.128674}, journal = {Neurocomputing}, month = {1}, pages = {128674}, shortjournal = {Neurocomputing}, title = {A novel deep learning algorithm for phaeocystis counting and density estimation based on feature reconstruction and multispectral generator}, volume = {611}, year = {2025}, }
• (2025). Pose-oriented scene-adaptive matching for abnormal event detection. NEUCOM, 611, 128673. (www)

  @article{ , abstract = {For intelligent surveillance systems, abnormal event detection automatically analyses surveillance video sequences and detects abnormal objects or unusual human actions at the frame level. Due to the lack of labelled data, most approaches are semi-supervised based on reconstruction or prediction methods. However, these methods may not generalize well to unseen scene contexts. To address this issue, we present a novel self and mutual scene-adaptive matching method for abnormal event detection. In the framework, we propose synergistic pose estimation and object detection, which effectively integrates human pose and object detection information to improve pose estimation accuracy. Then, the poses are resized to reduce the spatial distance between the source and target domains. The improved pose sequences are further fed into a spatio-temporal graph convolutional network to extract the geometric features. Finally, the features are embedded in a clustering layer to classify action types and compute normality scores. The training data is taken from the training part of common video anomaly detection datasets: UCSD PED1 & PED2, CHUK Avenue, and ShanghaiTech Campus. The proposed framework is evaluated on video sequences with unseen scene contexts in the UCSD PED2 and ShanghaiTech Campus datasets. The detection accuracy and efficiency are also evaluated in detail, and the proposed method for abnormal event detection achieves the highest AUC performance, 84.6%, on the ShanghaiTech Campus dataset and relatively high AUC performance, 96.9% and 74.8%, on UCSD PED2 & PED1 datasets. Compared with other state-of-the-art works, the performance analysis and results confirm the robustness and effectiveness of our proposed framework for cross-scene abnormal event detection.}, archive = {J_NEUCOM}, author = {Yuxing Yang and Leiyu Xie and Zeyu Fu and Jiawei Yan and Syed Mohsen Naqvi}, doi = {10.1016/j.neucom.2024.128673}, journal = {Neurocomputing}, month = {1}, pages = {128673}, shortjournal = {Neurocomputing}, title = {Pose-oriented scene-adaptive matching for abnormal event detection}, volume = {611}, year = {2025}, }
• (2025). A flexible and efficient algorithm for high dimensional support vector regression. NEUCOM, 611, 128671. (www)

  @article{ , abstract = {In high dimensional statistical learning, variable selection and handling highly correlated phenomena are two crucial topics. Elastic-net regularization can automatically perform variable selection and tends to either simultaneously select or remove highly correlated variables. Consequently, it has been widely applied in machine learning. In this paper, we incorporate elastic-net regularization into the support vector regression model, introducing the Elastic-net Support Vector Regression (En-SVR) model. Due to the inclusion of elastic-net regularization, the En-SVR model possesses the capability of variable selection, addressing high dimensional and highly correlated statistical learning problems. However, the optimization problem for the En-SVR model is rather complex, and common methods for solving the En-SVR model are challenging. Nevertheless, we observe that the optimization problem for the En-SVR model can be reformulated as a convex optimization problem where the objective function is separable into multiple blocks and connected by an inequality constraint. Therefore, we employ a novel and efficient Alternating Direction Method of Multipliers (ADMM) algorithm to solve the En-SVR model, and provide a complexity analysis as well as convergence analysis for the algorithm. Furthermore, extensive numerical experiments validate the outstanding performance of the En-SVR model in high dimensional statistical learning and the efficiency of this novel ADMM algorithm.}, archive = {J_NEUCOM}, author = {Menglei Yang and Hao Liang and Xiaofei Wu and Zhimin Zhang}, doi = {10.1016/j.neucom.2024.128671}, journal = {Neurocomputing}, month = {1}, pages = {128671}, shortjournal = {Neurocomputing}, title = {A flexible and efficient algorithm for high dimensional support vector regression}, volume = {611}, year = {2025}, }
• (2025). Hierarchical keypoints feature alignment for domain adaptive pose estimation. NEUCOM, 611, 128670. (www)

  @article{ , abstract = {Unsupervised Domain Adaptation (UDA) is a popular research topic in computer vision. One of the most common strategies in this field is to reduce domain shifts through global feature alignment. Unfortunately, we observe its failure on adaptive human pose estimation. From our analysis, we find out two major reasons: the extreme imbalance between keypoints and non-keypoints regions and the high diversity of human skeleton structures. To address these problems, we propose a simple yet effective approach named PoseDA. Our idea is to let the alignment focus on the feature of keypoint regions and enforce the model to learn domain-invariant representations for keypoint prediction. The key component of our PoseDA is a Hierarchical Feature Selection of Keypoints Regions (HFS) module, which consists of Coarse Feature Selection of Keypoints Regions (CFS) and Reliable Feature Selection of Keypoints Regions (RFS). By using HFS, we obtain reliable and compact keypoints features, allowing our model to achieve more effective feature alignment. Under three UDA scenarios, i.e. JTA → MPII, SURREAL → LSP, SURREAL → Human3.6, our PoseDA establishes new state-of-the-art performance. In particular, our PoseDA outperforms the previous best UDA methods by over 7% w.r.t. PCKh on JTA → MPII.}, archive = {J_NEUCOM}, author = {Jie Xu and Yunan Liu and Jian Yang and Shanshan Zhang}, doi = {10.1016/j.neucom.2024.128670}, journal = {Neurocomputing}, month = {1}, pages = {128670}, shortjournal = {Neurocomputing}, title = {Hierarchical keypoints feature alignment for domain adaptive pose estimation}, volume = {611}, year = {2025}, }
• (2025). Hybrid lane change strategy of autonomous vehicles based on SOAR cognitive architecture and deep reinforcement learning. NEUCOM, 611, 128669. (www)

  @article{ , abstract = {Research on lane change strategies for autonomous vehicles holds paramount importance in optimizing traffic flow efficiency, enhancing driving safety, and adapting to complex traffic environments. While numerous rule-based or machine-learning approaches have been explored to tackle the challenge of lane change on highways, they frequently exhibit limited performance owing to the complexity of driving environments. This study proposes a novel lane change strategy for autonomous vehicles, which utilizes a hybrid framework integrating the SOAR cognitive architecture and deep reinforcement learning (DRL) to address the lane change challenge on highways. First, we introduce a rule extraction algorithm, the RuleCOSI+, which is based on tree ensemble algorithms, designed to extract concise lane change rules from large-scale human driving data. These straightforward rules, together with traffic regulations and safety rules, constitute the long-term memory of the SOAR cognitive architecture, enabling transparent decision-making processes. Next, by analyzing the clipping mechanism of the proximal policy optimization (PPO) algorithm, we propose an Adaptive Clipping PPO (ACPPO) algorithm which is based on the importance of samples. This algorithm adopts different clipping strategies for SOAR samples and ACPPO samples during the training process, enabling the algorithm to more effectively utilize samples with different levels of importance. Then, we propose a hybrid decision-making algorithm: SOAR-ACPPO, which combines the SOAR cognitive architecture with the ACPPO algorithm. This algorithm leverages SOAR’s prior knowledge to effectively and safely guide agent learning. Finally, by selecting appropriate intervention probability and weaning strategy, the system avoids inappropriate knowledge intervention and ensures adequate environment exploration. Simulation experiments conducted using the CARLA simulator illustrate that the proposed strategy not only improves model learning efficiency but also enhances driving efficiency and safety. Additionally, it demonstrates a certain degree of human-like characteristics and interpretability.}, archive = {J_NEUCOM}, author = {Rongliang Zhou and Haotian Cao and Jiakun Huang and Xiaolin Song and Jing Huang and Zhi Huang}, doi = {10.1016/j.neucom.2024.128669}, journal = {Neurocomputing}, month = {1}, pages = {128669}, shortjournal = {Neurocomputing}, title = {Hybrid lane change strategy of autonomous vehicles based on SOAR cognitive architecture and deep reinforcement learning}, volume = {611}, year = {2025}, }
• (2025). Adaptive predefined-time consensus for nonlinear multi-agent systems with input delay and performance constraints. NEUCOM, 611, 128668. (www)

  @article{ , abstract = {This paper investigates the predefined-time tracking consensus issue for a class of non strict-feedback nonlinear multi-agent systems (MASs) under input delay and error performance constraints by adaptive backstepping control strategy. Firstly, the impact of input delay is eliminated by Pade approximation. With the help of an intermediate state variable, the considered MASs with input delay are revised as the independent-delay systems. Secondly, under constrained distributed error and corrected n th virtual error, the ln -type barrier Lyapunov function (BLF) is adopted to solve the error performance constrains and reduce the complexity of differentiation. Meanwhile, an adaptive predefined-time controller is designed to realize the tracking purpose by backstepping method and radial basis function neural networks (RBFNNs) approximation. Finally, a numerical simulation example and a practical application are presented to test the effectiveness for the proposed control strategy.}, archive = {J_NEUCOM}, author = {Xinzhao Hou and Huaiqin Wu and Jinde Cao}, doi = {10.1016/j.neucom.2024.128668}, journal = {Neurocomputing}, month = {1}, pages = {128668}, shortjournal = {Neurocomputing}, title = {Adaptive predefined-time consensus for nonlinear multi-agent systems with input delay and performance constraints}, volume = {611}, year = {2025}, }
• (2025). HiNER: Hierarchical feature fusion for chinese named entity recognition. NEUCOM, 611, 128667. (www)

  @article{ , abstract = {Named Entity Recognition (NER) aims to extract structured entity information from unstructured textual data by identifying entity boundaries and categories. Chinese NER is more challenging than that of English due to the complex structure and ambiguous word boundaries, as well as nested and discontinuous occurrences of entities. Previous Chinese NER methods are limited by their character-based approach and dependence on external lexical information, which is often non-contextualized, leading to the introduction of noise and potentially compromising model performance. This paper proposes a novel Chinese NER model, HiNER, which leverages external semantic enhancement and hierarchical attention fusion. Specifically, we initially formulate the Chinese NER as a character–character relation classification task, thoroughly taking into account the cases of nested and discontinuous entities. Then, by incorporating syntactic information, we develop a Triformer module that is used to better integrate Chinese character, lexical, and syntactic embeddings, carefully considering the impact of external semantic enhancement on the original text embeddings and reducing extrinsic information interference to some extent. In addition, through the fusion of local and global attention mechanisms, the representation of character–character relationships is enhanced, allowing for the effective capture of semantic features at various hierarchical levels within the Chinese context. We conduct extensive experiments on seven Chinese NER datasets, and the results indicate that the HiNER model achieves state-of-the-art (SOTA) performance. The outcomes also confirm that external semantic enhancement and hierarchical attention fusion can provide better assistance in accomplishing the Chinese NER task.}, archive = {J_NEUCOM}, author = {Shuxiang Hou and Yurong Qian and Jiaying Chen and Jigui Zhao and Huiyong Lv and Jiyuan Zhang and Hongyong Leng and Mengnan Ma}, doi = {10.1016/j.neucom.2024.128667}, journal = {Neurocomputing}, month = {1}, pages = {128667}, shortjournal = {Neurocomputing}, title = {HiNER: Hierarchical feature fusion for chinese named entity recognition}, volume = {611}, year = {2025}, }
• (2025). EntroCap: Zero-shot image captioning with entropy-based retrieval. NEUCOM, 611, 128666. (www)

  @article{ , abstract = {State-of-the-art image captioning often relies on supervised models trained on domain-specific image–text pairs, which can be expensive and time-consuming to annotate. Additionally, these models struggle with performance drops when applied to data from different distributions. To overcome these challenges, zero-shot image captioning has emerged, utilizing models trained on unlabeled image and text data. Pre-trained vision-language and generative language models have shown promise in this area, but existing methods do not fully leverage the cross-modal representation capabilities of large-scale models and tend to overlook smaller objects in images. We propose EntroCap, a novel zero-shot image captioning method that incorporates an Entropy-based Retrieval Strategy. During training, our Hierarchical Projector optimizes CLIP’s joint embeddings to capture comprehensive global signals with multi-level representations. During inference, the Entropy-based Retrieval Strategy adjusts prediction logits to better include smaller, less frequent objects. Additionally, our Balancing Gate mechanism guides the language model to produce more detailed and contextually relevant captions by balancing local and global signals. Extensive experiments on MSCOCO, Flickr30k, and NoCaps demonstrate that EntroCap outperforms previous zero-shot approaches in both in-domain and cross-domain settings. We also conduct experiments to quantitatively validate EntroCap’s effectiveness in describing small objects.}, archive = {J_NEUCOM}, author = {Jie Yan and Yuxiang Xie and Shiwei Zou and Yingmei Wei and Xidao Luan}, doi = {10.1016/j.neucom.2024.128666}, journal = {Neurocomputing}, month = {1}, pages = {128666}, shortjournal = {Neurocomputing}, title = {EntroCap: Zero-shot image captioning with entropy-based retrieval}, volume = {611}, year = {2025}, }
• (2025). Dynamic event-triggered finite-horizon robust suboptimal control of multi-player systems with input disturbances. NEUCOM, 611, 128665. (www)

  @article{ , abstract = {This paper presents an adaptive dynamic event-triggered robust finite-horizon suboptimal control scheme for multi-player nonzero-sum game of nonlinear systems with input disturbances based on dynamic programming (ADP) and integral sliding mode (ISM) control techniques. Firstly, dynamic event-triggered adaptive ISM controllers are designed to handle input disturbances that can force the system state to remain on the sliding manifold and relax the known upper bounds condition of input disturbances. Corresponding event-triggered condition is employed for ISM controller of each player to reduce the update frequency. Then, utilizing the obtained ISM dynamics, we employ ADP-based single critic neural networks to approximate the time-varying solution of the Hamilton–Jacobi equations. Thus, the event-triggered finite-horizon suboptimal controllers can be obtained, and corresponding state-based dynamic event-triggered condition is designed to improve the resource utilization. The stability property of the closed-loop system is demonstrated through the Lyapunov technique. Finally, the effectiveness of the proposed control scheme is confirmed through two simulation examples.}, archive = {J_NEUCOM}, author = {Haoming Zou and Guoshan Zhang and Zhiguo Yan and Wanquan Liu}, doi = {10.1016/j.neucom.2024.128665}, journal = {Neurocomputing}, month = {1}, pages = {128665}, shortjournal = {Neurocomputing}, title = {Dynamic event-triggered finite-horizon robust suboptimal control of multi-player systems with input disturbances}, volume = {611}, year = {2025}, }
• (2025). Towards faster yet accurate video prediction for resource-constrained platforms. NEUCOM, 611, 128663. (www)

  @article{ , abstract = {Video prediction has emerged as a critical task in the field of computer vision. While conventional deep learning-based video prediction models excel in accuracy, they often impose significant computational overhead. Moreover, they tend to disregard crucial factors like inference latency and memory consumption, which may render them unsuitable for real-time video prediction applications. This issue is further exacerbated by the necessity to deploy many of these applications on resource-limited embedded devices. In this study, we introduce an innovative faster yet accurate video prediction (FAVP) model designed to mitigate inference latency and memory footprint, all the while maintaining predictive accuracy. Our model employs an encoder-convertor-decoder architecture with a novel multi-layer convertor. Each layer within the convertor uses consistent kernel sizes to precisely capture dimensional features, while varying kernel sizes between layers effectively capture both local and global features. This combination of uniformity within layers and diversity between layers enhances the model’s ability to grasp dynamic changes. Furthermore, we demonstrate that replacing conventional large-kernel convolutions with involutions significantly trims the model’s parameter count and inference latency, without compromising prediction accuracy. Through comprehensive experiments conducted on resource-adequate x86 platforms, utilizing KTH, TrafficBJ, Human3.6M, Moving MNIST, and KTH-Enhanced datasets, we showcase our model’s exceptional prediction performance, particularly in terms of inference latency. Additionally, we evaluate the model on resource-constrained NVIDIA Jetson Nano platform using KITTI & CalTech Pedestrian dataset. The results underscore our model’s superior inference speed and its competence in meeting real-time video prediction requirements, even under stringent resource limitations.}, archive = {J_NEUCOM}, author = {Junhong Zhu and Junyu Lai and Lianqiang Gan and Huashuo Liu and Lianli Gao}, doi = {10.1016/j.neucom.2024.128663}, journal = {Neurocomputing}, month = {1}, pages = {128663}, shortjournal = {Neurocomputing}, title = {Towards faster yet accurate video prediction for resource-constrained platforms}, volume = {611}, year = {2025}, }
• (2025). Fractional-order spike-timing-dependent gradient descent for multi-layer spiking neural networks. NEUCOM, 611, 128662. (www)

  @article{ , abstract = {Accumulated detailed knowledge about the neuronal activities in human brains has brought more attention to bio-inspired spiking neural networks (SNNs). In contrast to non-spiking deep neural networks (DNNs), SNNs can encode and transmit spatiotemporal information more efficiently by exploiting biologically realistic and low-power event-driven neuromorphic architectures. However, the supervised learning of SNNs still remains a challenge because the spike-timing-dependent plasticity (STDP) of connected spiking neurons is difficult to implement and interpret in existing backpropagation learning schemes. This paper proposes a fractional-order spike-timing-dependent gradient descent (FO-STDGD) learning model by considering a derived nonlinear activation function that describes the relationship between the quasi-instantaneous firing rate and the temporal membrane potentials of nonleaky integrate-and-fire neurons. The training strategy can be generalized to any fractional orders between 0 and 2 since the FO-STDGD incorporates the fractional gradient descent method into the calculation of spike-timing-dependent loss gradients. The proposed FO-STDGD model is tested on the MNIST and DVS128 Gesture datasets and its accuracy under different network structure and fractional orders is analyzed. It can be found that the classification accuracy increases as the fractional order increases, and specifically, the case of fractional order 1.9 improves by 155 % relative to the case of fractional order 1 (traditional gradient descent). In addition, our scheme demonstrates the state-of-the-art computational efficacy for the same SNN structure and training epochs.}, archive = {J_NEUCOM}, author = {Yi Yang and Richard M. Voyles and Haiyan H. Zhang and Robert A. Nawrocki}, doi = {10.1016/j.neucom.2024.128662}, journal = {Neurocomputing}, month = {1}, pages = {128662}, shortjournal = {Neurocomputing}, title = {Fractional-order spike-timing-dependent gradient descent for multi-layer spiking neural networks}, volume = {611}, year = {2025}, }
• (2025). Relationship detection for manipulation in object stacking scene with fully connected CRF. NEUCOM, 611, 128661. (www)

  @article{ , abstract = {Relationship detection for manipulation in object stacking scenes is an actively studied problem in recent years. Since the stacked objects may occlude and block each other, it is necessary to model relationships among them for safety. Almost all existing solutions to this problem model object relationships separately and ignore their dependencies on each other. In this paper, we propose a fully connected Conditional Random Field (CRF) to impose global constraints on the object stacking scenes. Since the inference of fully connected CRF is usually intractable, we propose efficient methods for the exact inference and the variational inference to solve the proposed CRFs. Furthermore, the variational inference is implemented as an RNN for end-to-end training. Experimental results on Visual Manipulation Relationship Dataset VMRD and REGRAD suggest that the proposed CRF can, with a clear margin, improve the accuracy of manipulation relationship detection in object stacking scenes. Finally, the proposed models are also deployed on a physical robot to build a relation-aware grasping system. Results demonstrate that our algorithms can be generalized to real-world scenes well and help the robot perform safe grasping in object stacking scenes.}, archive = {J_NEUCOM}, author = {Mengyuan Ding and Chenjie Yang and Jianyi Liu and Xuguang Lan and Nanning Zheng}, doi = {10.1016/j.neucom.2024.128661}, journal = {Neurocomputing}, month = {1}, pages = {128661}, shortjournal = {Neurocomputing}, title = {Relationship detection for manipulation in object stacking scene with fully connected CRF}, volume = {611}, year = {2025}, }
• (2025). Event-based adaptive formation and tracking control with predetermined performance for nonlinear multi-agent systems. NEUCOM, 611, 128660. (www)

  @article{ , abstract = {The article focuses on the event-based predetermined performance formation tracking control for uncertain nonlinear multi-agent systems. Each agent is subject to actuator saturation constraint and full-state constraints. To meet the full-state constraints, a barrier function-based nonlinear mapping method is employed instead of the barrier Lyapunov function, such that the undesirable “feasibility conditions” are eradicated. To overcome actuator saturation, the auxiliary systems are utilized to let the controller be designed under the framework of backstepping. The approximation capacity of RBF NNs and adaptive methods are adopted, which can solve unknown uncertainties. An event-triggered predetermined performance formation tracking control strategy is designed, which can guarantee that the agents form the prescribed formation shape in a predetermined settling time and have a fine transient and steady-state tracking performance, while reducing the communication burden from the controller to the actuator. At last, an example of unmanned aerial vehicles is given to test the availability of the results.}, archive = {J_NEUCOM}, author = {Ru Chang and Yujie Liu and Xiaobo Chi and Changyin Sun}, doi = {10.1016/j.neucom.2024.128660}, journal = {Neurocomputing}, month = {1}, pages = {128660}, shortjournal = {Neurocomputing}, title = {Event-based adaptive formation and tracking control with predetermined performance for nonlinear multi-agent systems}, volume = {611}, year = {2025}, }
• (2025). Pure anomaly detection via self-supervised deep metric learning with adaptive margin. NEUCOM, 611, 128659. (www)

  @article{ , abstract = {We address the problem of anomaly detection (AD) by a deep network pretrained using self-supervised learning for an auxiliary geometric transformation (GT) classification task. Our key contribution is a novel loss function that augments the standard cross-entropy by an additional term that plays a significant role in the later stages of self-supervised learning. The proposed enabling innovation is a triplet centre loss with an adaptive margin and a learnable metric, which relentlessly drives the GT classes to exhibit continuously improving compactness and inter-class separation. The pretrained network is finetuned for the downstream task using non-anomalous data only, and a GT model for the data is constructed. Anomalies are detected by fusing the output of several decision functions defined using the learnt GT class model. In contrast to the majority of existing methods, our approach strictly adheres to the pure AD design philosophy, which relies on the use of purely non-anomalous data for the design. Extensive experiments on four publicly available AD datasets demonstrate the effectiveness of the proposed contributions and lead to significant performance gains compared to the state-of-the-art (1.8% on F-MNIST, 1.0% on CIFAR-10, 1.2% on CIFAR-100, and 1.7% on CatVsDog). https://github.com/12sf12/Deep-Anomaly-Detection}, archive = {J_NEUCOM}, author = {Soroush Fatemifar and Muhammad Awais and Ali Akbari and Josef Kittler}, doi = {10.1016/j.neucom.2024.128659}, journal = {Neurocomputing}, month = {1}, pages = {128659}, shortjournal = {Neurocomputing}, title = {Pure anomaly detection via self-supervised deep metric learning with adaptive margin}, volume = {611}, year = {2025}, }
• (2025). QSIM: A quantum-inspired hierarchical semantic interaction model for text classification. NEUCOM, 611, 128658. (www)

  @article{ , abstract = {Semantic interaction modeling is a fundamental technology in natural language understanding that guides models to extract deep semantic information from text. Currently, the attention mechanism is one of the most effective techniques in semantic interaction modeling, which learns word-level attention representation by measuring the relevance between different words. However, the attention mechanism is limited to word-level semantic interaction, it cannot meet the needs of fine-grained interactive information for some text classification tasks. In recent years, quantum-inspired language modeling methods have successfully constructed quantized representations of language systems in Hilbert spaces, which use density matrices to achieve fine-grained semantic interaction modeling. This paper presents a Q uantum-inspired hierarchical S emantic I nteraction M odel ( QSIM ), which follows the sememe-word-sentence language construction principle and utilizes quantum entanglement theory to capture hierarchical semantic interaction information in Hilbert space. Our work builds on the idea of the attention mechanism and extends it. Specifically, we explore the original semantic space from a quantum theory perspective and derive the core semantic space using the Schmidt decomposition technique, where: (1) Sememe is represented as the unit vector in the two-dimensional minimum semantic space; (2) Word is represented as reduced density matrices in the core semantic space, where Schmidt coefficients quantify sememe-level semantic interaction. Compared to density matrices, reduced density matrices capture fine-grained semantic interaction information with lower computational cost; (3) Sentence is represented as quantum superposition states of words, and the degree of word-level semantic interaction is measured using entanglement entropy. To evaluate the model’s performance, we conducted experiments on 15 text classification datasets. The experimental results demonstrate that our model is superior to classical neural network models and traditional quantum-inspired language models. Furthermore, the experiment also confirms two distinct advantages of QISM: (1) flexibility , as it can be integrated into various mainstream neural network text classification architectures; and (2) practicability , as it alleviates the problem of parameter growth inherent in density matrix calculation in quantum language model.}, archive = {J_NEUCOM}, author = {Hui Gao and Peng Zhang and Jing Zhang and Chang Yang}, doi = {10.1016/j.neucom.2024.128658}, journal = {Neurocomputing}, month = {1}, pages = {128658}, shortjournal = {Neurocomputing}, title = {QSIM: A quantum-inspired hierarchical semantic interaction model for text classification}, volume = {611}, year = {2025}, }
• (2025). Spiking neural networks for event-based action recognition: A new task to understand their advantage. NEUCOM, 611, 128657. (www)

  @article{ , abstract = {Spiking Neural Networks (SNN) are characterised by their unique temporal dynamics, but the properties and advantages of such computations are still not well understood. In order to provide answers, in this work we demonstrate how Spiking neurons can enable temporal feature extraction in feed-forward neural networks without the need for recurrent synapses, and how recurrent SNNs can achieve comparable results to LSTM with a smaller number of parameters. This shows how their bio-inspired computing principles can be successfully exploited beyond energy efficiency gains and evidences their differences with respect to conventional artificial neural networks. These results are obtained through a new task, DVS-Gesture-Chain (DVS-GC), which allows, for the first time, to evaluate the perception of temporal dependencies in a real event-based action recognition dataset. Our study proves how the widely used DVS Gesture benchmark can be solved by networks without temporal feature extraction when its events are accumulated in frames, unlike the new DVS-GC which demands an understanding of the order in which events happen. Furthermore, this setup allowed us to reveal the role of the leakage rate in spiking neurons for temporal processing tasks and demonstrated the benefits of ”hard reset” mechanisms. Additionally, we also show how time-dependent weights and normalisation can lead to understanding order by means of temporal attention. Code for the DVS-GC task is available.}, archive = {J_NEUCOM}, author = {Alex Vicente-Sola and Davide L. Manna and Paul Kirkland and Gaetano Di Caterina and Trevor J. Bihl}, doi = {10.1016/j.neucom.2024.128657}, journal = {Neurocomputing}, month = {1}, pages = {128657}, shortjournal = {Neurocomputing}, title = {Spiking neural networks for event-based action recognition: A new task to understand their advantage}, volume = {611}, year = {2025}, }
• (2025). Saliency and location aware pruning of deep visual detectors for autonomous driving. NEUCOM, 611, 128656. (www)

  @article{ , abstract = {Despite the remarkable achievements of deep neural networks, their high computational complexity limits their wide use in many real-world embedded applications, such as autonomous driving perception. While current neural network pruning approaches can reduce model complexity to various extents, they often adopt local or ad hoc importance measures that are not directly related to the final task. More importantly, most of them focus on classification tasks and do not take location information into consideration during pruning. To address these issues, we present a novel channel importance measure that incorporates detection-related saliency and location awareness, specifically designed for the pruning of self-driving visual detectors. Our comprehensive experiments on the KITTI and COCO_traffic datasets demonstrate that our pruning method achieves significant reductions in model size and computational operations with little performance degradation. Notably, it outperforms other state-of-the-art methods across various pruning rates and base detectors. Our pruned YOLOX-S model with 40.2% fewer parameters even improves the original model’s mAP by 1.8% on KITTI. Moreover, we experimentally highlight the potential of our pruning approach in effectively detecting small-scale objects.}, archive = {J_NEUCOM}, author = {Jung Im Choi and Qing Tian}, doi = {10.1016/j.neucom.2024.128656}, journal = {Neurocomputing}, month = {1}, pages = {128656}, shortjournal = {Neurocomputing}, title = {Saliency and location aware pruning of deep visual detectors for autonomous driving}, volume = {611}, year = {2025}, }
• (2025). A two-stage strategy for brain-inspired unsupervised learning in spiking neural networks. NEUCOM, 611, 128655. (www)

  @article{ , abstract = {Spiking Neural Networks (SNNs) are emerging as a promising model in the field of neuromorphic computing, offering possibilities for efficient bio-inspired computing. Unsupervised learning in SNNs plays a crucial role in exploring neural network dynamics and achieving efficient artificial intelligence systems. However, current unsupervised learning in SNNs mostly relies on single-stage learning strategies, which limits the possibility of combining diverse learning mechanisms and network structures, resulting in networks that lack flexibility and adaptability for low-power task processing. Inspired by the multi-stage and multi-strategy learning in the human brain, this paper proposes a two-stage learning strategy from coarse to fine to enhance the efficiency of unsupervised learning in SNNs. To quickly extract structural features from spike data in the first stage, we propose the Spiking-SOM algorithm, which effectively learns patterns in spike data and achieves clustering effects. Additionally, we integrate Rate encoding and Temporal encoding through a visual attention mechanism, where this hybrid encoding helps meet the need for rapid information processing in the first stage, while also aiding in detailed feature representation in the second stage. Compared to the traditional single-stage learning framework, this work maintains high accuracy in unsupervised classification while significantly reducing computational costs and speeding up training. This exploration offers a new direction for future learning frameworks in SNNs, helping to further improve the energy efficiency of the network.}, archive = {J_NEUCOM}, author = {Zhen Cao and Chuanfeng Ma and Biao Hou and Xiaoyu Chen and Leida Li and Hao Zhu and Dou Quan and Licheng Jiao}, doi = {10.1016/j.neucom.2024.128655}, journal = {Neurocomputing}, month = {1}, pages = {128655}, shortjournal = {Neurocomputing}, title = {A two-stage strategy for brain-inspired unsupervised learning in spiking neural networks}, volume = {611}, year = {2025}, }
• (2025). LinFa-Q: Accurate Q-learning with linear function approximation. NEUCOM, 611, 128654. (www)

  @article{ , abstract = {Although Q-learning has achieved remarkable success in some practical cases, it often suffers from the overestimation problem in stochastic environments, which is commonly viewed as a shortcoming of Q-learning. Overestimated values are introduced by estimations on the next state Q value, which is well-known as the maximization bias. In this paper, we propose a more accurate method for estimating the Q value by the Q value decomposition and re-evaluation with similar samples based on linear function approximation. Specifically, we reform the parameterized incremental update formula of Q-learning and also demonstrate that the new formula is equivalent to the original one. Moreover, we propose a new parameterized incremental update formula of Q-learning to address the overestimation problem and present the more accurate computing method, which can be used in problems with continuous state spaces and stochastic environments. Experimentally, when compared with Doubly Bounded Q-learning and other Q-learning based methods, the new algorithm has more than 31% improvement of performance in Mountain Car and Cart Pole. Furthermore, the algorithm is robust to the learning rate and its memory capacity. Finally, the practical applicability of our algorithm is discussed through an analysis of time consumption.}, archive = {J_NEUCOM}, author = {Zhechao Wang and Qiming Fu and Jianping Chen and Quan Liu and You Lu and Hongjie Wu and Fuyuan Hu}, doi = {10.1016/j.neucom.2024.128654}, journal = {Neurocomputing}, month = {1}, pages = {128654}, shortjournal = {Neurocomputing}, title = {LinFa-Q: Accurate Q-learning with linear function approximation}, volume = {611}, year = {2025}, }
• (2025). Federated learning data protection scheme based on personalized differential privacy in psychological evaluation. NEUCOM, 611, 128653. (www)

  @article{ , abstract = {Federated learning enables multi-party model training by utilizing shared models instead of raw data, allowing for effective use of user data while ensuring privacy protection. However, the training process still has potential threats. Guided by the safe sharing and analysis of data in psychological evaluation, a federated learning data protection scheme based on differential privacy FL-PDP is proposed. First, local model parameters are perturbed using personalized differential privacy techniques to provide varying levels of privacy protection. Second, the privacy configuration selected by the user is divided into multiple levels and uploaded to the server after encryption to avoid targeted attacks in the uplink channel. Third, a probability-based selection method is employed for model aggregation to mitigate the impact of high privacy protection requirements on the model aggregation process. Finally, after aggregation is completed on the server side, noise that meets the global differential privacy threshold is added to quantify the level of global privacy protection. Experimental results indicate that this algorithm achieves high model accuracy, low communication latency, and scalability while ensuring privacy protection.}, archive = {J_NEUCOM}, author = {Xueyan Liu and Jia Wang and Xin Xiong and Hao Sun}, doi = {10.1016/j.neucom.2024.128653}, journal = {Neurocomputing}, month = {1}, pages = {128653}, shortjournal = {Neurocomputing}, title = {Federated learning data protection scheme based on personalized differential privacy in psychological evaluation}, volume = {611}, year = {2025}, }
• (2025). Optimal deception attacks under energy harvesting constraints for cyber–physical systems. NEUCOM, 611, 128652. (www)

  @article{ , abstract = {This paper investigates the design of optimal deception attacks against remote state estimation, where the attacker is equipped with energy harvesting capability to supply more energy resources for implementing the attacks. To derive the optimal attack parameters and energy allocation strategy collaboratively, the attack design problem is transformed into a multi-variable constrained optimization problem which is solved by employing the proposed stepwise optimization method without sacrificing the optimality. Specifically, the analytical expression of the optimal attack parameters is firstly derived based on the presented data isolation technique, which removes the conservatism of the existing results where the attack schemes are merely suboptimal or the attack targets are limited to the systems without redundant sensors. Then, by analyzing the intrinsic property of the average estimation error, the optimal allocation of the dynamic attack energy is obtained to maximize the attack effect. Finally, the simulations are given to verify the developed results.}, archive = {J_NEUCOM}, author = {Xuan Liu and Guang-Hong Yang and Georgi Marko Dimirovski and Yi-Gang Li}, doi = {10.1016/j.neucom.2024.128652}, journal = {Neurocomputing}, month = {1}, pages = {128652}, shortjournal = {Neurocomputing}, title = {Optimal deception attacks under energy harvesting constraints for cyber–physical systems}, volume = {611}, year = {2025}, }
• (2025). Metaphor generation based on noval evaluation method. NEUCOM, 611, 128651. (www)

  @article{ , abstract = {Metaphor generation is a difficult research area to study. In this task, the generated content must maintain certain elements of the original content, such as verbs, adjectives, and compound phrases that convey a deeper meaning than their literal meanings. However, there is a lack of a large and accurate parallel corpus for metaphor generation, which makes it difficult to train effective models. As a result, most current studies on metaphor generation rely on constructing pseudo-parallel corpora, which are often of low quality and lack diversity. To address this problem, we propose the creation of a large-scale annotated parallel corpus for metaphor generation. This corpus consists of 18,000 samples that have been carefully screened by professionals and cover metaphorical sentences that include verbs, nouns, and compound phrases. In addition, we propose an automatic evaluation metric called the MS-score, which is based on metaphoricity and similarity, to assess the quality of metaphors generated by the model. We validate the consistency of the automatic evaluation metric with manual evaluation metrics and use the MS-score to re-rank the generated metaphors.}, archive = {J_NEUCOM}, author = {Chang Su and Xingyue Wang and Yongzhu Chang and Kechun Wu and Yijiang Chen}, doi = {10.1016/j.neucom.2024.128651}, journal = {Neurocomputing}, month = {1}, pages = {128651}, shortjournal = {Neurocomputing}, title = {Metaphor generation based on noval evaluation method}, volume = {611}, year = {2025}, }
• (2025). Supervised learning of spatial features with STDP and homeostasis using spiking neural networks on SpiNNaker. NEUCOM, 611, 128650. (www)

  @article{ , abstract = {Artificial Neural Networks (ANN) have gained significant popularity thanks to their ability to learn using the well-known backpropagation algorithm. Conversely, Spiking Neural Networks (SNNs), despite having broader capabilities than ANNs, have always posed challenges in the training phase. This paper shows a new method to perform supervised learning on SNNs, using Spike Timing Dependent Plasticity (STDP) and homeostasis, aiming at training the network to identify spatial patterns. Spatial patterns refer to spike patterns without a time component, where all spike events occur simultaneously. The method is tested using the SpiNNaker digital architecture. A SNN is trained to recognise one or multiple patterns and performance metrics are extracted to measure the performance of the network. Some considerations are drawn from the results showing that, in the case of a single trained pattern, the network behaves as the ideal detector, with 100% accuracy in detecting the trained pattern. However, as the number of trained patterns on a single network increases, the accuracy of identification is linked to the similarities between these patterns. This method of training an SNN to detect spatial patterns may be applied to pattern recognition in static images or traffic analysis in computer networks, where each network packet represents a spatial pattern. It will be stipulated that the homeostatic factor may enable the network to detect patterns with some degree of similarity, rather than only perfectly matching patterns. The principles outlined in this article serve as the fundamental building blocks for more complex systems that utilise both spatial and temporal patterns by converting specific features of input signals into spikes. One example of such a system is a computer network packet classifier, tasked with real-time identification of packet streams based on features within the packet content.}, archive = {J_NEUCOM}, author = {Sergio Davies and Andrew Gait and Andrew Rowley and Alessandro Di Nuovo}, doi = {10.1016/j.neucom.2024.128650}, journal = {Neurocomputing}, month = {1}, pages = {128650}, shortjournal = {Neurocomputing}, title = {Supervised learning of spatial features with STDP and homeostasis using spiking neural networks on SpiNNaker}, volume = {611}, year = {2025}, }
• (2025). Consistency prototype module and motion compensation for few-shot action recognition (CLIP-CPM2C). NEUCOM, 611, 128649. (www)

  @article{ , abstract = {Recently, few-shot action recognition has progressed significantly, as it has learned the feature discriminability and designed suitable comparison methods. Still, there are the following restrictions. (a) Previous works are mainly based on visual mono-modal. Although some multi-modal works use labels as supplementary to construct prototypes of support videos, they cannot use this information for query videos. The labels are not used efficiently. (b) Most of the works ignore the motion feature of video, although the motion features are essential for distinguishing. We proposed a Consistency Prototype and Motion Compensation Network(CLIP-CPM 2 C) based on CLIP to address these issues. Firstly, in order to make the amount of information between the prototype and the query more similar, we propose a novel method to compensate for the text(prompt) information of query videos when text(prompt) does not exist, which depends on a Consistency Loss. Secondly, we use the differential features of the adjacent frames in two directions as the motion features, which explicitly embeds the network with motion dynamics, and then apply the Consistency Loss to the motion features. Extensive experiments on standard benchmark datasets demonstrate that the proposed method can compete with state-of-the-art results. Our code is available at the URL: https://github.com/cofly2014/CPM-MC.git .}, archive = {J_NEUCOM}, author = {Fei Guo and YiKang Wang and Han Qi and Li Zhu and Jing Sun}, doi = {10.1016/j.neucom.2024.128649}, journal = {Neurocomputing}, month = {1}, pages = {128649}, shortjournal = {Neurocomputing}, title = {Consistency prototype module and motion compensation for few-shot action recognition (CLIP-CPM2C)}, volume = {611}, year = {2025}, }
• (2025). Finite-time synchronization for a fully complex-valued BAM inertial neural network with proportional delays via non-reduced order and non-separation approach. NEUCOM, 611, 128648. (www)

  @article{ , abstract = {This paper focuses on finite-time synchronization (FinTS) for fully complex-valued bidirectional associative memory inertial neural networks (FCV-BAM-INNs) with proportional delays via non-reduced order and non-separation approach. Firstly, a model of the FCV-BAM-INNs with proportional delays is established. Secondly, by using the non-reduced order transformation method of inertial neural networks (INNs) and the non-separation method of complex-valued neural networks (CVNNs), an adjustable complex feedback control law with memory is designed to reach the FinTS of the drive–response FCV-BAM-INNs with proportional delays. Thirdly, some new delay-independent algebraic sufficient criteria are formed to guarantee FinTS of FCV-BAM-INNs with proportional delays by applying non smooth theory and inequalities in the complex domain, which are flexible and adjustable and easy to implement in applications. Moreover, the settling time of FinTS theoretically is calculated. Finally, some numerical simulations and an application on image encryption and decryption as evidences of the validity and availability of the results.}, archive = {J_NEUCOM}, author = {Liyan Duan and Junmin Li}, doi = {10.1016/j.neucom.2024.128648}, journal = {Neurocomputing}, month = {1}, pages = {128648}, shortjournal = {Neurocomputing}, title = {Finite-time synchronization for a fully complex-valued BAM inertial neural network with proportional delays via non-reduced order and non-separation approach}, volume = {611}, year = {2025}, }
• (2025). Exploring fusion domain: Advancing infrared and visible image fusion via IDFFN-GAN. NEUCOM, 611, 128647. (www)

  @article{ , abstract = {Infrared (IR) and visible (VI) images are crucial in applications such as surveillance and night vision, where each modality provides complementary information—IR captures thermal details, while VI captures textures. Fusing these images is essential to combine their strengths, resulting in a more comprehensive and informative image. In this work, we introduce the “fusion domain” concept, a unique distribution that optimally blends IR and VI features. Our primary contribution is developing the Intermediate Domain Feature Fusion Network (IDFFN), which employs an MLP to learn the optimal domain factor for this fusion. Integrated with the IDFFN-GAN’s dual discriminators, our approach refines the fusion process to produce images that better preserve thermal and texture details from the proposed fusion and gradient domains. Experimental results demonstrate that our method achieves a 5% boost in Entropy (EN), a 50% rise in Spatial Frequency (SF), a 58% improvement in Average Gradient (AG), a 10% enhancement in Standard Deviation (SD), and a remarkable 198% increase in N A B / F compared to state-of-the-art techniques, ensuring superior retention of specific features from both modalities.}, archive = {J_NEUCOM}, author = {Juan Li and Xiaoqian Shi and Yanan Li and Huabing Zhou}, doi = {10.1016/j.neucom.2024.128647}, journal = {Neurocomputing}, month = {1}, pages = {128647}, shortjournal = {Neurocomputing}, title = {Exploring fusion domain: Advancing infrared and visible image fusion via IDFFN-GAN}, volume = {611}, year = {2025}, }
• (2025). MFGCN: Multimodal fusion graph convolutional network for speech emotion recognition. NEUCOM, 611, 128646. (www)

  @article{ , abstract = {Speech emotion recognition (SER) is challenging owing to the complexity of emotional representation. Hence, this article focuses on multimodal speech emotion recognition that analyzes the speaker’s sentiment state via audio signals and textual content. Existing multimodal approaches utilize sequential networks to capture the temporal dependency in various feature sequences, ignoring the underlying relations in acoustic and textual modalities. Moreover, current feature-level and decision-level fusion methods have unresolved limitations. Therefore, this paper develops a novel multimodal fusion graph convolutional network that comprehensively executes information interactions within and between the two modalities. Specifically, we construct the intra-modal relations to excavate exclusive intrinsic characteristics in each modality. For the inter-modal fusion, a multi-perspective fusion mechanism is devised to integrate the complementary information between the two modalities. Substantial experiments on the IEMOCAP and RAVDESS datasets and experimental results demonstrate that our approach achieves superior performance.}, archive = {J_NEUCOM}, author = {Xin Qi and Yujun Wen and Pengzhou Zhang and Heyan Huang}, doi = {10.1016/j.neucom.2024.128646}, journal = {Neurocomputing}, month = {1}, pages = {128646}, shortjournal = {Neurocomputing}, title = {MFGCN: Multimodal fusion graph convolutional network for speech emotion recognition}, volume = {611}, year = {2025}, }
• (2025). Secret image restoration with interpolation and social network search. NEUCOM, 611, 128643. (www)

  @article{ , abstract = {Image restoration is an essential task in image processing. Due to the secret image extracted from the corrupted stego-image and its pixels may suffer different levels of damage, the type of noise is particular. Most of the existing methods have a challenge in directly removing this type of noise. Hence, we propose a corrupted secret image restoration algorithm with interpolation and social network search. Firstly, a corrupted secret image is divided into several blocks, and the convex hull of each block is calculated. For the corrupted pixels in the convex hull, the natural neighbor interpolation and bi-harmonic spline interpolation are utilized to generate the estimation values of corrupted pixels according to the density of the trusted pixels. For the corrupted pixels outside the convex hull, the social network search algorithm is applied to generate the estimated values of corrupted pixels. Finally, all corrupted pixels are recovered by the ones of their candidate values closest to their estimated values. The performance of the various methodologies was evaluated using the USCSIPI database, employing peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indices, which demonstrated an enhancement of more than 0.78 dB and 0.0431, respectively. Experiments show that our algorithm outperforms other state-of-the-art (SOTA) algorithms.}, archive = {J_NEUCOM}, author = {Jianzhong Yang and Xianquan Zhang and Chunqiang Yu and Xuemao Zhang and Guoxiang Li and Zhenjun Tang}, doi = {10.1016/j.neucom.2024.128643}, journal = {Neurocomputing}, month = {1}, pages = {128643}, shortjournal = {Neurocomputing}, title = {Secret image restoration with interpolation and social network search}, volume = {611}, year = {2025}, }
• (2025). An efficient multimodal attentional principal component analysis for continual learning-based dynamic process monitoring. NEUCOM, 611, 128642. (www)

  @article{ , abstract = {Traditional multimode process monitoring methods extract features from time series data. Due to the catastrophic forgetting effect, data-driven multimode dynamic process monitoring is challenging based on a single monitoring model paradigm, i.e. the learned knowledge from previous modes may diminish as operating conditions undergo changes between modes, yet it is impractical to access all past data to retrain the model. In this work, a novel efficient method of multimodal attentional principal component analysis (M-APCA) with continual learning ability is introduced. Under the assumption that data from successive modes are received sequentially, dynamic process data are modeled using an attention mechanism to capture the relationship between data and the latent space, whereby meaningful information is concentrated as dynamic features which are extracted via a vector autoregressive model. In order to overcome the catastrophic forgetting problem, the idea of replay continual learning is employed. Specifically, past modes’ data which are significant to reflect the operating conditions, are selected and stored. These are repeatedly used in tandem with sequential data as replay data. Two types of attention mechanisms are considered and analyzed, each of which is specifically designed to learn from data in an unsupervised manner, so the overall algorithm is efficient both in time and storage costs. The proposed attentional principal component analysis and M-APCA are analyzed against several state-of-the-art methods to highlight the virtues of the proposed method. Compared with multimode monitoring methods, the effectiveness is demonstrated through case studies of: a continuous stirred tank heater, the Tennessee Eastman process and a practical coal pulverizing system.}, archive = {J_NEUCOM}, author = {Jingxin Zhang and Haikun Wei and Kanjian Zhang and James Xiao and Xia Hong}, doi = {10.1016/j.neucom.2024.128642}, journal = {Neurocomputing}, month = {1}, pages = {128642}, shortjournal = {Neurocomputing}, title = {An efficient multimodal attentional principal component analysis for continual learning-based dynamic process monitoring}, volume = {611}, year = {2025}, }
• (2025). Double kernel and minimum variance embedded broad learning system based autoencoder for one-class classification. NEUCOM, 611, 128638. (www)

  @article{ , abstract = {One-class classification methods are often used for anomaly detection in healthcare, quality control in manufacturing, and fraud detection in financial services. Particularly in medical datasets, instances of rare cancers are considerably fewer in comparison to those of healthy individuals. One-class classification methods, which concentrate on learning models using just samples from a specified class, have drawn increasing attention as a solution to this problem. To enhance the model’s ability to capture essential feature information, we propose a novel strategy in this paper. This strategy involves the initial application of double kernel mapping, translating the original features into a new feature space. The autoencoder based on the Broad Learning System is then modified to incorporate the minimum variance so that the model can learn the potential feature information. We validated the effectiveness of our proposed strategy using UCI dataset and concordia digits dataset. To compare the performance with other one-class methods, the average F 1 scores and the average G − m e a n are employed as comparative measures. Based on experimental results, it is evident that DKVBLS-AE not only outperforms but also demonstrates greater stability compared to the existing methods.}, archive = {J_NEUCOM}, author = {Ningxia He and Junwei Duan and Jun Lyu}, doi = {10.1016/j.neucom.2024.128638}, journal = {Neurocomputing}, month = {1}, pages = {128638}, shortjournal = {Neurocomputing}, title = {Double kernel and minimum variance embedded broad learning system based autoencoder for one-class classification}, volume = {611}, year = {2025}, }
• (2025). 3D visual grounding-audio: 3D scene object detection based on audio. NEUCOM, 611, 128637. (www)

  @article{ , abstract = {3D Visual Grounding (3DVG) is a prevalent multi-modal information fusion task capable of accurately localizing target objects referenced in natural language descriptions within a point cloud scene. Nevertheless, the stringent demands for input and output devices present substantial hurdles for the application and integration of 3DVG in fields like remote robotic control and telemedicine. To address this challenge, we introduce several innovative approaches. Firstly, we have initiated a novel multi-modal task, termed 3D Visual Grounding-Audio (3DVG-Audio), which is based on the fusion of audio and point cloud. To the best of our knowledge, this represents the first instance of an Audio-Point Cloud multi-modal task. 3DVG-Audio achieves precise localization of audio-mentioned objects within the point cloud by utilizing the point cloud in conjunction with the corresponding audio input. Secondly, building upon the ScanRefer, we have developed a novel dataset named 3DVG-AudioSet, specifically designed for the training and evaluation of the 3DVG-Audio method. Finally, we have crafted a tailored loss function to further enhance the performance of 3DVG-Audio and introduced a method named AP-Refer, which serves as a benchmark for the task. Extensive experimental results demonstrate the potential for deep integration of audio and point cloud to tackle complex real-world challenges. AP-Refer has successfully addressed the 3DVG-Audio, circumventing the limitations of conventional 3DVG methods, and exhibits significant application potential.}, archive = {J_NEUCOM}, author = {Can Zhang and Zeyu Cai and Xunhao Chen and Feipeng Da and Shaoyan Gai}, doi = {10.1016/j.neucom.2024.128637}, journal = {Neurocomputing}, month = {1}, pages = {128637}, shortjournal = {Neurocomputing}, title = {3D visual grounding-audio: 3D scene object detection based on audio}, volume = {611}, year = {2025}, }
• (2025). Unsupervised anomaly detection of multivariate time series based on multi-standard fusion. NEUCOM, 611, 128634. (www)

  @article{ , abstract = {With the wide application of the Internet of Things, industrial systems generate a large amount of multivariate time series data every day. Industrial data often contains a wide variety of anomalies. The existing anomaly detection methods have some shortcomings in dealing with large-scale, high-dimensional, and real-time industrial data. Particularly in the absence of prior knowledge, unsupervised learning methods are easily influenced by noise. Therefore, we propose a new unsupervised anomaly detection algorithm of multivariate time series based on multi-standard fusion (MSAD). Specifically, MSAD first classifies the data by analyzing the data density and sample spacing, thereby converting unsupervised anomaly detection into weakly supervised anomaly detection. Then, based on the characteristics of the data and deep learning technology, the degree of abnormality in the sample is comprehensively measured from four perspectives: holistic, local, time, and deep feature learning. At this stage, MSAD 's ability to learn edge features of data has been improved. Finally, based on MSAD, we design an anomaly detection method for industrial data streams in combination with data filtering techniques. The experimental results show that MSAD can detect anomalies more accurately and show stronger robustness to the influence of noise.}, archive = {J_NEUCOM}, author = {Huixin Tian and Hao Kong and Shikang Lu and Kun Li}, doi = {10.1016/j.neucom.2024.128634}, journal = {Neurocomputing}, month = {1}, pages = {128634}, shortjournal = {Neurocomputing}, title = {Unsupervised anomaly detection of multivariate time series based on multi-standard fusion}, volume = {611}, year = {2025}, }
• (2025). SPMGAE: Self-purified masked graph autoencoders release robust expression power. NEUCOM, 611, 128631. (www)

  @article{ , abstract = {To tackle the scarcity of labeled graph data, graph self-supervised learning (SSL) has branched into two paradigms: Generative methods and Contrastive methods. Inspired by MAE and BERT in computer vision (CV) and natural language processing (NLP), masked graph autoencoders (MGAEs) are gaining popularity in the generative genre. However, prevailing MGAEs are mostly designed under the assumption that the data has high homophilic score and is out of adversarial distortion. When people deliberately improve the performance on homophilic graph datasets, they ignore a critical issue that both internal heterophily and artificial attack noise are quite common in the real world. Therefore, when data itself is highly heterophilic or confronted with attacks, they merely have no defensive capability. Especially under self-supervised conditions, it is much more difficult to detect internal heterophily and resist artificial attacks. In this paper, we propose a Self-Purified Masked Graph Autoencoder (SPMGAE) to make up for the shortcomings of prevailing MGAEs in terms of robustness. SPMGAE first utilizes a self-purified module to prune raw graph data and separate perturbation information. The purified graph provides a robust graph structure for the entire pre-training process. Next, the encoding module reuses perturbation information for auxiliary training to enhance robustness, while the decoding module reconstructs the effective graph data at a finer granularity. Extensive experiments on homophilic and heterophilic datasets attacked by various attack methods demonstrate SPMGAE has a considerable robust expressive ability. Especially on small datasets with large perturbations, the improvement of defensive performance could reaches 10%–25%.}, archive = {J_NEUCOM}, author = {Shuhan Song and Ping Li and Ming Dun and Yuan Zhang and Huawei Cao and Xiaochun Ye}, doi = {10.1016/j.neucom.2024.128631}, journal = {Neurocomputing}, month = {1}, pages = {128631}, shortjournal = {Neurocomputing}, title = {SPMGAE: Self-purified masked graph autoencoders release robust expression power}, volume = {611}, year = {2025}, }
• (2025). SCGC: Self-supervised contrastive graph clustering. NEUCOM, 611, 128629. (www)

  @article{ , abstract = {Graph clustering discovers groups or communities within networks. Increasingly, models use autoencoders to achieve effective clustering combined with Graph Neural Networks (GNN) for structure incorporation. However, GNNs based on convolution or attention variants lack dynamic fusion, suffer from over-smoothing, noise, node heterophily, are computationally expensive and typically require the complete graph being present. Instead, we propose SCGC, capable of dynamic soft structure fusion via augmentation-less edge-contrastive loss. Further, we propose SCGC*, with a more expressive novel distance metric, Influence, and our Influence Augmented Contrastive (IAC) loss, requiring only half the model parameters. Our models, SCGC and SCGC*, dynamically fuse discriminative node representations, jointly refine soft cluster assignments, completely eliminate convolutions and attention of traditional GNNs, use only simple linear units, and yet efficiently incorporate structure. They are impervious to layer depth; robust to over-smoothing, incorrect edges and heterophily; scalable by batching; augmentation-less; relaxes the homophily assumption and trivially parallelizable. We improve significantly over the state-of-the-art on a wide range of benchmarks, including images, sensor data, text, and citation networks, with superb efficiency. Specifically, 20% on ARI and 18% on NMI for DBLP; overall 55% reduction in training time and overall, 81% reduction on inference time. code: https://github.com/gayanku/SCGC .}, archive = {J_NEUCOM}, author = {Gayan K. Kulatilleke and Marius Portmann and Shekhar S. Chandra}, doi = {10.1016/j.neucom.2024.128629}, journal = {Neurocomputing}, month = {1}, pages = {128629}, shortjournal = {Neurocomputing}, title = {SCGC: Self-supervised contrastive graph clustering}, volume = {611}, year = {2025}, }
• (2025). ShaderNN: A lightweight and efficient inference engine for real-time applications on mobile GPUs. NEUCOM, 611, 128628. (www)

  @article{ , abstract = {Inference using deep neural networks on mobile devices has been an active area of research in recent years. The design of a deep learning inference framework targeted for mobile devices needs to consider various factors, such as the limited computational capacity of the devices, low power budget, varied memory access methods, and I/O bus bandwidth governed by the underlying processor’s architecture. Furthermore, integrating an inference framework with time-sensitive applications — such as games and video-based software to perform tasks like ray tracing denoising and video processing — introduces the need to minimize data movement between processors and increase data locality in the target processor. In this paper, we propose Shader Neural Network (ShaderNN), an OpenGL-based, fast, and power-efficient inference framework designed for mobile devices to address these challenges. Our contributions include the following: (1) the texture-based input/output provides an efficient, zero-copy integration with real-time graphics pipelines or image processing applications, thereby saving expensive data transfers between CPU and GPU, which are unavoidable in most existing inference engines; (2) we are the first to leverage fragment shaders based on the OpenGL backend in neural network inference operators, which has an advantage in deploying parametrically small neural network models; (3) a hybrid implementation of the compute shader and fragment shader is proposed that enables layer-level shader selection to boost performance; and (4) we utilize OpenGL features — such as normalization, interpolation and texture padding — to improve performance. Experiments illustrate the favorable performance of ShaderNN over other popular on-device deep learning frameworks such as TensorFlow-Lite on the latest mobile devices powered by Qualcomm and MediaTek chips. A case study further demonstrates the usability and integration of the ShaderNN framework with a media processing Android application seamlessly. ShaderNN is available open source at Github ( https://github.com/inferenceengine/shadernn ).}, archive = {J_NEUCOM}, author = {Jing Xie and Yuzhong Yan and Abhishek Saxena and Qiang Qiu and Jiangong Chen and Hongyu Sun and Rong Chen and Shuvra S. Bhattacharyya}, doi = {10.1016/j.neucom.2024.128628}, journal = {Neurocomputing}, month = {1}, pages = {128628}, shortjournal = {Neurocomputing}, title = {ShaderNN: A lightweight and efficient inference engine for real-time applications on mobile GPUs}, volume = {611}, year = {2025}, }
• (2025). Multi-view clustering with adaptive anchor and bipartite graph learning. NEUCOM, 611, 128627. (www)

  @article{ , abstract = {Multi-view subspace clustering optimizes and integrates the graph structure information of each view. At present, the subspace clustering methods based on the abstract graph have better performance and improve the clustering results. Although the existing clustering algorithms have achieved excellent results, there are still four deficiencies: 1) Expensive time overhead, with most algorithms having a high time complexity; 2) Using fixed anchor points and the separation of anchor graph learning from subsequent graph construction, resulting in inadequate use of underlying graph structure between views; 3) Multi-view data have inevitable noise and in the original high-dimensional space data fusion may cause the loss of important information; 4) Most algorithms require additional post-processing to obtain clustering labels. To solve the above problems, we innovate a novel anchor-adaptive multi-view clustering algorithm based on a bipartite graph . Specifically, anchor graph learning and subspace graph construction are adaptively combined into a unified optimization framework. The projection matrix, consensus anchor matrix, and similarity matrix optimize each other to promote the clustering effect and structure a constrained bipartite graph to describe the relationship between representative points and sample points. At the same time, connectivity constraints are used to ensure that the connected components represent clusters directly. Our method has linear time complexity about sample size. Comparison experiments with the state-of-the-art algorithms demonstrate the effectiveness of the proposed method on various benchmark datasets. Moreover, our method is robust to noisy data and suitable for large-scale datasets.}, archive = {J_NEUCOM}, author = {Shibing Zhou and Xi Wang and Mingrui Yang and Wei Song}, doi = {10.1016/j.neucom.2024.128627}, journal = {Neurocomputing}, month = {1}, pages = {128627}, shortjournal = {Neurocomputing}, title = {Multi-view clustering with adaptive anchor and bipartite graph learning}, volume = {611}, year = {2025}, }
• (2025). Beyond the visible: A survey on cross-spectral face recognition. NEUCOM, 611, 128626. (www)

  @article{ , abstract = {Cross-spectral face recognition (CFR) refers to recognizing individuals using face images stemming from different spectral bands, such as infrared versus visible. While CFR is inherently more challenging than classical face recognition due to significant variation in facial appearance caused by the modality gap , it is useful in many scenarios including night-vision biometrics and detecting presentation attacks. Recent advances in deep neural networks (DNNs) have resulted in significant improvement in the performance of CFR systems. Given these developments, the contributions of this survey are three-fold. First, we provide an overview of CFR, by formalizing the CFR problem and presenting related applications. Secondly, we discuss the appropriate spectral bands for face recognition and discuss recent CFR methods, placing emphasis on deep neural networks . In particular we describe techniques that have been proposed to extract and compare heterogeneous features emerging from different spectral bands. We also discuss the datasets that have been used for evaluating CFR methods. Finally, we discuss the challenges and future lines of research on this topic.}, archive = {J_NEUCOM}, author = {David Anghelone and Cunjian Chen and Arun Ross and Antitza Dantcheva}, doi = {10.1016/j.neucom.2024.128626}, journal = {Neurocomputing}, month = {1}, pages = {128626}, shortjournal = {Neurocomputing}, title = {Beyond the visible: A survey on cross-spectral face recognition}, volume = {611}, year = {2025}, }
• (2025). Knowledge complementation based dual-table filling approach for aspect sentiment triplet extraction. NEUCOM, 611, 128625. (www)

  @article{ , abstract = {Aspect Sentiment Triplet Extraction (ASTE) task aims to extract aspect terms, opinion terms, and the corresponding sentiment polarity of aspect-opinion pairs from review sentences. Current outperforming table-filling approaches focus on enhancing the relation representation within a single table, with various algorithms designed to exploit deeper information and additional knowledge. However, due to the sequential execution of these algorithms on a single table, the granularity of the knowledge evolves unidirectionally, inevitably leading to the loss of knowledge at different granularities from the previous stage of the entire process. To overcome this limitation, we propose a simple yet effective end-to-end model for the ASTE task, namely D ual- T able F illing (DTF), which constructs two complementary tables instead of just one to capture different granularities of knowledge in a sentence. Specifically, the main table exploits latent knowledge and hidden dependencies, while the assistant table captures more fine-grained contextual knowledge. This structure allows the model to utilize knowledge at multiple granularities more efficiently while maintaining simplicity. In addition, a Syntactic Enhanced Encoder (SEE) is developed to acquire syntactic-enhanced knowledge from sentences, enhancing the comprehension of sentence structure and word relationships. Experimental results on public benchmark datasets show the effectiveness and exceptional performance of our proposed model.}, archive = {J_NEUCOM}, author = {Bao Wang and Baoxing Jiang and Guangjin Wang and Liang Shi and Peiyu Liu}, doi = {10.1016/j.neucom.2024.128625}, journal = {Neurocomputing}, month = {1}, pages = {128625}, shortjournal = {Neurocomputing}, title = {Knowledge complementation based dual-table filling approach for aspect sentiment triplet extraction}, volume = {611}, year = {2025}, }
• (2025). Prompt-supervised dynamic attention graph convolutional network for skeleton-based action recognition. NEUCOM, 611, 128623. (www)

  @article{ , abstract = {Skeleton-based action recognition is a core task in the field of video understanding. Skeleton sequences are characterized by high information density, low redundancy, and clear structural information, thereby facilitating the analysis of complex relationships among human behaviors more readily than other modalities. Although existing studies have encoded skeleton data and achieved positive outcomes, they have often overlooked the precise high-level semantic information inherent in the action descriptions. To address this issue, this paper proposes a prompt-supervised dynamic attention graph convolutional network (PDA-GCN). Specifically, the PDA-GCN incorporates a prompt supervision (PS) module that leverages a pre-trained large-scale language model (LLM) as a knowledge engine and retains the generated text features as prompts to provide additional supervision during model training, enhancing the model’s ability to discern analogous actions with negligible computational cost. In addition, for the purpose of bolstering the learning of discriminative features, a dynamic attention graph convolution (DA-GC) module is presented. This module utilizes self-attention mechanism to adaptively infer intrinsic relationships between joints and integrates dynamic convolution to strengthen the emphasis on local information. This dual focus on both global context and local details further amplifies the efficiency and effectiveness of the model. Extensive experiments, conducted on the widely-used skeleton-based action recognition datasets NTU RGB+D 60 and NTU RGB+D 120, demonstrate that the PDA-GCN surpasses known state-of-the-art methods, achieving accuracies of 93.4% on the NTU RGB+D 60 cross-subject split and 90.7% on the NTU RGB+D 120 cross-subject split.}, archive = {J_NEUCOM}, author = {Shasha Zhu and Lu Sun and Zeyuan Ma and Chenxi Li and Dongzhi He}, doi = {10.1016/j.neucom.2024.128623}, journal = {Neurocomputing}, month = {1}, pages = {128623}, shortjournal = {Neurocomputing}, title = {Prompt-supervised dynamic attention graph convolutional network for skeleton-based action recognition}, volume = {611}, year = {2025}, }
• (2025). Multi-model based MPC to improve the ramp rate for renewable energy consumption. NEUCOM, 611, 128619. (www)

  @article{ , abstract = {Accommodating mass renewable energy into the power grid requires coal-fired power plants to trace reference trajectory, which varies rapidly and is prescribed by the power dispatching system. The Coordinating Control System (CCS) is impotent because decoupling nets work well only on one linearized point of this nonlinear system. This paper presents a multi-observer prediction framework based on model predictive control (MPC) to address the tricky problem. In this framework, multiple state observers predict the output states within the predictive horizon. The arbitration mechanism chooses the right observer on the actual outputs of the system. The picked observer will calculate control inputs subject to energy constraints. Then, the issue is transformed into a quadratic programming formulation. Numerical computation based on gradient descent is used to obtain the optimal control inputs. The experimental results demonstrate that the outputs of coal-fired plants perform more flexibly under the proposed framework than CCS. This approach can improve the capability of equipment.}, archive = {J_NEUCOM}, author = {Shiquan Ma and Jinliang Ding}, doi = {10.1016/j.neucom.2024.128619}, journal = {Neurocomputing}, month = {1}, pages = {128619}, shortjournal = {Neurocomputing}, title = {Multi-model based MPC to improve the ramp rate for renewable energy consumption}, volume = {611}, year = {2025}, }
• (2025). Decoupled semantic graph neural network for knowledge graph embedding. NEUCOM, 611, 128614. (www)

  @article{ , abstract = {Knowledge graph embedding (KGE) learns an embedding space for a more accurate representation of entities and relations. Although the KGE model has proliferated, it often fails to fully capture the rich semantics in knowledge graphs. Some studies attempt to apply decoupling methods to decompose node representations, yet they neglect the semantic noise generated during the decoupled process. To address these issues, we introduce the decoupled semantic graph network for KGE (named DSGNet), which employs a novel approach that combines semantic decoupling with structural awareness. DSGNet begins by projecting the knowledge graph into distinct semantic spaces, ensuring minimal correlation between them to achieve effective semantic decoupling. To mitigate semantic noise, a top- k sampling technique is applied to the decoupled graphs. DSGNet then aggregates these different semantic graphs using a relation-aware aggregation module, followed by a multi-layer aggregation process for enhanced node representation. The extensive experiments demonstrate that DSGNet performs competitively on two popular datasets. We make our code publicly available at https://github.com/HubuKG/DSGNet .}, archive = {J_NEUCOM}, author = {Zhifei Li and Wei Huang and Xuchao Gong and Xiangyu Luo and Kui Xiao and Honglian Deng and Miao Zhang and Yan Zhang}, doi = {10.1016/j.neucom.2024.128614}, journal = {Neurocomputing}, month = {1}, pages = {128614}, shortjournal = {Neurocomputing}, title = {Decoupled semantic graph neural network for knowledge graph embedding}, volume = {611}, year = {2025}, }
• (2025). ESCAPE: Energy-based selective adaptive correction for out-of-distribution 3D human pose estimation. NEUCOM, 611, 128605. (www)

  @article{ , abstract = {Despite recent advances in human pose estimation (HPE), poor generalization to out-of-distribution (OOD) data remains a difficult problem. While previous works have proposed Test-Time Adaptation (TTA) to bridge the train-test domain gap by refining network parameters at inference, the absence of ground-truth annotations makes it highly challenging and existing methods typically increase inference times by one or more orders of magnitude. We observe that (1) not every test time sample is OOD, and (2) HPE errors are significantly larger on distal keypoints (wrist, ankle). To this end, we propose ESCAPE: a lightweight correction and selective adaptation framework which applies a fast, forward-pass correction on most data while reserving costly TTA for OOD data. The free energy function is introduced to separate OOD samples from incoming data and a correction network is trained to estimate the errors of pretrained backbone HPE predictions on the distal keypoints. For OOD samples, we propose a novel self-consistency adaptation loss to update the correction network by leveraging the constraining relationship between distal keypoints and proximal keypoints (shoulders, hips), via a second “reverse” network. ESCAPE improves the distal MPJPE of five popular HPE models by up to 7% on unseen data, achieves state-of-the-art results on two popular HPE benchmarks, and is significantly faster than existing adaptation methods.}, archive = {J_NEUCOM}, author = {Luke Bidulka and Mohsen Gholami and Jiannan Zheng and Martin J. McKeown and Z. Jane Wang}, doi = {10.1016/j.neucom.2024.128605}, journal = {Neurocomputing}, month = {1}, pages = {128605}, shortjournal = {Neurocomputing}, title = {ESCAPE: Energy-based selective adaptive correction for out-of-distribution 3D human pose estimation}, volume = {611}, year = {2025}, }
• (2025). Mask-free iterative refinement network for weakly-supervised few-shot semantic segmentation. NEUCOM, 611, 128600. (www)

  @article{ , abstract = {The dependence on densely labeled samples in Few-shot Semantic Segmentation (FSS) poses challenges in terms of sample annotation. This paper introduces a more challenging few-shot semantic segmentation task that only utilizes intra-class images as weak supervision. To address this task, we propose a Mask-free Iterative Refinement Network consisting of a mask generation module (MGM) and an iterative refinement module (IRM), respectively addressing the inherent two challenges of locating segmented objects and deriving class-specific features in the absence of support mask and semantic labels. MGM generates pseudo-masks for the support image without requiring any training, which provides a rough estimation of the object locations, serving as an initial guidance for segmentation. IRM is designed to capture the co-occurrence class-specific information between the support and query images for segmentation and progressively refine the predicted mask in a bootstrap manner with the guidance of the class-specific information. Experimental results on three FSS benchmarks, i.e., FSS-1000, PASCAL-5i, and COCO-20i demonstrate that our proposed method achieves comparable or decent performance compared to existing zero-shot and supervised segmentation methods.}, archive = {J_NEUCOM}, author = {Shanjuan Chen and Yunlong Yu and Yingming Li and Ziqian Lu and Yulin Zhou}, doi = {10.1016/j.neucom.2024.128600}, journal = {Neurocomputing}, month = {1}, pages = {128600}, shortjournal = {Neurocomputing}, title = {Mask-free iterative refinement network for weakly-supervised few-shot semantic segmentation}, volume = {611}, year = {2025}, }
• (2025). Multi-view clustering based on pairwise co-regularization and robust dual graph non-negative matrix factorization. NEUCOM, 611, 128594. (www)

  @article{ , abstract = {Multi-view clustering has been a hotspot in the field of machine learning and pattern recognition, and methods based on non-negative matrix factorization have gained attention for their simplicity and interpretability. Despite these methods achieving great clustering performance, there may also be some limitations, such as the full structural information of data and the similarity between different views not being considered. This paper proposes a novel multi-view clustering algorithm based on pairwise co-regularization and robust dual graph non-negative matrix factorization. Firstly, the L 2 , p -norm is applied in non-negative matrix factorization framework to enhance model robustness. Next, pairwise co-regularization is utilized to extract the inter-view information of all views. Then, graph dual regularization is applied to preserve the structure information of the data and feature spaces. Lastly, an auto-weighted strategy is introduced to assign appropriate weights to each view. In addition, an iterative updating optimization scheme for the proposed algorithm is developed, and the convergence proof of the scheme is provided. The experimental results on twelve real-world datasets show that the proposed algorithm is both effective and efficient, compared with three classical and six state-of-the-art algorithms.}, archive = {J_NEUCOM}, author = {Huiguo Tang and Shihu Liu and Yi Tang and Fusheng Yu}, doi = {10.1016/j.neucom.2024.128594}, journal = {Neurocomputing}, month = {1}, pages = {128594}, shortjournal = {Neurocomputing}, title = {Multi-view clustering based on pairwise co-regularization and robust dual graph non-negative matrix factorization}, volume = {611}, year = {2025}, }
• (2025). Decision-making model for selecting products through online product reviews utilizing natural language processing techniques. NEUCOM, 611, 128593. (www)

  @article{ , abstract = {Online product reviews are evaluations of products shared by customers on various online platforms, such as electronic commerce websites, social media, or dedicated review sites. These reviews offer valuable insights and opinions about products, potentially influencing the purchasing decisions of other prospective buyers. However, due to the vast volume of content provided by internet product evaluations, it can be challenging for new buyers to conduct a comprehensive qualitative assessment of competitive products. Utilizing natural language processing techniques, analyzing online product reviews posted on social media platforms can aid customers in making informed purchasing decisions. Hence, this article employs a 2-tuple linguistic q -rung orthopair fuzzy set for product selection based on online product reviews. In this study, we introduce a novel hybrid approach combining the Criteria Importance through Inter-Criteria Correlation (CRITIC) method with the Weighted Aggregated Sum Product Assessment (WASPAS) method. This hybrid method assists potential customers in evaluating alternative products by considering consumer opinions regarding product performance in the 2-tuple linguistic q -rung orthopair fuzzy environment. Moreover, for multi-attribute group decision-making problems, we develop a weighted average aggregation operator based on the 2-tuple linguistic q -rung orthopair fuzzy set. Finally, we apply the proposed approach to a real evaluation decision, validating its validity and practicality through parameter and comparison analyses.}, archive = {J_NEUCOM}, author = {Sumera Naz and Aqsa Shafiq and Shariq Aziz Butt and Rabia Tasneem and Dragan Pamucar and Zhoe Comas Gonzalez}, doi = {10.1016/j.neucom.2024.128593}, journal = {Neurocomputing}, month = {1}, pages = {128593}, shortjournal = {Neurocomputing}, title = {Decision-making model for selecting products through online product reviews utilizing natural language processing techniques}, volume = {611}, year = {2025}, }
• (2025). Reinforcement learning-based robust formation control for multi-UAV systems with switching communication topologies. NEUCOM, 611, 128591. (www)

  @article{ , abstract = {This paper introduces a novel optimal robust formation method for quadcopter multiple unmanned aerial vehicle (multi-UAV) systems. Firstly, a reinforcement learning (RL) algorithm based on a unique gradient descent training approach is proposed to solve the Hamilton–Jacobi–Bellman (HJB) equation, which can effectively eliminate the requirement of the Persistent Excitation (PE) condition. Secondly, the robustness of the controlled system is emphasized, and an Uncertainty and Disturbance Estimator (UDE) observer is developed to suppress model uncertainty and external disturbances through filtering techniques. Furthermore, a switched sliding mode control technique according to the average dwell time (ADT) is employed to convert switching communication topology between UAVs dynamically, and the stability analysis of the corresponding closed-loop control systems is then performed by the use of Lyapunov analysis. Finally, the simulation examples are provided to verify the effectiveness of the designed control strategy.}, archive = {J_NEUCOM}, author = {Hongsheng Sha and Rongwei Guo and Jin Zhou and Xiaojin Zhu and Jinchen Ji and Zhonghua Miao}, doi = {10.1016/j.neucom.2024.128591}, journal = {Neurocomputing}, month = {1}, pages = {128591}, shortjournal = {Neurocomputing}, title = {Reinforcement learning-based robust formation control for multi-UAV systems with switching communication topologies}, volume = {611}, year = {2025}, }
• (2025). The underlying mechanisms of alignment in error backpropagation through arbitrary weights. NEUCOM, 611, 128587. (www)

  @article{ , abstract = {Understanding the mechanisms by which plasticity in millions of synapses in the brain is orchestrated to achieve behavioral and cognitive goals is a fundamental question in neuroscience. In this regard, insights from learning methods in artificial neural networks (ANNs) and in particular supervised learning using backpropagation (BP) seem inspiring. However, the implementation of BP requires exact matching between forward and backward weights, which is unrealistic given the known connectivity pattern in the brain (known as the “weight transport problem”). Notably, it has been shown that under certain conditions, error B ack P ropagation T hrough A rbitrary W eights (BP-TAW) can lead to a partial alignment between forward and backward weights (weight alignment or WA). This learning algorithm, which is also known as feedback alignment (FA), can result in surprisingly good degrees of accuracy in simple classification tasks. However, the underlying mechanisms and mathematical basis of WA are not thoroughly understood. In this work, we demonstrate the mathematical basis of WA and answer the question of why and in what conditions WA occurs. We show that the occurrence of WA in ANNs is induced by statistical properties of the output and error signals of neurons, such as autocorrelation and cross-correlation, and can happen even in the absence of learning or reduction of the loss function. Moreover, we show that WA can be improved significantly by limiting the norm of input weights to neurons and that such a weight normalization (WN) method can improve the classification accuracy of BP-TAW. The findings presented can be used to further improve the performance of BP-TAW and open new ways for exploring possible learning mechanisms in biological neural networks without exact matching between forward and backward weights.}, archive = {J_NEUCOM}, author = {Alireza Rahmansetayesh and Ali Ghazizadeh and Farokh Marvasti}, doi = {10.1016/j.neucom.2024.128587}, journal = {Neurocomputing}, month = {1}, pages = {128587}, shortjournal = {Neurocomputing}, title = {The underlying mechanisms of alignment in error backpropagation through arbitrary weights}, volume = {611}, year = {2025}, }
• (2025). TUMbRAIN: A transformer with a unified mobile residual attention inverted network for diagnosing brain tumors from magnetic resonance scans. NEUCOM, 611, 128583. (www)

  @article{ , abstract = {Diagnosing tumors in Magnetic Resonance Imaging (MRI) brain scans is challenging and can lead to errors, even for radiologists. Deep learning, mainly through deep convolutional neural networks, has assisted in automating the diagnosis of these scans. However, there is still room for improvement. Researchers have shown that transformer models hold promise but often remain underutilized due to their need for large amounts of data and complexity compared to traditional neural networks. This paper introduces a new hybrid model called TUMbRAIN (Transformer with a Unified Mobile Residual Attention Inverted Network), which combines a lightweight transformer with a deep convolutional neural network to address these issues. TUMbRAIN incorporates innovative components designed for this purpose, such as the expanded inverted residual block and the unified self-attention mechanism. The results demonstrate that TUMbRAIN outperforms many existing state-of-the-art neural network models, achieving an impressive overall accuracy of 97.94 % with only 1.04 million parameters. These results suggest that hybrid transformer models like TUMbRAIN could significantly advance the automated diagnosis of brain tumors from MRI scans. The study also offers new insights into effectively integrating transformers into traditional neural network architectures, resulting in a cost-effective and accurate deep learning solution for medical imaging. By incorporating these advanced components, TUMbRAIN enhances support for radiological practice through improved diagnostic accuracy and efficiency.}, archive = {J_NEUCOM}, author = {Francis Jesmar P. Montalbo}, doi = {10.1016/j.neucom.2024.128583}, journal = {Neurocomputing}, month = {1}, pages = {128583}, shortjournal = {Neurocomputing}, title = {TUMbRAIN: A transformer with a unified mobile residual attention inverted network for diagnosing brain tumors from magnetic resonance scans}, volume = {611}, year = {2025}, }
• (2025). Analysis of regularized federated learning. NEUCOM, 611, 128579. (www)

  @article{ , abstract = {Federated learning is an efficient machine learning tool for dealing with heterogeneous big data and privacy protection. Federated learning methods with regularization can control the level of communications between the central and local machines. Stochastic gradient descent is often used for implementing such methods on heterogeneous big data, to reduce the communication costs. In this paper, we consider such an algorithm called Loopless Local Gradient Descent which has advantages in reducing the expected communications by controlling a probability level. We improve the method by allowing flexible step sizes and carry out novel analysis for the convergence of the algorithm in a non-convex setting in addition to the standard strongly convex setting. In the non-convex setting, we derive rates of convergence when the smooth objective function satisfies a Polyak-Łojasiewicz condition. When the objective function is strongly convex, a sufficient and necessary condition for the convergence in expectation is presented.}, archive = {J_NEUCOM}, author = {Langming Liu and Ding-Xuan Zhou}, doi = {10.1016/j.neucom.2024.128579}, journal = {Neurocomputing}, month = {1}, pages = {128579}, shortjournal = {Neurocomputing}, title = {Analysis of regularized federated learning}, volume = {611}, year = {2025}, }
• (2025). Cooperative output regulation problem with faults based on adaptive event-triggered mechanism. NEUCOM, 611, 128381. (www)

  @article{ , abstract = {This paper focuses on the cooperative output regulation issue of heterogeneous linear multiagent systems (MAS)s under directed graph, where there are communication link faults and actuator faults. Firstly, a novel adaptive event-triggered compensator is constructed to estimate the state of the exosystem in the presence of communication link faults. Then, dynamic event-triggered controller is constructed to compensate for the impact of actuator faults on system stability. Note that the novel dual event-triggered mechanisms are designed to reduce communication frequency, and the MAS does not show the Zeno behavior. Finally, simulation experiment is introduced to verify the validity of the proposed method.}, archive = {J_NEUCOM}, author = {Xin Wang and Dongsheng Yang and Weihua Li and Beijia Zhao}, doi = {10.1016/j.neucom.2024.128381}, journal = {Neurocomputing}, month = {1}, pages = {128381}, shortjournal = {Neurocomputing}, title = {Cooperative output regulation problem with faults based on adaptive event-triggered mechanism}, volume = {611}, year = {2025}, }